Lessons learned from translating AI from development to deployment in healthcare

Bias recognition and mitigation strategies in artificial intelligence healthcare applications

Artificial intelligence (AI) is delivering value across all aspects of clinical practice. However, bias may exacerbate healthcare disparities. This review examines the origins of bias in healthcare AI, strategies for mitigation, and responsibilities of relevant stakeholders towards achieving fair and equitable use. We highlight the importance of systematically identifying bias and engaging relevant mitigation activities throughout the AI model lifecycle, from model conception through to deployment and longitudinal surveillance.

Federally Qualified Health Centers as a Model to Improve Vision Health: A Systematic Review

Importance

Disparities in eye health are associated with lower-income and minoritized populations, many of whom seek care at federally qualified health centers (FQHCs).

Objective

To examine the literature addressing vision and eye health care provided at FQHCs, identify barriers to providing care at FQHCs, and highlight recommendations on how FQHCs can decrease disparities in eye health.

Evidence Review

A systematic review of Embase, SCOPUS, and PubMed was performed, and articles regarding eye and vision health at FQHCs within the US published between January 1, 1965, and July 14, 2023, were included. This review was conducted using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. Structured data and case studies were extracted and collated using an a priori method to reduce bias.

Findings

The systematic review yielded 423 unique articles, with 43 meeting inclusion criteria. Only 18.3% to 29% of FQHCs reported on-site vision services with the remainder relying on external referrals to vision specialists. Primary eye conditions evaluated included diabetic retinopathy (26 studies), general eye health (11 studies), and glaucoma (6 studies). Telehealth vision initiatives were an important method to expand access (18 studies). Other topics included economic analysis (5 studies) and policy suggestions (3 studies) to increase vision services at FQHCs. Systemic barriers to accessing care at FQHCs were the lack of eye clinicians available to provide services, the cost of resources, and limited reimbursement to implement screening programs. Patient barriers to accessing care included financial constraints for specialist care, limited awareness of the importance of eye examinations, and difficulty navigating the insurance system.

Conclusions and Relevance

Findings of this systematic review suggest that FQHCs are well positioned to increase vision services and thus improve vision health equity, serving populations who are at a higher risk for vision disorders. Results find systemic and patient-level barriers to vision health that may need to be addressed. Policy leaders could leverage existing gaps for purposeful advocacy, set standards and metrics for vision health at FQHCs, promote novel models of care, and encourage collaboration of eye clinicians with partnering FQHCs.

The Pivotal Role of Baseline LDCT for Lung Cancer Screening in the Era of Artificial Intelligence

In this narrative review, we address the ongoing challenges of lung cancer (LC) screening using chest low-dose computerized tomography (LDCT) and explore the contributions of artificial intelligence (AI), in overcoming them. We focus on evaluating the initial (baseline) LDCT examination, which provides a wealth of information relevant to the screening participant's health. This includes the detection of large-size prevalent LC and small-size malignant nodules that are typically diagnosed as LCs upon growth in subsequent annual LDCT scans. Additionally, the baseline LDCT examination provides valuable information about smoking-related comorbidities, including cardiovascular disease, chronic obstructive pulmonary disease, and interstitial lung disease (ILD), by identifying relevant markers. Notably, these comorbidities, despite the slow progression of their markers, collectively exceed LC as ultimate causes of death at follow-up in LC screening participants. Computer-assisted diagnosis tools currently improve the reproducibility of radiologic readings and reduce the false negative rate of LDCT. Deep learning (DL) tools that analyze the radiomic features of lung nodules are being developed to distinguish between benign and malignant nodules. Furthermore, AI tools can predict the risk of LC in the years following a baseline LDCT. AI tools that analyze baseline LDCT examinations can also compute the risk of cardiovascular disease or death, paving the way for personalized screening interventions. Additionally, DL tools are available for assessing osteoporosis and ILD, which helps refine the individual's current and future health profile. The primary obstacles to AI integration into the LDCT screening pathway are the generalizability of performance and the explainability.

Clinically Significant Prostate Cancer Prediction Using Multimodal Deep Learning with Prostate-Specific Antigen Restriction

Prostate cancer (PCa) is a clinically heterogeneous disease. Predicting clinically significant PCa with low-intermediate prostate-specific antigen (PSA), which often includes aggressive cancers, is imperative. This study evaluated the predictive accuracy of deep learning analysis using multimodal medical data focused on clinically significant PCa in patients with PSA ≤ 20 ng/mL. Our cohort study included 178 consecutive patients who underwent ultrasound-guided prostate biopsy. Deep learning analyses were applied to predict clinically significant PCa. We generated receiver operating characteristic curves and calculated the corresponding area under the curve (AUC) to assess the prediction. The AUC of the integrated medical data using our multimodal deep learning approach was 0.878 (95% confidence interval [CI]: 0.772-0.984) in all patients without PSA restriction. Despite the reduced predictive ability of PSA when restricted to PSA ≤ 20 ng/mL (n = 122), the AUC was 0.862 (95% CI: 0.723-1.000), complemented by imaging data. In addition, we assessed clinical presentations and images belonging to representative false-negative and false-positive cases. Our multimodal deep learning approach assists physicians in determining treatment strategies by predicting clinically significant PCa in patients with PSA ≤ 20 ng/mL before biopsy, contributing to personalized medical workflows for PCa management.

Addressing the Generalizability of AI in Radiology Using a Novel Data Augmentation Framework with Synthetic Patient Image Data: Proof-of-Concept and External Validation for Classification Tasks in Multiple Sclerosis

Artificial intelligence (AI) models often face performance drops after deployment to external datasets. This study evaluated the potential of a novel data augmentation framework based on generative adversarial networks (GANs) that creates synthetic patient image data for model training to improve model generalizability. Model development and external testing were performed for a given classification task, namely the detection of new fluid-attenuated inversion recovery lesions at MRI during longitudinal follow-up of patients with multiple sclerosis (MS). An internal dataset of 669 patients with MS (n = 3083 examinations) was used to develop an attention-based network, trained both with and without the inclusion of the GAN-based synthetic data augmentation framework. External testing was performed on 134 patients with MS from a different institution, with MR images acquired using different scanners and protocols than images used during training. Models trained using synthetic data augmentation showed a significant performance improvement when applied on external data (area under the receiver operating characteristic curve [AUC], 83.6% without synthetic data vs 93.3% with synthetic data augmentation; P = .03), achieving comparable results to the internal test set (AUC, 95.0%; P = .53), whereas models without synthetic data augmentation demonstrated a performance drop upon external testing (AUC, 93.8% on internal dataset vs 83.6% on external data; P = .03). Data augmentation with synthetic patient data substantially improved performance of AI models on unseen MRI data and may be extended to other clinical conditions or tasks to mitigate domain shift, limit class imbalance, and enhance the robustness of AI applications in medical imaging. Keywords: Brain, Brain Stem, Multiple Sclerosis, Synthetic Data Augmentation, Generative Adversarial Network Supplemental material is available for this article. © RSNA, 2024.

From Data Integration to Precision Medicine: A Value-Based Healthcare Approach for Sarcoma Care

The transformation of healthcare from a fee-for-service model to value-based care is particularly crucial in managing complex and rare diseases like sarcoma, where data fragmentation and variability present significant challenges. This manuscript reviews strategies for structured and harmonized data integration-a critical precursor to precision medicine in sarcoma care. We demonstrate how standardizing data formats, ontologies, and coding systems enable seamless integration of clinical, economic, and patient-reported outcomes across institutions, paving the way for comprehensive predictive analytics. By establishing robust value-based healthcare (VBHC) frameworks through digital transformation and predictive models, including digital twins, we create the foundation for personalized sarcoma treatment and real-world-time clinical decision-making. The manuscript also addresses practical challenges, including the need for system standardization, overcoming regulatory and privacy concerns, and managing high costs. We propose actionable strategies to overcome these barriers and discuss the role of advanced analytics and future research directions that further enhance VBHC and precision medicine. This work outlines the necessary steps to build a cohesive, data-driven approach that supports the transition to precision medicine, fundamentally improving outcomes for sarcoma patients.

Diagnostic and Prognostic Electrocardiogram-Based Models for Rapid Clinical Applications

Leveraging artificial intelligence (AI) for the analysis of electrocardiograms (ECGs) has the potential to transform diagnosis and estimate the prognosis of not only cardiac but, increasingly, noncardiac conditions. In this review, we summarize clinical studies and AI-enhanced ECG-based clinical applications in the early detection, diagnosis, and estimating prognosis of cardiovascular diseases in the past 5 years (2019-2023). With advancements in deep learning and the rapid increased use of ECG technologies, a large number of clinical studies have been published. However, most of these studies are single-centre, retrospective, proof-of-concept studies that lack external validation. Prospective studies that progress from development toward deployment in clinical settings account for < 15% of the studies. Successful implementations of ECG-based AI applications that have received approval from the Food and Drug Administration have been developed through commercial collaborations, with approximately half of them being for mobile or wearable devices. The field is in its early stages, and overcoming several obstacles is essential, such as prospective validation in multicentre large data sets, addressing technical issues, bias, privacy, data security, model generalizability, and global scalability. This review concludes with a discussion of these challenges and potential solutions. By providing a holistic view of the state of AI in ECG analysis, this review aims to set a foundation for future research directions, emphasizing the need for comprehensive, clinically integrated, and globally deployable AI solutions in cardiovascular disease management.

Artificial intelligence for retinal diseases

PURPOSE

To discuss the worldwide applications and potential impact of artificial intelligence (AI) for the diagnosis, management and analysis of treatment outcomes of common retinal diseases.

METHODS

We performed an online literature review, using PubMed Central (PMC), of AI applications to evaluate and manage retinal diseases. Search terms included AI for screening, diagnosis, monitoring, management, and treatment outcomes for age-related macular degeneration (AMD), diabetic retinopathy (DR), retinal surgery, retinal vascular disease, retinopathy of prematurity (ROP) and sickle cell retinopathy (SCR). Additional search terms included AI and color fundus photographs, optical coherence tomography (OCT), and OCT angiography (OCTA). We included original research articles and review articles.

RESULTS

Research studies have investigated and shown the utility of AI for screening for diseases such as DR, AMD, ROP, and SCR. Research studies using validated and labeled datasets confirmed AI algorithms could predict disease progression and response to treatment. Studies showed AI facilitated rapid and quantitative interpretation of retinal biomarkers seen on OCT and OCTA imaging. Research articles suggest AI may be useful for planning and performing robotic surgery. Studies suggest AI holds the potential to help lessen the impact of socioeconomic disparities on the outcomes of retinal diseases.

CONCLUSIONS

AI applications for retinal diseases can assist the clinician, not only by disease screening and monitoring for disease recurrence but also in quantitative analysis of treatment outcomes and prediction of treatment response. The public health impact on the prevention of blindness from DR, AMD, and other retinal vascular diseases remains to be determined.

Enhancing psychiatric rehabilitation outcomes through a multimodal multitask learning model based on BERT and TabNet: An approach for personalized treatment and improved decision-making

Evaluating the rehabilitation status of individuals with serious mental illnesses (SMI) necessitates a comprehensive analysis of multimodal data, including unstructured text records and structured diagnostic data. However, progress in the effective assessment of rehabilitation status remains limited. Our study develops a deep learning model integrating Bidirectional Encoder Representations from Transformers (BERT) and TabNet through a late fusion strategy to enhance rehabilitation prediction, including referral risk, dangerous behaviors, self-awareness, and medication adherence, in patients with SMI. BERT processes unstructured textual data, such as doctor's notes, whereas TabNet manages structured diagnostic information. The model's interpretability function serves to assist healthcare professionals in understanding the model's predictive decisions, improving patient care. Our model exhibited excellent predictive performance for all four tasks, with an accuracy exceeding 0.78 and an area under the curve of 0.70. In addition, a series of tests proved the model's robustness, fairness, and interpretability. This study combines multimodal and multitask learning strategies into a model and applies it to rehabilitation assessment tasks, offering a promising new tool that can be seamlessly integrated with the clinical workflow to support the provision of optimized patient care.

Sample Size Considerations for Fine-Tuning Large Language Models for Named Entity Recognition Tasks: Methodological Study

BACKGROUND

Large language models (LLMs) have the potential to support promising new applications in health informatics. However, practical data on sample size considerations for fine-tuning LLMs to perform specific tasks in biomedical and health policy contexts are lacking.

OBJECTIVE

This study aims to evaluate sample size and sample selection techniques for fine-tuning LLMs to support improved named entity recognition (NER) for a custom data set of conflicts of interest disclosure statements.

METHODS

A random sample of 200 disclosure statements was prepared for annotation. All "PERSON" and "ORG" entities were identified by each of the 2 raters, and once appropriate agreement was established, the annotators independently annotated an additional 290 disclosure statements. From the 490 annotated documents, 2500 stratified random samples in different size ranges were drawn. The 2500 training set subsamples were used to fine-tune a selection of language models across 2 model architectures (Bidirectional Encoder Representations from Transformers [BERT] and Generative Pre-trained Transformer [GPT]) for improved NER, and multiple regression was used to assess the relationship between sample size (sentences), entity density (entities per sentence [EPS]), and trained model performance (F1-score). Additionally, single-predictor threshold regression models were used to evaluate the possibility of diminishing marginal returns from increased sample size or entity density.

RESULTS

Fine-tuned models ranged in topline NER performance from F1-score=0.79 to F1-score=0.96 across architectures. Two-predictor multiple linear regression models were statistically significant with multiple R2 ranging from 0.6057 to 0.7896 (all P<.001). EPS and the number of sentences were significant predictors of F1-scores in all cases ( P<.001), except for the GPT-2_large model, where EPS was not a significant predictor (P=.184). Model thresholds indicate points of diminishing marginal return from increased training data set sample size measured by the number of sentences, with point estimates ranging from 439 sentences for RoBERTa_large to 527 sentences for GPT-2_large. Likewise, the threshold regression models indicate a diminishing marginal return for EPS with point estimates between 1.36 and 1.38.

CONCLUSIONS

Relatively modest sample sizes can be used to fine-tune LLMs for NER tasks applied to biomedical text, and training data entity density should representatively approximate entity density in production data. Training data quality and a model architecture's intended use (text generation vs text processing or classification) may be as, or more, important as training data volume and model parameter size.

The Evolving Regulatory Paradigm of AI in MedTech: A Review of Perspectives and Where We Are Today

Artificial intelligence (AI)-enabled technologies in the MedTech sector hold the promise to transform healthcare delivery by improving access, quality, and outcomes. As the regulatory contours of these technologies are being defined, there is a notable lack of literature on the key stakeholders such as the organizations and interest groups that have a significant input in shaping the regulatory framework. This article explores the perspectives and contributions of these stakeholders in shaping the regulatory paradigm of AI-enabled medical technologies. The formation of an AI regulatory framework requires the convergence of ethical, regulatory, technical, societal, and practical considerations. These multiple perspectives contribute to the various dimensions of an evolving regulatory paradigm. From the global governance guidelines set by the World Health Organization (WHO) to national regulations, the article sheds light not just on these multiple perspectives but also on their interconnectedness in shaping the regulatory landscape of AI.

Developing a practical neurodevelopmental prediction model for targeting high-risk very preterm infants during visit after NICU: a retrospective national longitudinal cohort study

BACKGROUND

Follow-up visits for very preterm infants (VPI) after hospital discharge is crucial for their neurodevelopmental trajectories, but ensuring their attendance before 12 months corrected age (CA) remains a challenge. Current prediction models focus on future outcomes at discharge, but post-discharge data may enhance predictions of neurodevelopmental trajectories due to brain plasticity. Few studies in this field have utilized machine learning models to achieve this potential benefit with transparency, explainability, and transportability.

METHODS

We developed four prediction models for cognitive or motor function at 24 months CA separately at each follow-up visits, two for the 6-month and two for the 12-month CA visits, using hospitalized and follow-up data of VPI from the Taiwan Premature Infant Follow-up Network from 2010 to 2017. Regression models were employed at 6 months CA, defined as a decline in The Bayley Scales of Infant Development 3rd edition (BSIDIII) composite score > 1 SD between 6- and 24-month CA. The delay models were developed at 12 months CA, defined as a BSIDIII composite score < 85 at 24 months CA. We used an evolutionary-derived machine learning method (EL-NDI) to develop models and compared them to those built by lasso regression, random forest, and support vector machine.

RESULTS

One thousand two hundred forty-four VPI were in the developmental set and the two validation cohorts had 763 and 1347 VPI, respectively. EL-NDI used only 4-10 variables, while the others required 29 or more variables to achieve similar performance. For models at 6 months CA, the area under the receiver operating curve (AUC) of EL-NDI were 0.76-0.81(95% CI, 0.73-0.83) for cognitive regress with 4 variables and 0.79-0.83 (95% CI, 0.76-0.86) for motor regress with 4 variables. For models at 12 months CA, the AUC of EL-NDI were 0.75-0.78 (95% CI, 0.72-0.82) for cognitive delay with 10 variables and 0.73-0.82 (95% CI, 0.72-0.85) for motor delay with 4 variables.

CONCLUSIONS

Our EL-NDI demonstrated good performance using simpler, transparent, explainable models for clinical purpose. Implementing these models for VPI during follow-up visits may facilitate more informed discussions between parents and physicians and identify high-risk infants more effectively for early intervention.

Deep learning-aided decision support for diagnosis of skin disease across skin tones

Although advances in deep learning systems for image-based medical diagnosis demonstrate their potential to augment clinical decision-making, the effectiveness of physician-machine partnerships remains an open question, in part because physicians and algorithms are both susceptible to systematic errors, especially for diagnosis of underrepresented populations. Here we present results from a large-scale digital experiment involving board-certified dermatologists (n = 389) and primary-care physicians (n = 459) from 39 countries to evaluate the accuracy of diagnoses submitted by physicians in a store-and-forward teledermatology simulation. In this experiment, physicians were presented with 364 images spanning 46 skin diseases and asked to submit up to four differential diagnoses. Specialists and generalists achieved diagnostic accuracies of 38% and 19%, respectively, but both specialists and generalists were four percentage points less accurate for the diagnosis of images of dark skin as compared to light skin. Fair deep learning system decision support improved the diagnostic accuracy of both specialists and generalists by more than 33%, but exacerbated the gap in the diagnostic accuracy of generalists across skin tones. These results demonstrate that well-designed physician-machine partnerships can enhance the diagnostic accuracy of physicians, illustrating that success in improving overall diagnostic accuracy does not necessarily address bias.

An intentional approach to managing bias in general purpose embedding models

Advances in machine learning for health care have brought concerns about bias from the research community; specifically, the introduction, perpetuation, or exacerbation of care disparities. Reinforcing these concerns is the finding that medical images often reveal signals about sensitive attributes in ways that are hard to pinpoint by both algorithms and people. This finding raises a question about how to best design general purpose pretrained embeddings (GPPEs, defined as embeddings meant to support a broad array of use cases) for building downstream models that are free from particular types of bias. The downstream model should be carefully evaluated for bias, and audited and improved as appropriate. However, in our view, well intentioned attempts to prevent the upstream components-GPPEs-from learning sensitive attributes can have unintended consequences on the downstream models. Despite producing a veneer of technical neutrality, the resultant end-to-end system might still be biased or poorly performing. We present reasons, by building on previously published data, to support the reasoning that GPPEs should ideally contain as much information as the original data contain, and highlight the perils of trying to remove sensitive attributes from a GPPE. We also emphasise that downstream prediction models trained for specific tasks and settings, whether developed using GPPEs or not, should be carefully designed and evaluated to avoid bias that makes models vulnerable to issues such as distributional shift. These evaluations should be done by a diverse team, including social scientists, on a diverse cohort representing the full breadth of the patient population for which the final model is intended.

Experiences of using artificial intelligence in healthcare: a qualitative study of UK clinician and key stakeholder perspectives

OBJECTIVES

Artificial intelligence (AI) is a rapidly developing field in healthcare, with tools being developed across various specialties to support healthcare professionals and reduce workloads. It is important to understand the experiences of professionals working in healthcare to ensure that future AI tools are acceptable and effectively implemented. The aim of this study was to gain an in-depth understanding of the experiences and perceptions of UK healthcare workers and other key stakeholders about the use of AI in the National Health Service (NHS).

DESIGN

A qualitative study using semistructured interviews conducted remotely via MS Teams. Thematic analysis was carried out.

SETTING

NHS and UK higher education institutes.

PARTICIPANTS

Thirteen participants were recruited, including clinical and non-clinical participants working for the NHS and researchers working to develop AI tools for healthcare settings.

RESULTS

Four core themes were identified: positive perceptions of AI; potential barriers to using AI in healthcare; concerns regarding AI use and steps needed to ensure the acceptability of future AI tools. Overall, we found that those working in healthcare were generally open to the use of AI and expected it to have many benefits for patients and facilitate access to care. However, concerns were raised regarding the security of patient data, the potential for misdiagnosis and that AI could increase the burden on already strained healthcare staff.

CONCLUSION

This study found that healthcare staff are willing to engage with AI research and incorporate AI tools into care pathways. Going forward, the NHS and AI developers will need to collaborate closely to ensure that future tools are suitable for their intended use and do not negatively impact workloads or patient trust. Future AI studies should continue to incorporate the views of key stakeholders to improve tool acceptability.

TRIAL REGISTRATION NUMBER

NCT05028179; ISRCTN15113915; IRAS ref: 293515.

Artificial intelligence and digital health in global eye health: opportunities and challenges

Global eye health is defined as the degree to which vision, ocular health, and function are maximised worldwide, thereby optimising overall wellbeing and quality of life. Improving eye health is a global priority as a key to unlocking human potential by reducing the morbidity burden of disease, increasing productivity, and supporting access to education. Although extraordinary progress fuelled by global eye health initiatives has been made over the last decade, there remain substantial challenges impeding further progress. The accelerated development of digital health and artificial intelligence (AI) applications provides an opportunity to transform eye health, from facilitating and increasing access to eye care to supporting clinical decision making with an objective, data-driven approach. Here, we explore the opportunities and challenges presented by digital health and AI in global eye health and describe how these technologies could be leveraged to improve global eye health. AI, telehealth, and emerging technologies have great potential, but require specific work to overcome barriers to implementation. We suggest that a global digital eye health task force could facilitate coordination of funding, infrastructural development, and democratisation of AI and digital health to drive progress forwards in this domain.

Machine learning models for diagnosis and prognosis of Parkinson's disease using brain imaging: general overview, main challenges, and future directions

Parkinson's disease (PD) is a progressive and complex neurodegenerative disorder associated with age that affects motor and cognitive functions. As there is currently no cure, early diagnosis and accurate prognosis are essential to increase the effectiveness of treatment and control its symptoms. Medical imaging, specifically magnetic resonance imaging (MRI), has emerged as a valuable tool for developing support systems to assist in diagnosis and prognosis. The current literature aims to improve understanding of the disease's structural and functional manifestations in the brain. By applying artificial intelligence to neuroimaging, such as deep learning (DL) and other machine learning (ML) techniques, previously unknown relationships and patterns can be revealed in this high-dimensional data. However, several issues must be addressed before these solutions can be safely integrated into clinical practice. This review provides a comprehensive overview of recent ML techniques analyzed for the automatic diagnosis and prognosis of PD in brain MRI. The main challenges in applying ML to medical diagnosis and its implications for PD are also addressed, including current limitations for safe translation into hospitals. These challenges are analyzed at three levels: disease-specific, task-specific, and technology-specific. Finally, potential future directions for each challenge and future perspectives are discussed.