Randomized Clinical Trials of Machine Learning Interventions in Health Care: A Systematic Review

Re-engineering the clinical approach to suspected cardiac chest pain assessment in the emergency department by expediting research evidence to practice using artificial intelligence. (RAPIDx AI)-a cluster randomized study design

BACKGROUND

Clinical work-up for suspected cardiac chest pain is resource intensive. Despite expectations, high-sensitivity cardiac troponin assays have not made decision making easier. The impact of recently validated rapid triage protocols including the 0-hour/1-hour hs-cTn protocols on care and outcomes may be limited by the heterogeneity in interpretation of troponin profiles by clinicians. We have developed machine learning (ML) models which digitally phenotype myocardial injury and infarction with a high predictive performance and provide accurate risk assessment among patients presenting to EDs with suspected cardiac symptoms. The use of these models may support clinical decision-making and allow the synthesis of an evidence base particularly in non-T1MI patients however prospective validation is required.

OBJECTIVE

We propose that integrating validated real-time artificial intelligence (AI) methods into clinical care may better support clinical decision-making and establish the foundation for a self-learning health system.

DESIGN

This prospective, multicenter, open-label, cluster-randomized clinical trial within blinded endpoint adjudication across 12 hospitals (n = 20,000) will randomize sites to the clinical decision-support tool or continue current standard of care. The clinical decision support tool will utilize ML models to provide objective patient-specific diagnostic probabilities (ie, likelihood for Type 1 myocardial infarction [MI] versus Type 2 MI/Acute Myocardial Injury versus Chronic Myocardial Injury etc.) and prognostic assessments. The primary outcome is the composite of cardiovascular mortality, new or recurrent MI and unplanned hospital re-admission at 12 months post index presentation.

SUMMARY

Supporting clinicians with a decision support tool that utilizes AI has the potential to provide better diagnostic and prognostic assessment thereby improving clinical efficiency and establish a self-learning health system continually improving risk assessment, quality and safety.

TRIAL REGISTRATION

ANZCTR, Registration Number: ACTRN12620001319965, https://www.anzctr.org.au/.

Triaging mammography with artificial intelligence: an implementation study

PURPOSE

Many breast centers are unable to provide immediate results at the time of screening mammography which results in delayed patient care. Implementing artificial intelligence (AI) could identify patients who may have breast cancer and accelerate the time to diagnostic imaging and biopsy diagnosis.

METHODS

In this prospective randomized, unblinded, controlled implementation study we enrolled 1000 screening participants between March 2021 and May 2022. The experimental group used an AI system to prioritize a subset of cases for same-visit radiologist evaluation, and same-visit diagnostic workup if necessary. The control group followed the standard of care. The primary operational endpoints were time to additional imaging (TA) and time to biopsy diagnosis (TB).

RESULTS

The final cohort included 463 experimental and 392 control participants. The one-sided Mann-Whitney U test was employed for analysis of TA and TB. In the control group, the TA was 25.6 days [95% CI 22.0-29.9] and TB was 55.9 days [95% CI 45.5-69.6]. In comparison, the experimental group's mean TA was reduced by 25% (6.4 fewer days [one-sided 95% CI > 0.3], p<0.001) and mean TB was reduced by 30% (16.8 fewer days; 95% CI > 5.1], p=0.003). The time reduction was more pronounced for AI-prioritized participants in the experimental group. All participants eventually diagnosed with breast cancer were prioritized by the AI.

CONCLUSIONS

Implementing AI prioritization can accelerate care timelines for patients requiring additional workup, while maintaining the efficiency of delayed interpretation for most participants. Reducing diagnostic delays could contribute to improved patient adherence, decreased anxiety and addressing disparities in access to timely care.

Health professionals' perspectives on the use of Artificial Intelligence in healthcare: A systematic review

INTRODUCTION

Artificial Intelligence (AI) is fast emerging as a crucial tool for improving patient care and treatment outcomes; however, concerns persist among health professionals about potential compromises in quality care and loss of jobs. The availability of systematic evidence on health professionals' perspectives on AI in healthcare is limited.

OBJECTIVE

This systematic review aims to document the perceived advantages and disadvantages associated with AI applications in healthcare.

METHOD

We conducted a comprehensive search across databases - Embase, PubMed/Medline, IEEE, and Epistemonikos up to November 2023, using 'Artificial Intelligence' AND 'health professionals' as key domains. We searched for studies that describe the perceptions of healthcare professionals towards AI in healthcare.

FINDINGS

We identified 3931 records. After screening, 25 articles were selected, and 11 were included in the final review. The studies highlight the benefits of AI in healthcare, such as consultation summaries, data management, patient triaging, and referrals, but also raise concerns about job loss, over-reliance, legal implications, and data privacy concerns.

CONCLUSION

AI enhances care delivery efficiency, and concerns arise due to knowledge and experience gaps. Therefore, healthcare workforce education and skill development are crucial for AI adoption, implementation, and future research.

Artificial Intelligence in Atrial Fibrillation: From Early Detection to Precision Therapy

Atrial fibrillation (AF) is the most prevalent cardiac arrhythmia, associated with significant morbidity, mortality, and healthcare burden. Despite advances in AF management, challenges persist in early detection, risk stratification, and treatment optimization, necessitating innovative solutions. Artificial intelligence (AI) has emerged as a transformative tool in AF care, leveraging machine learning and deep learning algorithms to enhance diagnostic accuracy, improve risk prediction, and guide therapeutic interventions. AI-powered electrocardiographic screening has demonstrated the ability to detect asymptomatic AF, while wearable photoplethysmography-based technologies have expanded real-time rhythm monitoring beyond clinical settings. AI-driven predictive models integrate electronic health records and multimodal physiological data to refine AF risk stratification, stroke prediction, and anticoagulation decision making. In the realm of treatment, AI is revolutionizing individualized therapy and optimizing anticoagulation management and catheter ablation strategies. Notably, AI-enhanced electroanatomic mapping and real-time procedural guidance hold promise for improving ablation success rates and reducing AF recurrence. Despite these advancements, the clinical integration of AI in AF management remains an evolving field. Future research should focus on large-scale validation, model interpretability, and regulatory frameworks to ensure widespread adoption. This review explores the current and emerging applications of AI in AF, highlighting its potential to enhance precision medicine and patient outcomes.

Real-time surveillance system for patient deterioration: a pragmatic cluster-randomized controlled trial

The COmmunicating Narrative Concerns Entered by RNs (CONCERN) early warning system (EWS) uses real-time nursing surveillance documentation patterns in its machine learning algorithm to identify deterioration risk. We conducted a 1-year, multisite, pragmatic trial with cluster-randomization of 74 clinical units (37 intervention; 37 usual care) across 2 health systems. Eligible adult hospital encounters were included. We tested if outcomes differed between patients whose care teams were and patients whose care teams were not informed by the CONCERN EWS. Coprimary outcomes were in-hospital mortality (examined as instantaneous risk) and length of stay. Secondary outcomes were cardiopulmonary arrest, sepsis, unanticipated intensive care unit transfers and 30-day hospital readmission. Among 60,893 hospital encounters (33,024 intervention; 27,869 usual care), intervention group encounters had 35.6% decreased instantaneous risk of death (adjusted hazard ratio (HR), 0.64; 95% confidence interval (CI), 0.53-0.78; P < 0.0001), 11.2% decreased length of stay (adjusted incidence rate ratio, 0.91; 95% CI, 0.90-0.93; P < 0.0001), 7.5% decreased instantaneous risk of sepsis (adjusted HR, 0.93; 95% CI, 0.86-0.99; P = 0.0317) and 24.9% increased instantaneous risk of unanticipated intensive care unit transfer (adjusted HR, 1.25; 95% CI, 1.09-1.43; P = 0.0011) compared with usual-care group encounters. No adverse events were reported. A machine learning-based EWS, modeled on nursing surveillance patterns, decreased inpatient deterioration risk with statistical significance. ClinicalTrials.gov registration: NCT03911687 .

Artificial intelligence for individualized treatment of persistent atrial fibrillation: a randomized controlled trial

Although pulmonary vein isolation (PVI) has become the cornerstone ablation procedure for atrial fibrillation (AF), the optimal ablation procedure for persistent and long-standing persistent AF remains elusive. Targeting spatio-temporal electrogram dispersion in a tailored procedure has been suggested as a potentially beneficial alternative to a conventional PVI-only procedure. In this multicenter, randomized, controlled, double-blind, superiority trial, patients with drug-refractory persistent AF were randomly assigned to a tailored ablation procedure targeting areas of spatio-temporal dispersion, as detected by an artificial intelligence (AI) algorithm, in addition to PVI (tailored arm, n = 187, 23% women) or to a conventional PVI-only procedure (anatomical arm, n = 183, 19% women). The primary efficacy endpoint was freedom from documented AF with or without antiarrhythmic drugs at 12 months after a single ablation procedure. Secondary endpoints included freedom from any atrial arrhythmic events, and the secondary composite safety endpoint consisted of death, cerebrovascular events, or treatment-related serious adverse events. One year post-procedure, the trial met its primary efficacy endpoint, which was achieved in 88% of patients in the tailored arm compared with 70% of patients in the anatomical arm (log-rank P < 0.0001 for superiority). However, no significant difference between arms was observed for the freedom from any atrial arrhythmia endpoint after one ablation. The safety endpoint did not differ between arms, with procedure and ablation times being twice as long in the tailored arm. These results show that AI-guided ablation of spatio-temporal dispersion areas in addition to PVI is superior to PVI alone in eliminating AF at 1-year follow-up in patients with persistent and long-standing persistent AF. Ablation of subsequent organized atrial tachycardias may be needed to maintain sinus rhythm long term. ClinicalTrials.gov identifier: NCT04702451 .

Artificial intelligence-based clinical decision support in the emergency department: A scoping review

OBJECTIVE

Artificial intelligence (AI)-based clinical decision support (CDS) has the potential to augment high-stakes clinical decisions in the emergency department (ED). However, its current usage and translation to implementation remains poorly understood. We asked: (1) What is the current landscape of AI-CDS for individual patient care in the ED? and (2) What phases of development have AI-CDS tools achieved?

METHODS

We performed a scoping review of AI for prognostic, diagnostic, and treatment decisions regarding individual ED patient care. We searched five databases (MEDLINE, EMBASE, Cochrane Central, Scopus, Web of Science) and gray literature sources from January 1, 2010, to December 11, 2023. We adhered to guidelines from the Joanna Briggs Institute and PRISMA Extension for Scoping Reviews. We published our protocol on Open Science Framework (DOI 10.17605/OSF.IO/FDZ3Y).

RESULTS

Of 5168 unique records identified, we selected 605 studies for inclusion. The majority (369, 61%) were published in 2021-2023. The studies ranged over a variety of clinical applications, patient populations, and AI model types. Prognostic outcomes were most commonly assessed (270, 44.6%), followed by diagnostic (193, 31.9%) and disposition (115, 19%). Most studies remained in the earliest phase of preclinical development (572, 94.5%) with few advancing to later phases (33, 5.5%).

CONCLUSIONS

By thoroughly mapping the landscape of AI-CDS in the ED, we demonstrate a rapidly increasing volume of studies covering a breadth of clinical applications, yet few have achieved advanced phases of testing or implementation. A more granular understanding of the barriers and facilitators to implementing AI-CDS in the ED is needed.

Computer-assisted prescription of erythropoiesis-stimulating agents in patients undergoing maintenance hemodialysis: a randomized control trial for artificial intelligence model selection

Objective

Machine learning (ML) algorithms are promising tools for managing anemia in hemodialysis (HD) patients. However, their efficacy in predicting erythropoiesis-stimulating agents (ESAs) doses remains uncertain. This study aimed to evaluate the effectiveness of a contemporary artificial intelligence (AI) model in prescribing ESA doses compared to physicians for HD patients.

Materials and Methods

This double-blinded control trial randomized participants into traditional doctor (Dr) and AI groups. In the Dr group, doses of ESA were determined by following clinical guideline recommendations, while in the AI group, they were predicted by the developed models named Random effects (REEM) trees, Mixed-effect random forest (MERF), Long short-term memory (LSTM) networks-I, and LSTM-II. The primary outcome was the capability to maintain patients' hemoglobin (Hb) value near 11 g/dL with a margin of 0.25 g/dL after treating the suggested ESA, with the secondary outcome being Hb value between 10 and 12 g/dL.

Results

A total of 124 participants were enrolled, with 104 completing the study. The mean Hb values were 10.8 and 10.9 g/dL in the AI and Dr groups, respectively, with 69.7% and 73.5% of participants in the respective groups maintaining Hb levels between 10 and 12 g/dL. Only the REEM trees model passed the non-inferiority test for the primary outcome with a margin of 0.25 g/dL and the secondary outcome with a margin of 15%. There was no difference in severe adverse events between the 2 groups.

Conclusion

The REEM trees AI model demonstrated non-inferiority to physicians in prescribing ESA doses for HD patients, maintaining Hb levels within the therapeutic target.

ClinicalTrialsgov Identifier

NCT04185519.

Using machine learning and single nucleotide polymorphisms for improving rheumatoid arthritis risk Prediction in postmenopausal women

Genetic factors contribute to 60-70% of the variability in rheumatoid arthritis (RA). However, few studies have used genetic variants to predict RA risk. This study aimed to enhance RA risk prediction by leveraging single nucleotide polymorphisms (SNPs) through machine-learning algorithms, utilizing Women's Health Initiative data. We developed four predictive models: 1) based on common RA risk factors, 2) model 1 incorporating polygenic risk scores (PRS) with principal components, 3) model 1 and SNPs after feature reduction, and 4) model 1 and SNPs with kernel principal component analysis. Each model was assessed using logistic regression (LR), random forest (RF), eXtreme Gradient Boosting (XGBoost), and support vector machine (SVM). Performance metrics included the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive and negative predictive values (PPV and NPV), and F1-score. The fourth model, integrating SNPs with XGBoost, outperformed all other models. In addition, the XGBoost model that combines genomic data with conventional phenotypic predictors significantly enhanced predictive accuracy, achieving the highest AUC of 0.90 and an F1 score of 0.83. The DeLong test confirmed significant differences in AUC between this model and the others (p-values < 0.0001), particularly highlighting its efficacy in utilizing complex genetic information. These findings emphasize the advantage of combining in-depth genomic data with advanced machine learning for RA risk prediction. The most robust performance of the XGBoost model, which integrated both conventional risk factors and individual SNPs, demonstrates its potential as a tool in personalized medicine for complex diseases like RA. This approach offers a more nuanced and effective RA risk assessment strategy, underscoring the need for further studies to extend broader applications.

Bridging the Gap: From AI Success in Clinical Trials to Real-World Healthcare Implementation-A Narrative Review

BACKGROUND

Artificial intelligence (AI) has demonstrated remarkable diagnostic accuracy in controlled clinical trials, sometimes rivaling or even surpassing experienced clinicians. However, AI's real-world effectiveness is frequently diminished when applied to diverse clinical settings, owing to methodological shortcomings, limited multicenter studies, and insufficient real-world validations.

OBJECTIVE

This narrative review critically examines the discrepancy between AI's robust performance in clinical trials and its inconsistent real-world implementation. Our goal is to synthesize methodological, ethical, and operational challenges impeding AI integration and propose a comprehensive framework to bridge this gap.

METHODS

We conducted a thematic synthesis of peer-reviewed studies from the PubMed, IEEE Xplore, and Scopus databases, targeting studies from 2014 to 2024. Included studies addressed diagnostic, therapeutic, or operational AI applications and related implementation challenges in healthcare. Non-peer-reviewed articles and studies without rigorous analysis were excluded.

RESULTS

Our synthesis identified key barriers to AI's real-world deployment, including algorithmic bias from homogeneous datasets, workflow misalignment, increased clinician workload, and ethical concerns surrounding transparency, accountability, and data privacy. Additionally, scalability remains a challenge due to interoperability issues, insufficient methodological rigor, and inconsistent reporting standards. To address these challenges, we introduce the AI Healthcare Integration Framework (AI-HIF), a structured model incorporating theoretical and operational strategies for responsible AI implementation in healthcare.

CONCLUSIONS

Translating AI from controlled environments to real-world clinical practice necessitates a multifaceted, interdisciplinary approach. Future research should prioritize large-scale pragmatic trials and observational studies to empirically validate the proposed AI Healthcare Integration Framework (AI-HIF) in diverse, real-world healthcare contexts.

Precision management in chronic disease: An AI empowered perspective on medicine-engineering crossover

Precision management of chronic diseases is crucial for improving patient quality of life and alleviating global health burdens. Advancements at the intersection of medicine and engineering, particularly through artificial intelligence (AI), have driven significant progress in precision care. From the perspective of the full life span management of chronic diseases, we focus on medicine-engineering crossover for monitoring chronic diseases, developing and implementing precision care plans, and evaluating care outcomes. Through an in-depth discussion, we address key issues such as AI's potential to enable precision care and the challenges associated with its implementation, including data accuracy, privacy concerns, and clinical adoption. Emphasizing the importance of nurses embracing new technologies and interdisciplinary collaboration, this paper highlights how technological innovation can improve chronic disease management, particularly by enhancing care efficiency and personalizing health interventions. We aim to support the development of integrated healthcare solutions that improve patient outcomes in chronic disease management.

Global output of clinical application research on artificial intelligence in the past decade: a scientometric study and science mapping

BACKGROUND

Artificial intelligence (AI) has shown immense potential in the field of medicine, but its actual effectiveness and safety still need to be validated through clinical trials. Currently, the research themes, methodologies, and development trends of AI-related clinical trials remain unclear, and further exploration of these studies will be crucial for uncovering AI's practical application potential and promoting its broader adoption in clinical settings.

OBJECTIVE

To analyze the current status, hotspots, and trends of published clinical research on AI applications.

METHODS

Publications related to AI clinical applications were retrieved from the Web of Science database. Relevant data were extracted using VOSviewer 1.6.17 to generate visual cooperation network maps for countries, organizations, authors, and keywords. Burst citation detection for keywords and citations was performed using CiteSpace 5.8.R3 to identify sudden surges in citation frequency within a short period, and the theme evolution was analyzed using SciMAT to track the development and trends of research topics over time.

RESULTS

A total of 22,583 articles were obtained from the Web of Science database. Seven-hundred and thirty-five AI clinical application research were published by 1764 institutions from 53 countries. The majority of publications were contributed by the United States, China, and the UK. Active collaborations were noted among leading authors, particularly those from developed countries. The publications mainly focused on evaluating the application value of AI technology in the fields of disease diagnosis and classification, disease risk prediction and management, assisted surgery, and rehabilitation. Deep learning and chatbot technologies were identified as emerging research hotspots in recent studies on AI applications.

CONCLUSIONS

A total of 735 articles on AI in clinical research were analyzed, with publication volume and citation counts steadily increasing each year. Institutions and researchers from the United States contributed the most to the research output in this field. Key areas of focus included AI applications in surgery, rehabilitation, disease diagnosis, risk prediction, and health management, with emerging trends in deep learning and chatbots. This study also provides detailed and intuitive information about important articles, journals, core authors, institutions, and topics in the field through visualization maps, which will help researchers quickly understand the current status, hotspots, and trends of artificial intelligence clinical application research. Future clinical trials of artificial intelligence should strengthen scientific design, ethical compliance, and interdisciplinary and international cooperation and pay more attention to its practical clinical value and reliable application in diverse scenarios.

Artificial intelligence in clinical medicine: a state-of-the-art overview of systematic reviews with methodological recommendations for improved reporting

Medicine has become increasingly receptive to the use of artificial intelligence (AI). This overview of systematic reviews (SRs) aims to categorise current evidence about it and identify the current methodological state of the art in the field proposing a classification of AI model (CLASMOD-AI) to improve future reporting. PubMed/MEDLINE, Scopus, Cochrane library, EMBASE and Epistemonikos databases were screened by four blinded reviewers and all SRs that investigated AI tools in clinical medicine were included. 1923 articles were found, and of these, 360 articles were examined via the full-text and 161 SRs met the inclusion criteria. The search strategy, methodological, medical and risk of bias information were extracted. The CLASMOD-AI was based on input, model, data training, and performance metric of AI tools. A considerable increase in the number of SRs was observed in the last five years. The most covered field was oncology accounting for 13.9% of the SRs, with diagnosis as the predominant objective in 44.4% of the cases). The risk of bias was assessed in 49.1% of included SRs, yet only 39.2% of these used tools with specific items to assess AI metrics. This overview highlights the need for improved reporting on AI metrics, particularly regarding the training of AI models and dataset quality, as both are essential for a comprehensive quality assessment and for mitigating the risk of bias using specialized evaluation tools.

Whole Slide Imaging, Artificial Intelligence, and Machine Learning in Pediatric and Perinatal Pathology: Current Status and Future Directions

The integration of artificial intelligence (AI) into healthcare is becoming increasingly mainstream. Leveraging digital technologies, such as AI and deep learning, impacts researchers, clinicians, and industry due to promising performance and clinical potential. Digital pathology is now a proven technology, enabling generation of high-resolution digital images from glass slides (whole slide images; WSI). WSIs facilitates AI-based image analysis to aid pathologists in diagnostic tasks, improve workflow efficiency, and address workforce shortages. Example applications include tumor segmentation, disease classification, detection, quantitation and grading, rare object identification, and outcome prediction. While advancements have occurred, integration of WSI-AI into clinical laboratories faces challenges, including concerns regarding evidence quality, regulatory adaptations, clinical evaluation, and safety considerations. In pediatric and developmental histopathology, adoption of AI could improve diagnostic efficiency, automate routine tasks, and address specific diagnostic challenges unique to the specialty, such as standardizing placental pathology and developmental autopsy findings, as well as mitigating staffing shortages in the subspeciality. Additionally, AI-based tools have potential to mitigate medicolegal implications by enhancing reproducibility and objectivity in diagnostic evaluations. An overview of recent developments and challenges in applying AI to pediatric and developmental pathology, focusing on machine learning methods applied to WSIs of pediatric pathology specimens is presented.

"Be Really Careful about That": Clinicians' Perceptions of an Intelligence Augmentation Tool for In-Hospital Deterioration Detection

OBJECTIVE

 This study aimed to explore clinicians' perceptions and preferences of prototype intelligence augmentation (IA)-based visualization displays of in-hospital deterioration risk scores to inform future user interface design and implementation in clinical care.

METHODS

 Prototype visualization displays incorporating an IA-based early warning score (EWS) for in-hospital deterioration were developed using cognitive theory and user-centered design principles. The displays featured variations of EWS and clinical data arranged in multipatient and single-patient views. Physician and nurse participants with at least 5 years of clinical experience were recruited to participate in semistructured qualitative interviews focused on understanding their experiences with IA and thoughts and preferences about the prototype displays. A thematic analysis was performed on these data.

RESULTS

 Six themes were identified: (1) clinicians perceive IA as valuable with some caveats related to function and context; (2) individual differences among users influence preferences for customizability; (3) EWS are particularly useful for patient triage; (4) need for patient-centered contextual information to complement EWS; (5) perspectives related to understanding the EWS composition; and (6) design preferences that focus on clarity for interpretation of information.

CONCLUSION

 This study demonstrates clinicians' interest in and reservations about IA tools for clinical deterioration. The findings underscore the importance of understanding clinicians' cognitive needs and framing IA-generated tools as complementary to support them. A clinician focuses on high-level pattern matching information, and clinician's comments related to the power of consistency with typical views (e.g., this is "how I usually see things"), and questions regarding support of score interpretation (e.g., age of the data, questions about what the model "knows") suggest some of the challenges of IA implementation. The findings also identify design implications including the need for contextualizing the EWS for the patient's specific situation, incorporating trend information, and explaining the display purpose for clinical use.

Machine Learning vs Traditional Approaches to Predict All-Cause Mortality for Acute Coronary Syndrome: A Systematic Review and Meta-analysis

BACKGROUND

Acute coronary syndrome (ACS) remains one of the leading causes of death globally. Accurate and reliable mortality risk prediction of ACS patients is essential for developing targeted treatment strategies and improve prognostication. Traditional models for risk stratification such as the GRACE and TIMI risk scores offer moderate discriminative value, and do not incorporate contemporary predictors of ACS prognosis. Machine learning (ML) models have emerged as an alternate method that may offer improved risk assessment. This review compares ML models with traditional risk scores for predicting all-cause mortality in patients with ACS.

METHODS

PubMed, Embase, Web of Science, Cochrane, CINAHL, Scopus, and IEEE XPlore databases were searched through October 30, 2024, as well as Google Scholar and manual screening of reference lists from included studies and the grey literature for studies comparing ML models with traditional statistical methods for event prediction of ACS patients. The primary outcome was comparative discrimination measured by C-statistics with 95% confidence intervals (CIs) in estimating risk of all-cause mortality.

RESULTS

Twelve studies were included (250,510 patients). The summary C-statistic of best-performing ML models across all end points was 0.88 (95% CI 0.86-0.91), compared with 0.82 (95% CI 0.80-0.85) for traditional methods. The difference in C-statistic between ML models and traditional methods was 0.06 (P < 0.0007). Five studies undertook external validation. The PROBAST tool demonstrated high risk of bias for all studies. Common sources of bias included reporting bias and selection bias. Best-performing ML models demonstrated superior discrimination of all-cause mortality for ACS patients compared with traditional risk scores.

CONCLUSIONS

Despite outperforming well established prognostic tools such as the GRACE and TIMI scores, current clinical applications of ML approaches remain uncertain, particularly in view of the need for greater model validation.

Validation and Recalibration of a Model for Predicting Surgical-Site Infection After Pelvic Organ Prolapse Surgery

INTRODUCTION AND HYPOTHESIS

The objective was to externally validate and recalibrate a previously developed model for predicting postoperative surgical-site infection (SSI) after pelvic organ prolapse (POP) surgery.

METHODS

This study utilized a previously validated model for predicting post-POP surgery SSI within 90 days of surgery using a Medicare population. For this study, the model was externally validated and recalibrated using the Premier Healthcare Database (PHD) and the National Surgical Quality Improvement Project (NSQIP) database. Discriminatory performance was assessed via the c-statistic and calibration was assessed using calibration curves. Methods of recalibration in the large and logistic recalibration were used to update the models.

RESULTS

The PHD contained 420,277 POP procedures meeting the inclusion criteria and 1.6% resulted in SSI. The NSQIP dataset contained 62,553 POP surgeries and 1.4% resulted in SSI. Discrimination of the original model was comparable with that seen in the initial validation (c-statistic = 0.57 in PHD, 0.59 in NSQIP vs 0.60 in the original Medicare data). Recalibration greatly improved model calibration when evaluated in NSQIP data.

CONCLUSION

A previously developed model for predicting SSI after POP surgery demonstrated stable discriminatory ability when externally validated on the PHD and NSQIP databases. Model recalibration was necessary to improve prediction. Prospective studies are needed to validate the clinical utility of such a model.

Artificial Intelligence in Pancreatic Imaging: A Systematic Review

The rising incidence of pancreatic diseases, including acute and chronic pancreatitis and various pancreatic neoplasms, poses a significant global health challenge. Pancreatic ductal adenocarcinoma (PDAC) for example, has a high mortality rate due to late-stage diagnosis and its inaccessible location. Advances in imaging technologies, though improving diagnostic capabilities, still necessitate biopsy confirmation. Artificial intelligence, particularly machine learning and deep learning, has emerged as a revolutionary force in healthcare, enhancing diagnostic precision and personalizing treatment. This narrative review explores Artificial intelligence's role in pancreatic imaging, its technological advancements, clinical applications, and associated challenges. Following the PRISMA-DTA guidelines, a comprehensive search of databases including PubMed, Scopus, and Cochrane Library was conducted, focusing on Artificial intelligence, machine learning, deep learning, and radiomics in pancreatic imaging. Articles involving human subjects, written in English, and published up to March 31, 2024, were included. The review process involved title and abstract screening, followed by full-text review and refinement based on relevance and novelty. Recent Artificial intelligence advancements have shown promise in detecting and diagnosing pancreatic diseases. Deep learning techniques, particularly convolutional neural networks (CNNs), have been effective in detecting and segmenting pancreatic tissues as well as differentiating between benign and malignant lesions. Deep learning algorithms have also been used to predict survival time, recurrence risk, and therapy response in pancreatic cancer patients. Radiomics approaches, extracting quantitative features from imaging modalities such as CT, MRI, and endoscopic ultrasound, have enhanced the accuracy of these deep learning models. Despite the potential of Artificial intelligence in pancreatic imaging, challenges such as legal and ethical considerations, algorithm transparency, and data security remain. This review underscores the transformative potential of Artificial intelligence in enhancing the diagnosis and treatment of pancreatic diseases, ultimately aiming to improve patient outcomes and survival rates.

Replication study of PD-L1 status prediction in NSCLC using PET/CT radiomics

This study investigates the predictive capability of radiomics in determining programmed cell death ligand 1 (PD-L1) expression (>=1%) status in non-small cell lung cancer (NSCLC) patients using a newly collected [18F]FDG PET/CT dataset. We aimed to replicate and validate the radiomics-based machine learning (ML) model proposed by Zhao et al. [1] predicting PD-L1 status from PET/CT-imaging. An independent cohort of 254 NSCLC patients underwent [18F]FDG PET/CT imaging, with primary tumor segmentation conducted using lung tissue window (LTW) and more conservative soft tissue window (STW) methods. Radiomics models ("Rad-score" and "complex model") and a clinical-stage model from Zhao et al. were evaluated via 10-fold cross-validation and AUC analysis, alongside a benchmark-study comparing different ML-model pipelines. Clinicopathological data were collected from medical records. On our data, the Rad-score model yielded mean AUCs of 0.593 (STW) and 0.573 (LTW), below Zhao et al.'s 0.761. The complex model achieved mean AUCs of 0.505 (STW) and 0.519 (LTW), lower than Zhao et al.'s 0.769. The clinical model showed a mean AUC of 0.555, below Zhao et al.'s 0.64. All models performed significantly lower than Zhao et al.'s findings. Our benchmark study on four ML pipelines revealed consistently low performance across all configurations. Our study failed to replicate original findings, suggesting poor model performance and questioning predictive value of radiomics features in classifying PD-L1 expression from PET/CT imaging. These results highlight challenges in replicating radiomics-based ML models and stress the need for rigorous validation.

Pathology Foundation Models

Pathology plays a crucial role in diagnosing and evaluating patient tissue samples obtained via surgeries and biopsies. The advent of whole slide scanners and the development of deep learning technologies have considerably advanced this field, promoting extensive research and development in pathology artificial intelligence (AI). These advancements have contributed to reduced workload of pathologists and supported decision-making in treatment plans. Large-scale AI models, known as foundation models (FMs), are more accurate and applicable to various tasks than traditional AI. Such models have recently emerged and expanded their application scope in healthcare. Numerous FMs have been developed in pathology, with reported applications in various tasks, such as disease and rare cancer diagnoses, patient survival prognosis prediction, biomarker expression prediction, and scoring of the immunohistochemical expression intensity. However, several challenges persist in the clinical application of FMs, which healthcare professionals, as users, must be aware of. Research to address these challenges is ongoing. In the future, the development of generalist medical AI, which integrates pathology FMs with FMs from other medical domains, is expected to progress, effectively utilizing AI in real clinical settings to promote precision and personalized medicine.

Benefits and harms associated with the use of AI-related algorithmic decision-making systems by healthcare professionals: a systematic review

Background

Despite notable advancements in artificial intelligence (AI) that enable complex systems to perform certain tasks more accurately than medical experts, the impact on patient-relevant outcomes remains uncertain. To address this gap, this systematic review assesses the benefits and harms associated with AI-related algorithmic decision-making (ADM) systems used by healthcare professionals, compared to standard care.

Methods

In accordance with the PRISMA guidelines, we included interventional and observational studies published as peer-reviewed full-text articles that met the following criteria: human patients; interventions involving algorithmic decision-making systems, developed with and/or utilizing machine learning (ML); and outcomes describing patient-relevant benefits and harms that directly affect health and quality of life, such as mortality and morbidity. Studies that did not undergo preregistration, lacked a standard-of-care control, or pertained to systems that assist in the execution of actions (e.g., in robotics) were excluded. We searched MEDLINE, EMBASE, IEEE Xplore, and Google Scholar for studies published in the past decade up to 31 March 2024. We assessed risk of bias using Cochrane's RoB 2 and ROBINS-I tools, and reporting transparency with CONSORT-AI and TRIPOD-AI. Two researchers independently managed the processes and resolved conflicts through discussion. This review has been registered with PROSPERO (CRD42023412156) and the study protocol has been published.

Findings

Out of 2,582 records identified after deduplication, 18 randomized controlled trials (RCTs) and one cohort study met the inclusion criteria, covering specialties such as psychiatry, oncology, and internal medicine. Collectively, the studies included a median of 243 patients (IQR 124-828), with a median of 50.5% female participants (range 12.5-79.0, IQR 43.6-53.6) across intervention and control groups. Four studies were classified as having low risk of bias, seven showed some concerns, and another seven were assessed as having high or serious risk of bias. Reporting transparency varied considerably: six studies showed high compliance, four moderate, and five low compliance with CONSORT-AI or TRIPOD-AI. Twelve studies (63%) reported patient-relevant benefits. Of those with low risk of bias, interventions reduced length of stay in hospital and intensive care unit (10.3 vs. 13.0 days, p = 0.042; 6.3 vs. 8.4 days, p = 0.030), in-hospital mortality (9.0% vs. 21.3%, p = 0.018), and depression symptoms in non-complex cases (45.1% vs. 52.3%, p = 0.03). However, harms were frequently underreported, with only eight studies (42%) documenting adverse events. No study reported an increase in adverse events as a result of the interventions.

Interpretation

The current evidence on AI-related ADM systems provides limited insights into patient-relevant outcomes. Our findings underscore the essential need for rigorous evaluations of clinical benefits, reinforced compliance with methodological standards, and balanced consideration of both benefits and harms to ensure meaningful integration into healthcare practice.

Funding

This study did not receive any funding.

Machine learning approaches for risk prediction after percutaneous coronary intervention: a systematic review and meta-analysis

Aims

Accurate prediction of clinical outcomes following percutaneous coronary intervention (PCI) is essential for mitigating risk and peri-procedural planning. Traditional risk models have demonstrated a modest predictive value. Machine learning (ML) models offer an alternative risk stratification that may provide improved predictive accuracy.

Methods and results

This study was reported according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses, Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies and Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis guidelines. PubMed, EMBASE, Web of Science, and Cochrane databases were searched until 1 November 2023 for studies comparing ML models with traditional statistical methods for event prediction after PCI. The primary outcome was comparative discrimination measured by C-statistics with 95% confidence intervals (CIs) between ML models and traditional methods in estimating the risk of all-cause mortality, major bleeding, and the composite outcome major adverse cardiovascular events (MACE). Thirty-four models were included across 13 observational studies (4 105 916 patients). For all-cause mortality, the pooled C-statistic for top-performing ML models was 0.89 (95%CI, 0.84-0.91), compared with 0.86 (95% CI, 0.80-0.93) for traditional methods (P = 0.54). For major bleeding, the pooled C-statistic for ML models was 0.80 (95% CI, 0.77-0.84), compared with 0.78 (95% CI, 0.77-0.79) for traditional methods (P = 0.02). For MACE, the C-statistic for ML models was 0.83 (95% CI, 0.75-0.91), compared with 0.71 (95% CI, 0.69-0.74) for traditional methods (P = 0.007). Out of all included models, only one model was externally validated. Calibration was inconsistently reported across all models. Prediction Model Risk of Bias Assessment Tool demonstrated a high risk of bias across all studies.

Conclusion

Machine learning models marginally outperformed traditional risk scores in the discrimination of MACE and major bleeding following PCI. While integration of ML algorithms into electronic healthcare systems has been hypothesized to improve peri-procedural risk stratification, immediate implementation in the clinical setting remains uncertain. Further research is required to overcome methodological and validation limitations.

QUAIDE - Quality assessment of AI preclinical studies in diagnostic endoscopy

Artificial intelligence (AI) holds significant potential for enhancing quality of gastrointestinal (GI) endoscopy, but the adoption of AI in clinical practice is hampered by the lack of rigorous standardisation and development methodology ensuring generalisability. The aim of the Quality Assessment of pre-clinical AI studies in Diagnostic Endoscopy (QUAIDE) Explanation and Checklist was to develop recommendations for standardised design and reporting of preclinical AI studies in GI endoscopy.The recommendations were developed based on a formal consensus approach with an international multidisciplinary panel of 32 experts among endoscopists and computer scientists. The Delphi methodology was employed to achieve consensus on statements, with a predetermined threshold of 80% agreement. A maximum three rounds of voting were permitted.Consensus was reached on 18 key recommendations, covering 6 key domains: data acquisition and annotation (6 statements), outcome reporting (3 statements), experimental setup and algorithm architecture (4 statements) and result presentation and interpretation (5 statements). QUAIDE provides recommendations on how to properly design (1. Methods, statements 1-14), present results (2. Results, statements 15-16) and integrate and interpret the obtained results (3. Discussion, statements 17-18).The QUAIDE framework offers practical guidance for authors, readers, editors and reviewers involved in AI preclinical studies in GI endoscopy, aiming at improving design and reporting, thereby promoting research standardisation and accelerating the translation of AI innovations into clinical practice.

ESR Essentials: how to get to valuable radiology AI: the role of early health technology assessment-practice recommendations by the European Society of Medical Imaging Informatics

AI tools in radiology are revolutionising the diagnosis, evaluation, and management of patients. However, there is a major gap between the large number of developed AI tools and those translated into daily clinical practice, which can be primarily attributed to limited usefulness and trust in current AI tools. Instead of technically driven development, little effort has been put into value-based development to ensure AI tools will have a clinically relevant impact on patient care. An iterative comprehensive value evaluation process covering the complete AI tool lifecycle should be part of radiology AI development. For value assessment of health technologies, health technology assessment (HTA) is an extensively used and comprehensive method. While most aspects of value covered by HTA apply to radiology AI, additional aspects, including transparency, explainability, and robustness, are unique to radiology AI and crucial in its value assessment. Additionally, value assessment should already be included early in the design stage to determine the potential impact and subsequent requirements of the AI tool. Such early assessment should be systematic, transparent, and practical to ensure all stakeholders and value aspects are considered. Hence, early value-based development by incorporating early HTA will lead to more valuable AI tools and thus facilitate translation to clinical practice. CLINICAL RELEVANCE STATEMENT: This paper advocates for the use of early value-based assessments. These assessments promote a comprehensive evaluation on how an AI tool in development can provide value in clinical practice and thus help improve the quality of these tools and the clinical process they support. KEY POINTS: Value in radiology AI should be perceived as a comprehensive term including health technology assessment domains and AI-specific domains. Incorporation of an early health technology assessment for radiology AI during development will lead to more valuable radiology AI tools. Comprehensive and transparent value assessment of radiology AI tools is essential for their widespread adoption.

Perceptions and attitudes toward artificial intelligence among frontline physicians and physicians' assistants in Kansas: a cross-sectional survey

Objective

This survey aims to understand frontline healthcare professionals' perceptions of artificial intelligence (AI) in healthcare and assess how AI familiarity influences these perceptions.

Materials and Methods

We conducted a survey from February to March 2023 of physicians and physician assistants registered with the Kansas State Board of Healing Arts. Participants rated their perceptions toward AI-related domains and constructs on a 5-point Likert scale, with higher scores indicating stronger agreement. Two sub-groups were created for analysis to assess the impact of participants' familiarity and experience with AI on the survey results.

Results

From 532 respondents, key concerns were Perceived Communication Barriers (median = 4.0, IQR = 2.8-4.8), Unregulated Standards (median = 4.0, IQR = 3.6-4.8), and Liability Issues (median = 4.0, IQR = 3.5-4.8). Lower levels of agreement were noted for Trust in AI Mechanisms (median = 3.0, IQR = 2.2-3.4), Perceived Risks of AI (median = 3.2, IQR = 2.6-4.0), and Privacy Concerns (median = 3.3, IQR = 2.3-4.0). Positive correlations existed between Intention to use AI and Perceived Benefits (r = 0.825) and Trust in AI Mechanisms (r = 0.777). Perceived risk negatively correlated with Intention to Use AI (r = -0.718). There was no difference in perceptions between AI experienced and AI naïve subgroups.

Discussion

The findings suggest that perceptions of benefits, trust, risks, communication barriers, regulation, and liability issues influence healthcare professionals' intention to use AI, regardless of their AI familiarity.

Conclusion

The study highlights key factors affecting AI adoption in healthcare from the frontline healthcare professionals' perspective. These insights can guide strategies for successful AI implementation in healthcare.

Artificial Intelligence-Assisted Colonoscopy for Polyp Detection : A Systematic Review and Meta-analysis

BACKGROUND

Randomized clinical trials (RCTs) of computer-aided detection (CADe) system-enhanced colonoscopy compared with conventional colonoscopy suggest increased adenoma detection rate (ADR) and decreased adenoma miss rate (AMR), but the effect on detection of advanced colorectal neoplasia (ACN) is unclear.

PURPOSE

To conduct a systematic review to compare performance of CADe-enhanced and conventional colonoscopy.

DATA SOURCES

Cochrane Library, Google Scholar, Ovid EMBASE, Ovid MEDLINE, PubMed, Scopus, and Web of Science Core Collection databases were searched through February 2024.

STUDY SELECTION

Published RCTs comparing CADe-enhanced and conventional colonoscopy.

DATA EXTRACTION

Average adenoma per colonoscopy (APC) and ACN per colonoscopy were primary outcomes. Adenoma detection rate, AMR, and ACN detection rate (ACN DR) were secondary outcomes. Balancing outcomes included withdrawal time and resection of nonneoplastic polyps (NNPs). Subgroup analyses were done by neural network architecture.

DATA SYNTHESIS

Forty-four RCTs with 36 201 cases were included. Computer-aided detection-enhanced colonoscopies have higher average APC (12 090 of 12 279 [0.98] vs. 9690 of 12 292 [0.78], incidence rate difference [IRD] = 0.22 [95% CI, 0.16 to 0.28]) and higher ADR (7098 of 16 253 [44.7%] vs. 5825 of 15 855 [36.7%], rate ratio [RR] = 1.21 [CI, 1.15 to 1.28]). Average ACN per colonoscopy was similar (1512 of 9296 [0.16] vs. 1392 of 9121 [0.15], IRD = 0.01 [CI, -0.01 to 0.02]), but ACN DR was higher with CADe system use (1260 of 9899 [12.7%] vs. 1119 of 9746 [11.5%], RR = 1.16 [CI, 1.02 to 1.32]). Using CADe systems resulted in resection of almost 2 extra NNPs per 10 colonoscopies and longer total withdrawal time (0.53 minutes [CI, 0.30 to 0.77]).

LIMITATION

Statistically significant heterogeneity in quality and sample size and inability to blind endoscopists to the intervention in included studies may affect the performance estimates.

CONCLUSION

Computer-aided detection-enhanced colonoscopies have increased APC and detection rate but no difference in ACN per colonoscopy and a small increase in ACN DR. There is minimal increase in procedure time and no difference in performance across neural network architectures.

PRIMARY FUNDING SOURCE

None. (PROSPERO: CRD42023422835).

Machine learning-augmented interventions in perioperative care: a systematic review and meta-analysis

BACKGROUND

We lack evidence on the cumulative effectiveness of machine learning (ML)-driven interventions in perioperative settings. Therefore, we conducted a systematic review to appraise the evidence on the impact of ML-driven interventions on perioperative outcomes.

METHODS

Ovid MEDLINE, CINAHL, Embase, Scopus, PubMed, and ClinicalTrials.gov were searched to identify randomised controlled trials (RCTs) evaluating the effectiveness of ML-driven interventions in surgical inpatient populations. The review was registered with PROSPERO (CRD42023433163) and conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. Meta-analysis was conducted for outcomes with two or more studies using a random-effects model, and vote counting was conducted for other outcomes.

RESULTS

Among 13 included RCTs, three types of ML-driven interventions were evaluated: Hypotension Prediction Index (HPI) (n=5), Nociception Level Index (NoL) (n=7), and a scheduling system (n=1). Compared with the standard care, HPI led to a significant decrease in absolute hypotension (n=421, P=0.003, I2=75%) and relative hypotension (n=208, P<0.0001, I2=0%); NoL led to significantly lower mean pain scores in the post-anaesthesia care unit (PACU) (n=191, P=0.004, I2=19%). NoL showed no significant impact on intraoperative opioid consumption (n=339, P=0.31, I2=92%) or PACU opioid consumption (n=339, P=0.11, I2=0%). No significant difference in hospital length of stay (n=361, P=0.81, I2=0%) and PACU stay (n=267, P=0.44, I2=0) was found between HPI and NoL.

CONCLUSIONS

HPI decreased the duration of intraoperative hypotension, and NoL decreased postoperative pain scores, but no significant impact on other clinical outcomes was found. We highlight the need to address both methodological and clinical practice gaps to ensure the successful future implementation of ML-driven interventions.

SYSTEMATIC REVIEW PROTOCOL

CRD42023433163 (PROSPERO).

Applications of Machine Learning on Electronic Health Record Data to Combat Antibiotic Resistance

There is growing excitement about the clinical use of artificial intelligence and machine learning (ML) technologies. Advancements in computing and the accessibility of ML frameworks enable researchers to easily train predictive models using electronic health record data. However, several practical factors must be considered when employing ML on electronic health record data. We provide a primer on ML and approaches commonly taken to address these challenges. To illustrate how these approaches have been applied to address antimicrobial resistance, we review the use of electronic health record data to construct ML models for predicting pathogen carriage or infection, optimizing empiric therapy, and aiding antimicrobial stewardship tasks. ML shows promise in promoting the appropriate use of antimicrobials, although clinical deployment is limited. We conclude by describing the potential dangers of, and barriers to, implementation of ML models in the clinic.

Artificial Intelligence and Radiologist Burnout

IMPORTANCE

Understanding the association of artificial intelligence (AI) with physician burnout is crucial for fostering a collaborative interactive environment between physicians and AI.

OBJECTIVE

To estimate the association between AI use in radiology and radiologist burnout.

DESIGN, SETTING, AND PARTICIPANTS

This cross-sectional study conducted a questionnaire survey between May and October 2023, using the national quality control system of radiology in China. Participants included radiologists from 1143 hospitals. Radiologists reporting regular or consistent AI use were categorized as the AI group. Statistical analysis was performed from October 2023 to May 2024.

EXPOSURE

AI use in radiology practice.

MAIN OUTCOMES AND MEASURES

Burnout was defined by emotional exhaustion (EE) or depersonalization according to the Maslach Burnout Inventory. Workload was assessed based on working hours, number of image interpretations, hospital level, device type, and role in the workflow. AI acceptance was determined via latent class analysis considering AI-related knowledge, attitude, confidence, and intention. Propensity score-based mixed-effect generalized linear logistic regression was used to estimate the associations between AI use and burnout and its components. Interactions of AI use, workload, and AI acceptance were assessed on additive and multiplicative scales.

RESULTS

Among 6726 radiologists included in this study, 2376 (35.3%) were female and 4350 (64.7%) were male; the median (IQR) age was 41 (34-48) years; 3017 were in the AI group (1134 [37.6%] female; median [IQR] age, 40 [33-47] years) and 3709 in the non-AI group (1242 [33.5%] female; median [IQR] age, 42 [34-49] years). The weighted prevalence of burnout was significantly higher in the AI group compared with the non-AI group (40.9% vs 38.6%; P < .001). After adjusting for covariates, AI use was significantly associated with increased odds of burnout (odds ratio [OR], 1.20; 95% CI, 1.10-1.30), primarily driven by its association with EE (OR, 1.21; 95% CI, 1.10-1.34). A dose-response association was observed between the frequency of AI use and burnout (P for trend < .001). The associations were more pronounced among radiologists with high workload and lower AI acceptance. A significant negative interaction was noted between high AI acceptance and AI use.

CONCLUSIONS AND RELEVANCE

In this cross-sectional study of radiologist burnout, frequent AI use was associated with an increased risk of radiologist burnout, particularly among those with high workload or lower AI acceptance. Further longitudinal studies are needed to provide more evidence.

Core Concepts in Pharmacoepidemiology: Principled Use of Artificial Intelligence and Machine Learning in Pharmacoepidemiology and Healthcare Research

Artificial intelligence (AI) and machine learning (ML) are important tools across many fields of health and medical research. Pharmacoepidemiologists can bring essential methodological rigor and study design expertise to the design and use of these technologies within healthcare settings. AI/ML-based tools also play a role in pharmacoepidemiology research, as we may apply them to answer our own research questions, take responsibility for evaluating medical devices with AI/ML components, or participate in interdisciplinary research to create new AI/ML algorithms. While epidemiologic expertise is essential to deploying AI/ML responsibly and ethically, the rapid advancement of these technologies in the past decade has resulted in a knowledge gap for many in the field. This article provides a brief overview of core AI/ML concepts, followed by a discussion of potential applications of AI/ML in pharmacoepidemiology research, and closes with a review of important concepts across application areas, including interpretability and fairness. This review is intended to provide an accessible, practical overview of AI/ML for pharmacoepidemiology research, with references to further, more detailed resources on fundamental topics.

Health Care Professionals' Experience of Using AI: Systematic Review With Narrative Synthesis

BACKGROUND

There has been a substantial increase in the development of artificial intelligence (AI) tools for clinical decision support. Historically, these were mostly knowledge-based systems, but recent advances include non-knowledge-based systems using some form of machine learning. The ability of health care professionals to trust technology and understand how it benefits patients or improves care delivery is known to be important for their adoption of that technology. For non-knowledge-based AI tools for clinical decision support, these issues are poorly understood.

OBJECTIVE

The aim of this study is to qualitatively synthesize evidence on the experiences of health care professionals in routinely using non-knowledge-based AI tools to support their clinical decision-making.

METHODS

In June 2023, we searched 4 electronic databases, MEDLINE, Embase, CINAHL, and Web of Science, with no language or date limit. We also contacted relevant experts and searched reference lists of the included studies. We included studies of any design that reported the experiences of health care professionals using non-knowledge-based systems for clinical decision support in their work settings. We completed double independent quality assessment for all included studies using the Mixed Methods Appraisal Tool. We used a theoretically informed thematic approach to synthesize the findings.

RESULTS

After screening 7552 titles and 182 full-text articles, we included 25 studies conducted in 9 different countries. Most of the included studies were qualitative (n=13), and the remaining were quantitative (n=9) and mixed methods (n=3). Overall, we identified 7 themes: health care professionals' understanding of AI applications, level of trust and confidence in AI tools, judging the value added by AI, data availability and limitations of AI, time and competing priorities, concern about governance, and collaboration to facilitate the implementation and use of AI. The most frequently occurring are the first 3 themes. For example, many studies reported that health care professionals were concerned about not understanding the AI outputs or the rationale behind them. There were issues with confidence in the accuracy of the AI applications and their recommendations. Some health care professionals believed that AI provided added value and improved decision-making, and some reported that it only served as a confirmation of their clinical judgment, while others did not find it useful at all.

CONCLUSIONS

Our review identified several important issues documented in various studies on health care professionals' use of AI tools in real-world health care settings. Opinions of health care professionals regarding the added value of AI tools for supporting clinical decision-making varied widely, and many professionals had concerns about their understanding of and trust in this technology. The findings of this review emphasize the need for concerted efforts to optimize the integration of AI tools in real-world health care settings.

TRIAL REGISTRATION

PROSPERO CRD42022336359; https://tinyurl.com/2yunvkmb.

A review on the efficacy of artificial intelligence for managing anxiety disorders

Anxiety disorders are psychiatric conditions characterized by prolonged and generalized anxiety experienced by individuals in response to various events or situations. At present, anxiety disorders are regarded as the most widespread psychiatric disorders globally. Medication and different types of psychotherapies are employed as the primary therapeutic modalities in clinical practice for the treatment of anxiety disorders. However, combining these two approaches is known to yield more significant benefits than medication alone. Nevertheless, there is a lack of resources and a limited availability of psychotherapy options in underdeveloped areas. Psychotherapy methods encompass relaxation techniques, controlled breathing exercises, visualization exercises, controlled exposure exercises, and cognitive interventions such as challenging negative thoughts. These methods are vital in the treatment of anxiety disorders, but executing them proficiently can be demanding. Moreover, individuals with distinct anxiety disorders are prescribed medications that may cause withdrawal symptoms in some instances. Additionally, there is inadequate availability of face-to-face psychotherapy and a restricted capacity to predict and monitor the health, behavioral, and environmental aspects of individuals with anxiety disorders during the initial phases. In recent years, there has been notable progress in developing and utilizing artificial intelligence (AI) based applications and environments to improve the precision and sensitivity of diagnosing and treating various categories of anxiety disorders. As a result, this study aims to establish the efficacy of AI-enabled environments in addressing the existing challenges in managing anxiety disorders, reducing reliance on medication, and investigating the potential advantages, issues, and opportunities of integrating AI-assisted healthcare for anxiety disorders and enabling personalized therapy.

Toward real-world deployment of machine learning for health care: External validation, continual monitoring, and randomized clinical trials

In this commentary, we elucidate three indispensable evaluation steps toward the real-world deployment of machine learning within the healthcare sector and demonstrate referable examples for diagnostic, therapeutic, and prognostic tasks. We encourage researchers to move beyond retrospective and within-sample validation, and step into the practical implementation at the bedside rather than leaving developed machine learning models in the dust of archived literature.

Conceptualizing bias in EHR data: A case study in performance disparities by demographic subgroups for a pediatric obesity incidence classifier

Electronic Health Records (EHRs) are increasingly used to develop machine learning models in predictive medicine. There has been limited research on utilizing machine learning methods to predict childhood obesity and related disparities in classifier performance among vulnerable patient subpopulations. In this work, classification models are developed to recognize pediatric obesity using temporal condition patterns obtained from patient EHR data in a U.S. study population. We trained four machine learning algorithms (Logistic Regression, Random Forest, Gradient Boosted Trees, and Neural Networks) to classify cases and controls as obesity positive or negative, and optimized hyperparameter settings through a bootstrapping methodology. To assess the classifiers for bias, we studied model performance by population subgroups then used permutation analysis to identify the most predictive features for each model and the demographic characteristics of patients with these features. Mean AUC-ROC values were consistent across classifiers, ranging from 0.72-0.80. Some evidence of bias was identified, although this was through the models performing better for minority subgroups (African Americans and patients enrolled in Medicaid). Permutation analysis revealed that patients from vulnerable population subgroups were over-represented among patients with the most predictive diagnostic patterns. We hypothesize that our models performed better on under-represented groups because the features more strongly associated with obesity were more commonly observed among minority patients. These findings highlight the complex ways that bias may arise in machine learning models and can be incorporated into future research to develop a thorough analytical approach to identify and mitigate bias that may arise from features and within EHR datasets when developing more equitable models.

Trial Factors Associated With Completion of Clinical Trials Evaluating AI: Retrospective Case-Control Study

BACKGROUND

Evaluation of artificial intelligence (AI) tools in clinical trials remains the gold standard for translation into clinical settings. However, design factors associated with successful trial completion and the common reasons for trial failure are unknown.

OBJECTIVE

This study aims to compare trial design factors of complete and incomplete clinical trials testing AI tools. We conducted a case-control study of complete (n=485) and incomplete (n=51) clinical trials that evaluated AI as an intervention of ClinicalTrials.gov.

METHODS

Trial design factors, including area of clinical application, intended use population, and intended role of AI, were extracted. Trials that did not evaluate AI as an intervention and active trials were excluded. The assessed trial design factors related to AI interventions included the domain of clinical application related to organ systems; intended use population for patients or health care providers; and the role of AI for different applications in patient-facing clinical workflows, such as diagnosis, screening, and treatment. In addition, we also assessed general trial design factors including study type, allocation, intervention model, masking, age, sex, funder, continent, length of time, sample size, number of enrollment sites, and study start year. The main outcome was the completion of the clinical trial. Odds ratio (OR) and 95% CI values were calculated for all trial design factors using propensity-matched, multivariable logistic regression.

RESULTS

We queried ClinicalTrials.gov on December 23, 2023, using AI keywords to identify complete and incomplete trials testing AI technologies as a primary intervention, yielding 485 complete and 51 incomplete trials for inclusion in this study. Our nested propensity-matched, case-control results suggest that trials conducted in Europe were significantly associated with trial completion when compared with North American trials (OR 2.85, 95% CI 1.14-7.10; P=.03), and the trial sample size was positively associated with trial completion (OR 1.00, 95% CI 1.00-1.00; P=.02).

CONCLUSIONS

Our case-control study is one of the first to identify trial design factors associated with completion of AI trials and catalog study-reported reasons for AI trial failure. We observed that trial design factors positively associated with trial completion include trials conducted in Europe and sample size. Given the promising clinical use of AI tools in health care, our results suggest that future translational research should prioritize addressing the design factors of AI clinical trials associated with trial incompletion and common reasons for study failure.

Patient Perspectives on AI for Mental Health Care: Cross-Sectional Survey Study

BACKGROUND

The application of artificial intelligence (AI) to health and health care is rapidly increasing. Several studies have assessed the attitudes of health professionals, but far fewer studies have explored the perspectives of patients or the general public. Studies investigating patient perspectives have focused on somatic issues, including those related to radiology, perinatal health, and general applications. Patient feedback has been elicited in the development of specific mental health care solutions, but broader perspectives toward AI for mental health care have been underexplored.

OBJECTIVE

This study aims to understand public perceptions regarding potential benefits of AI, concerns about AI, comfort with AI accomplishing various tasks, and values related to AI, all pertaining to mental health care.

METHODS

We conducted a 1-time cross-sectional survey with a nationally representative sample of 500 US-based adults. Participants provided structured responses on their perceived benefits, concerns, comfort, and values regarding AI for mental health care. They could also add free-text responses to elaborate on their concerns and values.

RESULTS

A plurality of participants (245/497, 49.3%) believed AI may be beneficial for mental health care, but this perspective differed based on sociodemographic variables (all P<.05). Specifically, Black participants (odds ratio [OR] 1.76, 95% CI 1.03-3.05) and those with lower health literacy (OR 2.16, 95% CI 1.29-3.78) perceived AI to be more beneficial, and women (OR 0.68, 95% CI 0.46-0.99) perceived AI to be less beneficial. Participants endorsed concerns about accuracy, possible unintended consequences such as misdiagnosis, the confidentiality of their information, and the loss of connection with their health professional when AI is used for mental health care. A majority of participants (80.4%, 402/500) valued being able to understand individual factors driving their risk, confidentiality, and autonomy as it pertained to the use of AI for their mental health. When asked who was responsible for the misdiagnosis of mental health conditions using AI, 81.6% (408/500) of participants found the health professional to be responsible. Qualitative results revealed similar concerns related to the accuracy of AI and how its use may impact the confidentiality of patients' information.

CONCLUSIONS

Future work involving the use of AI for mental health care should investigate strategies for conveying the level of AI's accuracy, factors that drive patients' mental health risks, and how data are used confidentially so that patients can determine with their health professionals when AI may be beneficial. It will also be important in a mental health care context to ensure the patient-health professional relationship is preserved when AI is used.

Machine-Learning for Phenotyping and Prognostication of Myocardial Infarction and Injury in Suspected Acute Coronary Syndrome

Background

Clinical work-up for suspected acute coronary syndrome (ACS) is resource intensive.

Objectives

This study aimed to develop a machine learning model for digitally phenotyping myocardial injury and infarction and predict 30-day events in suspected ACS patients.

Methods

Training and testing data sets, predominantly derived from electronic health records, included suspected ACS patients presenting to 6 and 26 South Australian hospitals, respectively. All index presentations and 30-day death and myocardial infarction (MI) were adjudicated using the Fourth Universal Definition of MI. We developed 2 diagnostic prediction models which phenotype myocardial injury and infarction according to the Fourth UDMI (chronic myocardial injury vs acute myocardial injury patterns, the latter further differentiated into acute non-ischaemic myocardial injury, Types 1 and 2 MI) using eXtreme Gradient Boosting (XGB) and deep-learning (DL). We also developed an event prediction model for risk prediction of 30-day death or MI using XGB. Analyses were performed in Python 3.6.

Results

The training and testing data sets had 6,722 and 8,869 participants, respectively. The diagnostic prediction XGB and deep learning models achieved an area under the curve of 99.2% ± 0.1% and 98.8% ± 0.2%, respectively, for differentiating an acute myocardial injury pattern from no injury or chronic myocardial injury pattern and achieved 95.5% ± 0.2% and 94.6% ± 0.9%, respectively, for differentiating type 1 MI from type 2 MI or acute nonischemic myocardial injury. The 30-day death/MI event prediction model achieved an area under the curve of 88.5% ± 0.5%.

Conclusions

Machine learning models can digitally phenotype suspected ACS patients at index presentation and predict subsequent events within 30 days. These models require external validation in a randomized clinical trial to evaluate their impact in clinical practice.

Assessing accuracy of ChatGPT in response to questions from day to day pharmaceutical care in hospitals

Background

The advent of Large Language Models (LLMs) such as ChatGPT introduces opportunities within the medical field. Nonetheless, use of LLM poses a risk when healthcare practitioners and patients present clinical questions to these programs without a comprehensive understanding of its suitability for clinical contexts.

Objective

The objective of this study was to assess ChatGPT's ability to generate appropriate responses to clinical questions that hospital pharmacists could encounter during routine patient care.

Methods

Thirty questions from 10 different domains within clinical pharmacy were collected during routine care. Questions were presented to ChatGPT in a standardized format, including patients' age, sex, drug name, dose, and indication. Subsequently, relevant information regarding specific cases were provided, and the prompt was concluded with the query "what would a hospital pharmacist do?". The impact on accuracy was assessed for each domain by modifying personification to "what would you do?", presenting the question in Dutch, and regenerating the primary question. All responses were independently evaluated by two senior hospital pharmacists, focusing on the availability of an advice, accuracy and concordance.

Results

In 77% of questions, ChatGPT provided an advice in response to the question. For these responses, accuracy and concordance were determined. Accuracy was correct and complete for 26% of responses, correct but incomplete for 22% of responses, partially correct and partially incorrect for 30% of responses and completely incorrect for 22% of responses. The reproducibility was poor, with merely 10% of responses remaining consistent upon regeneration of the primary question.

Conclusions

While concordance of responses was excellent, the accuracy and reproducibility were poor. With the described method, ChatGPT should not be used to address questions encountered by hospital pharmacists during their shifts. However, it is important to acknowledge the limitations of our methodology, including potential biases, which may have influenced the findings.

A novel way to prospectively evaluate of AI-enhanced ECG algorithms

Significant strides will be made in the field of computerized electrocardiology through the development of artificial intelligence (AI)-enhanced ECG (AI-ECG) algorithms. Yet, the scientific discourse has primarily relied upon on retrospective analyses for deriving and externally validating AI-ECG classification algorithms, an approach that fails to fully judge their real-world effectiveness or reveal potential unintended consequences. Prospective trials and analyses of AI-ECG algorithms will be crucial for assessing real-world diagnostic scenarios and understanding their practical utility and degree influence they confer onto clinicians. However, conducting such studies is challenging due to their resource-intensive nature and associated technical and logistical hurdles. To overcome these challenges, we propose an innovative approach to assess AI-ECG algorithms using a virtual testing environment. This strategy can yield critical insights into the practical utility and clinical implications of novel AI-ECG algorithms. Moreover, such an approach can enable an assessment of the influence of AI-ECG algorithms have their users. Herein, we outline a proposed randomized control trial for evaluating the diagnostic efficacy of new AI-ECG algorithm(s) specifically designed to differentiate between wide complex tachycardias into ventricular tachycardia and supraventricular wide complex tachycardia.

Prospective Human Validation of Artificial Intelligence Interventions in Cardiology: A Scoping Review

Background

Despite the potential of artificial intelligence (AI) in enhancing cardiovascular care, its integration into clinical practice is limited by a lack of evidence on its effectiveness with respect to human experts or gold standard practices in real-world settings.

Objectives

The purpose of this study was to identify AI interventions in cardiology that have been prospectively validated against human expert benchmarks or gold standard practices, assessing their effectiveness, and identifying future research areas.

Methods

We systematically reviewed Scopus and MEDLINE to identify peer-reviewed publications that involved prospective human validation of AI-based interventions in cardiology from January 2015 to December 2023.

Results

Of 2,351 initial records, 64 studies were included. Among these studies, 59 (92.2%) were published after 2020. A total of 11 (17.2%) randomized controlled trials were published. AI interventions in 44 articles (68.75%) reported definite clinical or operational improvements over human experts. These interventions were mostly used in imaging (n = 14, 21.9%), ejection fraction (n = 10, 15.6%), arrhythmia (n = 9, 14.1%), and coronary artery disease (n = 12, 18.8%) application areas. Convolutional neural networks were the most common predictive model (n = 44, 69%), and images were the most used data type (n = 38, 54.3%). Only 22 (34.4%) studies made their models or data accessible.

Conclusions

This review identifies the potential of AI in cardiology, with models often performing equally well as human counterparts for specific and clearly scoped tasks suitable for such models. Nonetheless, the limited number of randomized controlled trials emphasizes the need for continued validation, especially in real-world settings that closely examine joint human AI decision-making.

Reaching the ceiling? Empirical scaling behaviour for deep EEG pathology classification

Machine learning techniques, particularly deep convolutional neural networks (ConvNets), are increasingly being used to automate clinical EEG analysis, with the potential to reduce the clinical burden and improve patient care. However, further research is required before they can be used in clinical settings, particularly regarding the impact of the number of training samples and model parameters on their testing error. To address this, we present a comprehensive study of the empirical scaling behaviour of ConvNets for EEG pathology classification. We analysed the testing error with increasing the training samples and model size for four different ConvNet architectures. The focus of our experiments is width scaling, and we have increased the number of parameters to up to 1.8 million. Our evaluation was based on two publicly available datasets: the Temple University Hospital (TUH) Abnormal EEG Corpus and the TUH Abnormal Expansion Balanced EEG Corpus, which together contain 10,707 training samples. The results show that the testing error follows a saturating power-law with both model and dataset size. This pattern is consistent across different datasets and ConvNet architectures. Furthermore, empirically observed accuracies saturate at 85%-87%, which may be due to an imperfect inter-rater agreement on the clinical labels. The empirical scaling behaviour of the test performance with dataset and model size has significant implications for deep EEG pathology classification research and practice. Our findings highlight the potential of deep ConvNets for high-performance EEG pathology classification, and the identified scaling relationships provide valuable recommendations for the advancement of automated EEG diagnostics.

Poor performance of ChatGPT in clinical rule-guided dose interventions in hospitalized patients with renal dysfunction

PURPOSE

Clinical decision support systems (CDSS) are used to identify drugs with potential need for dose modification in patients with renal impairment. ChatGPT holds the potential to be integrated in the electronic health record (EHR) system to give such dosing advices. In this study, we aim to evaluate the performance of ChatGPT in clinical rule-guided dose interventions in hospitalized patients with renal impairment.

METHODS

This cross-sectional study was performed at Tergooi Medical Center, the Netherlands. CDSS alerts regarding renal dysfunction were collected from the electronic health record (EHR) during a 2-week period and were presented to ChatGPT and an expert panel. Alerts were presented with and without patient variables. To evaluate the performance, suggested medication interventions were compared.

RESULTS

In total, 172 CDDS alerts were generated for 80 patients. Indecisive responses by ChatGPT to alerts were excluded. For alerts presented without patient variables, ChatGPT provided "correct and identical" responses to 19.9%, "correct and different" responses to 26.7%, and "incorrect responses to 53.4% of the alerts. For alerts including patient variables, ChatGPT provided "correct and identical" responses to 16.7%, "correct and different" responses to 16.0%, and "incorrect responses to 67.3% of the alerts. Accuracy was better for newer drugs such as direct oral anticoagulants.

CONCLUSION

The performance of ChatGPT in clinical rule-guided dose interventions in hospitalized patients with renal dysfunction was poor. Based on these results, we conclude that ChatGPT, in its current state, is not appropriate for automatic integration into our EHR to handle CDSS alerts related to renal dysfunction.

The AI Future of Emergency Medicine

In the coming years, artificial intelligence (AI) and machine learning will likely give rise to profound changes in the field of emergency medicine, and medicine more broadly. This article discusses these anticipated changes in terms of 3 overlapping yet distinct stages of AI development. It reviews some fundamental concepts in AI and explores their relation to clinical practice, with a focus on emergency medicine. In addition, it describes some of the applications of AI in disease diagnosis, prognosis, and treatment, as well as some of the practical issues that they raise, the barriers to their implementation, and some of the legal and regulatory challenges they create.

Bridging Health Disparities in the Data-Driven World of Artificial Intelligence: A Narrative Review

BACKGROUND

Artificial intelligence (AI) holds exciting potential to revolutionize healthcare delivery in the United States. However, there are concerns about its potential to perpetuate disparities among historically marginalized populations.

OBJECTIVE

Following the guidelines of Preferred Reporting Items for Systematic Reviews and Meta-Analyses, we conducted a narrative review of current literature on AI and health disparities in the United States. We aimed to answer the question, Does AI have the potential to reduce or eliminate health disparities, or will its use further exacerbate these disparities?

METHODS

We searched the Ovid MEDLINE electronic database to identify and retrieve publications discussing AI and its impact on racial/ethnic health disparities. Articles were included if they discussed AI as a tool to mitigate racial health disparities with or without bias in developing and using AI.

RESULTS

This review included 65 articles. We identified six themes of limitations in AI that impact health equity: (1) biases in AI can perpetuate and exacerbate racial and ethnic inequities; (2) equity in algorithms should be a priority; (3) lack of diversity in the field of AI is concerning; (4) the need for regulation and testing algorithms for accuracy; (5) ethical standards for AI in health care are needed; and (6) the importance of promoting transparency and accountability.

CONCLUSIONS

While AI promises to enhance healthcare outcomes and address equity concerns, risks and challenges are associated with its implementation. To maximize the use of AI, it must be approached with an equity lens during all phases of development.

Application of Artificial Intelligence in Ophthalmology: An Updated Comprehensive Review

Artificial intelligence (AI) holds immense promise for transforming ophthalmic care through automated screening, precision diagnostics, and optimized treatment planning. This paper reviews recent advances and challenges in applying AI techniques such as machine learning and deep learning to major eye diseases. In diabetic retinopathy, AI algorithms analyze retinal images to accurately identify lesions, which helps clinicians in ophthalmology practice. Systems like IDx-DR (IDx Technologies Inc, USA) are FDA-approved for autonomous detection of referable diabetic retinopathy. For glaucoma, deep learning models assess optic nerve head morphology in fundus photographs to detect damage. In age-related macular degeneration, AI can quantify drusen and diagnose disease severity from both color fundus and optical coherence tomography images. AI has also been used in screening for retinopathy of prematurity, keratoconus, and dry eye disease. Beyond screening, AI can aid treatment decisions by forecasting disease progression and anti-VEGF response. However, potential limitations such as the quality and diversity of training data, lack of rigorous clinical validation, and challenges in regulatory approval and clinician trust must be addressed for the widespread adoption of AI. Two other significant hurdles include the integration of AI into existing clinical workflows and ensuring transparency in AI decision-making processes. With continued research to address these limitations, AI promises to enable earlier diagnosis, optimized resource allocation, personalized treatment, and improved patient outcomes. Besides, synergistic human-AI systems could set a new standard for evidence-based, precise ophthalmic care.

The Promises and Perils of Foundation Models in Dermatology

Foundation models (FM), which are large-scale artificial intelligence (AI) models that can complete a range of tasks, represent a paradigm shift in AI. These versatile models encompass large language models, vision-language models, and multimodal models. Although these models are often trained for broad tasks, they have been applied either out of the box or after additional fine tuning to tasks in medicine, including dermatology. From addressing administrative tasks to answering dermatology questions, these models are poised to have an impact on dermatology care delivery. As FMs become more ubiquitous in health care, it is important for clinicians and dermatologists to have a basic understanding of how these models are developed, what they are capable of, and what pitfalls exist. In this paper, we present a comprehensive yet accessible overview of the current state of FMs and summarize their current applications in dermatology, highlight their limitations, and discuss future developments in the field.

Efficient Normalized Conformal Prediction and Uncertainty Quantification for Anti-Cancer Drug Sensitivity Prediction with Deep Regression Forests

Deep learning models are being adopted and applied across various critical medical tasks, yet they are primarily trained to provide point predictions without providing degrees of confidence. Medical practitioner's trustworthiness of deep learning models is increased when paired with uncertainty estimations. Conformal Prediction has emerged as a promising method to pair machine learning models with prediction intervals, allowing for a view of the model's uncertainty. However, popular uncertainty estimation methods for conformal prediction fail to provide highly accurate heteroskedastic intervals. In this paper, we propose a method to estimate the uncertainty of each sample by calculating the variance obtained from a Deep Regression Forest. We show that the deep regression forest variance improves the efficiency and coverage of normalized inductive conformal prediction when applied on an anti-cancer drug sensitivity prediction task.

Detecting Algorithmic Errors and Patient Harms for AI-Enabled Medical Devices in Randomized Controlled Trials: Protocol for a Systematic Review

BACKGROUND

Artificial intelligence (AI) medical devices have the potential to transform existing clinical workflows and ultimately improve patient outcomes. AI medical devices have shown potential for a range of clinical tasks such as diagnostics, prognostics, and therapeutic decision-making such as drug dosing. There is, however, an urgent need to ensure that these technologies remain safe for all populations. Recent literature demonstrates the need for rigorous performance error analysis to identify issues such as algorithmic encoding of spurious correlations (eg, protected characteristics) or specific failure modes that may lead to patient harm. Guidelines for reporting on studies that evaluate AI medical devices require the mention of performance error analysis; however, there is still a lack of understanding around how performance errors should be analyzed in clinical studies, and what harms authors should aim to detect and report.

OBJECTIVE

This systematic review will assess the frequency and severity of AI errors and adverse events (AEs) in randomized controlled trials (RCTs) investigating AI medical devices as interventions in clinical settings. The review will also explore how performance errors are analyzed including whether the analysis includes the investigation of subgroup-level outcomes.

METHODS

This systematic review will identify and select RCTs assessing AI medical devices. Search strategies will be deployed in MEDLINE (Ovid), Embase (Ovid), Cochrane CENTRAL, and clinical trial registries to identify relevant papers. RCTs identified in bibliographic databases will be cross-referenced with clinical trial registries. The primary outcomes of interest are the frequency and severity of AI errors, patient harms, and reported AEs. Quality assessment of RCTs will be based on version 2 of the Cochrane risk-of-bias tool (RoB2). Data analysis will include a comparison of error rates and patient harms between study arms, and a meta-analysis of the rates of patient harm in control versus intervention arms will be conducted if appropriate.

RESULTS

The project was registered on PROSPERO in February 2023. Preliminary searches have been completed and the search strategy has been designed in consultation with an information specialist and methodologist. Title and abstract screening started in September 2023. Full-text screening is ongoing and data collection and analysis began in April 2024.

CONCLUSIONS

Evaluations of AI medical devices have shown promising results; however, reporting of studies has been variable. Detection, analysis, and reporting of performance errors and patient harms is vital to robustly assess the safety of AI medical devices in RCTs. Scoping searches have illustrated that the reporting of harms is variable, often with no mention of AEs. The findings of this systematic review will identify the frequency and severity of AI performance errors and patient harms and generate insights into how errors should be analyzed to account for both overall and subgroup performance.

TRIAL REGISTRATION

PROSPERO CRD42023387747; https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=387747.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

PRR1-10.2196/51614.