Reporting of prognostic clinical prediction models based on machine learning methods in oncology needs to be improved

Methodological and reporting quality of machine learning studies on cancer diagnosis, treatment, and prognosis

This study aimed to evaluate the quality and transparency of reporting in studies using machine learning (ML) in oncology, focusing on adherence to the Consolidated Reporting Guidelines for Prognostic and Diagnostic Machine Learning Models (CREMLS), TRIPOD-AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis), and PROBAST (Prediction Model Risk of Bias Assessment Tool). The literature search included primary studies published between February 1, 2024, and January 31, 2025, that developed or tested ML models for cancer diagnosis, treatment, or prognosis. To reflect the current state of the rapidly evolving landscape of ML applications in oncology, fifteen most recent articles in each category were selected for evaluation. Two independent reviewers screened studies and extracted data on study characteristics, reporting quality (CREMLS and TRIPOD+AI), risk of bias (PROBAST), and ML performance metrics. The most frequently studied cancer types were breast cancer (n=7/45; 15.6%), lung cancer (n=7/45; 15.6%), and liver cancer (n=5/45; 11.1%). The findings indicate several deficiencies in reporting quality, as assessed by CREMLS and TRIPOD+AI. These deficiencies primarily relate to sample size calculation, reporting on data quality, strategies for handling outliers, documentation of ML model predictors, access to training or validation data, and reporting on model performance heterogeneity. The methodological quality assessment using PROBAST revealed that 89% of the included studies exhibited a low overall risk of bias, and all studies have shown a low risk of bias in terms of applicability. Regarding the specific AI models identified as the best-performing, Random Forest (RF) and XGBoost were the most frequently reported, each used in 17.8% of the studies (n = 8). Additionally, our study outlines the specific areas where reporting is deficient, providing researchers with guidance to improve reporting quality in these sections and, consequently, reduce the risk of bias in their studies.

Larger sample sizes are needed when developing a clinical prediction model using machine learning in oncology: methodological systematic review

BACKGROUND AND OBJECTIVES

Having a sufficient sample size is crucial when developing a clinical prediction model. We reviewed details of sample size in studies developing prediction models for binary outcomes using machine learning (ML) methods within oncology and compared the sample size used to develop the models with the minimum required sample size needed when developing a regression-based model (Nmin).

METHODS

We searched the Medline (via OVID) database for studies developing a prediction model using ML methods published in December 2022. We reviewed how sample size was justified. We calculated Nmin, which is the Nmin, and compared this with the sample size that was used to develop the models.

RESULTS

Only one of 36 included studies justified their sample size. We were able to calculate Nmin for 17 (47%) studies. 5/17 studies met Nmin, allowing to precisely estimate the overall risk and minimize overfitting. There was a median deficit of 302 participants with the event (n = 17; range: -21,331 to 2298) when developing the ML models. An additional three out of the 17 studies met the required sample size to precisely estimate the overall risk only.

CONCLUSION

Studies developing a prediction model using ML in oncology seldom justified their sample size and sample sizes were often smaller than Nmin. As ML models almost certainly require a larger sample size than regression models, the deficit is likely larger. We recommend that researchers consider and report their sample size and at least meet the minimum sample size required when developing a regression-based model.

Public Disclosure of Results From Artificial Intelligence/Machine Learning Research in Health Care: Comprehensive Analysis of ClinicalTrials.gov, PubMed, and Scopus Data (2010-2023)

BACKGROUND

Despite the rapid growth of research in artificial intelligence/machine learning (AI/ML), little is known about how often study results are disclosed years after study completion.

OBJECTIVE

We aimed to estimate the proportion of AI/ML research that reported results through ClinicalTrials.gov or peer-reviewed publications indexed in PubMed or Scopus.

METHODS

Using data from the Clinical Trials Transformation Initiative Aggregate Analysis of ClinicalTrials.gov, we identified studies initiated and completed between January 2010 and December 2023 that contained AI/ML-specific terms in the official title, brief summary, interventions, conditions, detailed descriptions, primary outcomes, or keywords. For 842 completed studies, we searched PubMed and Scopus for publications containing study identifiers and AI/ML-specific terms in relevant fields, such as the title, abstract, and keywords. We calculated disclosure rates within 3 years of study completion and median times to disclosure-from the "primary completion date" to the "results first posted date" on ClinicalTrials.gov or the earliest date of journal publication.

RESULTS

When restricted to studies completed before 2021, ensuring at least 3 years of follow-up in which to report results, 7.0% (22/316) disclosed results on ClinicalTrials.gov, 16.5% (52/316) in journal publications, and 20.6% (65/316) through either route within 3 years of completion. Higher disclosure rates were observed for trials: 11.0% (15/136) on ClinicalTrials.gov, 25.0% (34/136) in journal publications, and 30.1% (41/136) through either route. Randomized controlled trials had even higher disclosure rates: 12.2% (9/74) on ClinicalTrials.gov, 31.1% (23/74) in journal publications, and 36.5% (27/74) through either route. Nevertheless, most study findings (79.4%; 251/316) remained undisclosed 3 years after study completion. Trials using randomization (vs nonrandomized) or masking (vs open label) had higher disclosure rates and shorter times to disclosure. Most trials (85%; 305/357) had sample sizes of ≤1000, yet larger trials (n>1000) had higher publication rates (30.8%; 16/52) than smaller trials (n≤1000) (17.4%; 53/305). Hospitals (12.4%; 42/340), academia (15.1%; 39/259), and industry (13.7%; 20/146) published the most. High-income countries accounted for 82.4% (89/108) of all published studies. Of studies with disclosed results, the median times to report through ClinicalTrials.gov and in journal publications were 505 days (IQR 399-676) and 407 days (IQR 257-674), respectively. Open-label trials were common (60%; 214/357). Single-center designs were prevalent in both trials (83.3%; 290/348) and observational studies (82.3%; 377/458).

CONCLUSIONS

For nearly 80% of completed studies, findings remained undisclosed within the 3 years of follow-up, raising questions about the representativeness of publicly available evidence. While methodological rigor was generally associated with higher publication rates, the predominance of single-center designs and high-income countries may limit the generalizability of the results currently accessible.

Development of a Digestive Cancer Risk Score Based on Nutritional Predictors: A Risk Prediction Model in the Golestan Cohort Study

This study aimed to develop a non-laboratory simple and useful scoring system to predict risk of incident digestive cancers within the healthcare and clinical framework in Iranian population. The present study was conducted on the collected data from the Golestan Cohort Study. A total of 49,173 participants, aged 37-80 years, were recruited from Gonbad City and 326 rural villages and were followed from 2004 to 2021 in Iran. A non-laboratory model for prediction of the 15-year risk of digestive cancers by means of dietary predictors and formulating a simple and useful scoring system in Iranian population was done in this study. A total of 43550 participants (25249 women and 18301 men) were included in the final analysis. The model's discrimination and calibration were assessed by concordance statistic (C-statistic) and calibration plot, respectively. The model had an acceptable discrimination in both derivation (C-statistic: 0.76) and validation (C-statistic: 0.70) samples (p < 0.001). Also, the calibration of model in derivation and validation datasets was 0.88 and 0.91, respectively. As an assessment tool, the established simple and practical nutritional risk score is suitable for motivating at-risk individuals to change lifestyles and dietary patterns to reduce future risks and prevent health problems.

Overlooked and underpowered: a meta-research addressing sample size in radiomics prediction models for binary outcomes

OBJECTIVES

To investigate how studies determine the sample size when developing radiomics prediction models for binary outcomes, and whether the sample size meets the estimates obtained by using established criteria.

METHODS

We identified radiomics studies that were published from 01 January 2023 to 31 December 2023 in seven leading peer-reviewed radiological journals. We reviewed the sample size justification methods, and actual sample size used. We calculated and compared the actual sample size used to the estimates obtained by using three established criteria proposed by Riley et al. We investigated which characteristics factors were associated with the sufficient sample size that meets the estimates obtained by using established criteria proposed by Riley et al. RESULTS: We included 116 studies. Eleven out of one hundred sixteen studies justified the sample size, in which 6/11 performed a priori sample size calculation. The median (first and third quartile, Q1, Q3) of the total sample size is 223 (130, 463), and those of sample size for training are 150 (90, 288). The median (Q1, Q3) difference between total sample size and minimum sample size according to established criteria are -100 (-216, 183), and those differences between total sample size and a more restrictive approach based on established criteria are -268 (-427, -157). The presence of external testing and the specialty of the topic were associated with sufficient sample size.

CONCLUSION

Radiomics studies are often designed without sample size justification, whose sample size may be too small to avoid overfitting. Sample size justification is encouraged when developing a radiomics model.

KEY POINTS

Question Sample size justification is critical to help minimize overfitting in developing a radiomics model, but is overlooked and underpowered in radiomics research. Findings Few of the radiomics models justified, calculated, or reported their sample size, and most of them did not meet the recent formal sample size criteria. Clinical relevance Radiomics models are often designed without sample size justification. Consequently, many models are too small to avoid overfitting. It should be encouraged to justify, perform, and report the considerations on sample size when developing radiomics models.

The reporting quality and methodological quality of dynamic prediction models for cancer prognosis

BACKGROUND

To evaluate the reporting quality and methodological quality of dynamic prediction model (DPM) studies on cancer prognosis.

METHODS

Extensive search for DPM studies on cancer prognosis was conducted in MEDLINE, EMBASE, and the Cochrane Library databases. The Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) and the Prediction model Risk of Bias Assessment Tool (PROBAST) were used to assess reporting quality and methodological quality, respectively.

RESULTS

A total of 34 DPM studies were identified since the first publication in 2005, the main modeling methods for DPMs included the landmark model and the joint model. Regarding the reporting quality, the median overall TRIPOD adherence score was 75%. The TRIPOD items were poorly reported, especially the title (23.53%), model specification, including presentation (55.88%) and interpretation (50%) of the DPM usage, and implications for clinical use and future research (29.41%). Concerning methodological quality, most studies were of low quality (n = 30) or unclear (n = 3), mainly due to statistical analysis issues.

CONCLUSIONS

The Landmark model and joint model show potential in DPM. The suboptimal reporting and methodological qualities of current DPM studies should be improved to facilitate clinical application.

Externally validated and clinically useful machine learning algorithms to support patient-related decision-making in oncology: a scoping review

BACKGROUND

This scoping review systematically maps externally validated machine learning (ML)-based models in cancer patient care, quantifying their performance, and clinical utility, and examining relationships between models, cancer types, and clinical decisions. By synthesizing evidence, this study identifies, strengths, limitations, and areas requiring further research.

METHODS

The review followed the Joanna Briggs Institute's methodology, Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews guidelines, and the Population, Concept, and Context mnemonic. Searches were conducted across Embase, IEEE Xplore, PubMed, Scopus, and Web of Science (January 2014-September 2022), targeting English-language quantitative studies in Q1 journals (SciMago Journal and Country Ranking > 1) that used ML to evaluate clinical outcomes for human cancer patients with commonly available data. Eligible models required external validation, clinical utility assessment, and performance metric reporting. Studies involving genetics, synthetic patients, plants, or animals were excluded. Results were presented in tabular, graphical, and descriptive form.

RESULTS

From 4023 deduplicated abstracts and 636 full-text reviews, 56 studies (2018-2022) met the inclusion criteria, covering diverse cancer types and applications. Convolutional neural networks were most prevalent, demonstrating high performance, followed by gradient- and decision tree-based algorithms. Other algorithms, though underrepresented, showed promise. Lung and digestive system cancers were most frequently studied, focusing on diagnosis and outcome predictions. Most studies were retrospective and multi-institutional, primarily using image-based data, followed by text-based and hybrid approaches. Clinical utility assessments involved 499 clinicians and 12 tools, indicating improved clinician performance with AI assistance and superior performance to standard clinical systems.

DISCUSSION

Interest in ML-based clinical decision-making has grown in recent years alongside increased multi-institutional collaboration. However, small sample sizes likely impacted data quality and generalizability. Persistent challenges include limited international validation across ethnicities, inconsistent data sharing, disparities in validation metrics, and insufficient calibration reporting, hindering model comparison reliability.

CONCLUSION

Successful integration of ML in oncology decision-making requires standardized data and methodologies, larger sample sizes, greater transparency, and robust validation and clinical utility assessments.

OTHER

Financed by FCT-Fundação para a Ciência e a Tecnologia (Portugal, project LA/P/0063/2020, grant 2021.09040.BD) as part of CSS's Ph.D. This work was not registered.

Updating methods for artificial intelligence-based clinical prediction models: a scoping review

OBJECTIVES

To give an overview of methods for updating artificial intelligence (AI)-based clinical prediction models based on new data.

STUDY DESIGN AND SETTING

We comprehensively searched Scopus and Embase up to August 2022 for articles that addressed developments, descriptions, or evaluations of prediction model updating methods. We specifically focused on articles in the medical domain involving AI-based prediction models that were updated based on new data, excluding regression-based updating methods as these have been extensively discussed elsewhere. We categorized and described the identified methods used to update the AI-based prediction model as well as the use cases in which they were used.

RESULTS

We included 78 articles. The majority of the included articles discussed updating for neural network methods (93.6%) with medical images as input data (65.4%). In many articles (51.3%) existing, pretrained models for broad tasks were updated to perform specialized clinical tasks. Other common reasons for model updating were to address changes in the data over time and cross-center differences; however, more unique use cases were also identified, such as updating a model from a broad population to a specific individual. We categorized the identified model updating methods into four categories: neural network-specific methods (described in 92.3% of the articles), ensemble-specific methods (2.5%), model-agnostic methods (9.0%), and other (1.3%). Variations of neural network-specific methods are further categorized based on the following: (1) the part of the original neural network that is kept, (2) whether and how the original neural network is extended with new parameters, and (3) to what extent the original neural network parameters are adjusted to the new data. The most frequently occurring method (n = 30) involved selecting the first layer(s) of an existing neural network, appending new, randomly initialized layers, and then optimizing the entire neural network.

CONCLUSION

We identified many ways to adjust or update AI-based prediction models based on new data, within a large variety of use cases. Updating methods for AI-based prediction models other than neural networks (eg, random forest) appear to be underexplored in clinical prediction research.

PLAIN LANGUAGE SUMMARY

AI-based prediction models are increasingly used in health care, helping clinicians with diagnosing diseases, guiding treatment decisions, and informing patients. However, these prediction models do not always work well when applied to hospitals, patient populations, or times different from those used to develop the models. Developing new models for every situation is neither practical nor desired, as it wastes resources, time, and existing knowledge. A more efficient approach is to adjust existing models to new contexts ('updating'), but there is limited guidance on how to do this for AI-based clinical prediction models. To address this, we reviewed 78 studies in detail to understand how researchers are currently updating AI-based clinical prediction models, and the types of situations in which these updating methods are used. Our findings provide a comprehensive overview of the available methods to update existing models. This is intended to serve as guidance and inspiration for researchers. Ultimately, this can lead to better reuse of existing models and improve the quality and efficiency of AI-based prediction models in health care.

Seeing the random forest through the decision trees. Supporting learning health systems from histopathology with machine learning models: Challenges and opportunities

This paper discusses some overlooked challenges faced when working with machine learning models for histopathology and presents a novel opportunity to support "Learning Health Systems" with them. Initially, the authors elaborate on these challenges after separating them according to their mitigation strategies: those that need innovative approaches, time, or future technological capabilities and those that require a conceptual reappraisal from a critical perspective. Then, a novel opportunity to support "Learning Health Systems" by integrating hidden information extracted by ML models from digitalized histopathology slides with other healthcare big data is presented.

Several methods for assessing research waste in reviews with a systematic search: a scoping review

Background

Research waste is present in all study designs and can have significant consequences for science, including reducing the reliability of research findings and contributing to the inefficient use of resources. Estimates suggest that as much as 85% of all biomedical research is wasted. However, it is uncertain how avoidable research waste is assessed in specific types of study designs and what methods could be used to examine different aspects of research waste. We aimed to investigate which methods, systematic reviews, scoping reviews, and overviews of reviews discussing research waste, have used to assess avoidable research waste.

Materials and Methods

We published a protocol in the Open Science Framework prospectively (https://osf.io/2fbp4). We searched PubMed and Embase with a 30-year limit (January 1993-August 2023). The concept examined was how research waste and related synonyms (e.g., unnecessary, redundant, duplicate, etc.) were assessed in reviews with a systematic search: systematic, scoping, or overviews of reviews. We extracted data on the method used in the review to examine for research waste and for which study design this method was applied.

Results

The search identified 4,285 records of which 93 reviews with systematic searches were included. The reviews examined a median of 90 (range 10-6,781) studies, where the study designs most commonly included were randomized controlled trials (48%) and systematic reviews (33%). In the last ten years, the number of reports assessing research waste has increased. More than 50% of examined reviews reported evaluating methodological research waste among included studies, typically using tools such as one of Cochrane Risk of Bias tools (n = 8) for randomized controlled trials or AMSTAR 1 or 2 (n = 12) for systematic reviews. One fourth of reviews assessed reporting guideline adherence to e.g., CONSORT (n = 4) for randomized controlled trials or PRISMA (n = 6) for systematic reviews.

Conclusion

Reviews with systematic searches focus on methodological quality and reporting guideline adherence when examining research waste. However, this scoping review revealed that a wide range of tools are used, which may pose difficulties in comparing examinations and performing meta-research. This review aids researchers in selecting methodologies and contributes to the ongoing discourse on optimizing research efficiency.

Machine learning models including patient-reported outcome data in oncology: a systematic literature review and analysis of their reporting quality

PURPOSE

To critically examine the current state of machine learning (ML) models including patient-reported outcome measure (PROM) scores in cancer research, by investigating the reporting quality of currently available studies and proposing areas of improvement for future use of ML in the field.

METHODS

PubMed and Web of Science were systematically searched for publications of studies on patients with cancer applying ML models with PROM scores as either predictors or outcomes. The reporting quality of applied ML models was assessed utilizing an adapted version of the MI-CLAIM (Minimum Information about CLinical Artificial Intelligence Modelling) checklist. The key variables of the checklist are study design, data preparation, model development, optimization, performance, and examination. Reproducibility and transparency complement the reporting quality criteria.

RESULTS

The literature search yielded 1634 hits, of which 52 (3.2%) were eligible. Thirty-six (69.2%) publications included PROM scores as a predictor and 32 (61.5%) as an outcome. Results of the reporting quality appraisal indicate a potential for improvement, especially in the areas of model examination. According to the standards of the MI-CLAIM checklist, the reporting quality of ML models in included studies proved to be low. Only nine (17.3%) publications present a discussion about the clinical applicability of the developed model and reproducibility and only three (5.8%) provide a code to reproduce the model and the results.

CONCLUSION

The herein performed critical examination of the status quo of the application of ML models including PROM scores in published oncological studies allowed the identification of areas of improvement for reporting and future use of ML in the field.

Improvements Are Needed in the Adherence to the TRIPOD Statement for Clinical Prediction Models for Patients With Spinal Pain or Osteoarthritis: A Metaresearch Study

This metaresearch study aimed to evaluate the completeness of reporting of prediction model studies in patients with spinal pain or osteoarthritis (OA) in terms of adherence to the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) statement. We searched for prognostic and diagnostic prediction models in patients with spinal pain or OA in MEDLINE, Embase, Web of Science, and CINAHL. Using a standardized assessment form, we assessed the adherence to the TRIPOD of the included studies. Two independent reviewers performed the study selection and data extraction phases. We included 66 studies. Approximately 35% of the studies declared to have used the TRIPOD. The median adherence to the TRIPOD was 59% overall (interquartile range (IQR): 21.8), with the items of the methods and results sections having the worst reporting. Studies on neck pain had better adherence to the TRIPOD than studies on back pain and OA (medians of 76.5%, 59%, and 53%, respectively). External validation studies had the highest total adherence (median: 79.5%, IQR: 12.8) of all the study types. The median overall adherence was 4 points higher in studies that declared TRIPOD use than those that did not. Finally, we did not observe any improvement in adherence over the years. The adherence to the TRIPOD of prediction models in the spinal and OA fields is low, with the methods and results sections being the most poorly reported. Future studies on prediction models in spinal pain and OA should follow the TRIPOD to improve their reporting completeness. PERSPECTIVE: This article provides data about adherence to the TRIPOD statement in 66 prediction model studies for spinal pain or OA. The adherence to the TRIPOD statement was found to be low (median adherence of 59%). This inadequate reporting may negatively impact the effective use of the models in clinical practice.

Planned but ever published? A retrospective analysis of clinical prediction model studies registered on clinicaltrials.gov since 2000

OBJECTIVES

To describe the characteristics and publication outcomes of clinical prediction model studies registered on clinicaltrials.gov since 2000.

STUDY DESIGN AND SETTING

Observational studies registered on clinicaltrials.gov between January 1, 2000, and March 2, 2022, describing the development of a new clinical prediction model or the validation of an existing model for predicting individual-level prognostic or diagnostic risk were analyzed. Eligible clinicaltrials.gov records were classified by modeling study type (development, validation) and the model outcome being predicted (prognostic, diagnostic). Recorded characteristics included study status, sample size information, Medical Subject Headings, and plans to share individual participant data. Publication outcomes were analyzed by linking National Clinical Trial numbers for eligible records with PubMed abstracts.

RESULTS

Nine hundred twenty-eight records were analyzed from a possible 89,896 observational study records. Publications searches found 170 matching peer-reviewed publications for 137 clinicaltrials.gov records. The estimated proportion of records with 1 or more matching publications after accounting for time since study start was 2.8% at 2 years (95% CI: 1.7%, 3.9%), 12.3% at 5 years (9.8% to 14.9%) and 27% at 10 years (23% to 33%). Stratifying records by study start year indicated that publication proportions improved over time. Records tended to prioritize the development of new prediction models over the validation of existing models (76%; 704/928 vs. 24%; 182/928). At the time of download, 27% of records were marked as complete, 35% were still recruiting, and 14.7% had unknown status. Only 7.4% of records stated plans to share individual participant data.

CONCLUSION

Published clinical prediction model studies are only a fraction of overall research efforts, with many studies planned but not completed or published. Improving the uptake of study preregistration and follow-up will increase the visibility of planned research. Introducing additional registry features and guidance may improve the identification of clinical prediction model studies posted to clinical registries.

Identifying patients with undiagnosed small intestinal neuroendocrine tumours in primary care using statistical and machine learning: model development and validation study

BACKGROUND

Neuroendocrine tumours (NETs) are increasing in incidence, often diagnosed at advanced stages, and individuals may experience years of diagnostic delay, particularly when arising from the small intestine (SI). Clinical prediction models could present novel opportunities for case finding in primary care.

METHODS

An open cohort of adults (18+ years) contributing data to the Optimum Patient Care Research Database between 1st Jan 2000 and 30th March 2023 was identified. This database collects de-identified data from general practices in the UK. Model development approaches comprised logistic regression, penalised regression, and XGBoost. Performance (discrimination and calibration) was assessed using internal-external cross-validation. Decision analysis curves compared clinical utility.

RESULTS

Of 11.7 million individuals, 382 had recorded SI NET diagnoses (0.003%). The XGBoost model had the highest AUC (0.869, 95% confidence interval [CI]: 0.841-0.898) but was mildly miscalibrated (slope 1.165, 95% CI: 1.088-1.243; calibration-in-the-large 0.010, 95% CI: -0.164 to 0.185). Clinical utility was similar across all models.

DISCUSSION

Multivariable prediction models may have clinical utility in identifying individuals with undiagnosed SI NETs using information in their primary care records. Further evaluation including external validation and health economics modelling may identify cost-effective strategies for case finding for this uncommon tumour.

A machine learning predictive model for recurrence of resected distal cholangiocarcinoma: Development and validation of predictive model using artificial intelligence

INTRODUCTION

Distal Cholangiocarcinoma (dCCA) represents a challenge in hepatobiliary oncology, that requires nuanced post-resection prognostic modeling. Conventional staging criteria may oversimplify dCCA complexities, prompting the exploration of novel prognostic factors and methodologies, including machine learning algorithms. This study aims to develop a machine learning predictive model for recurrence after resected dCCA.

MATERIAL AND METHODS

This retrospective multicentric observational study included patients with dCCA from 13 international centers who underwent curative pancreaticoduodenectomy (PD). A LASSO-regularized Cox regression model was used to feature selection, examine the path of the coefficient and create a model to predict recurrence. Internal and external validation and model performance were assessed using the C-index score. Additionally, a web application was developed to enhance the clinical use of the algorithm.

RESULTS

Among 654 patients, LNR (Lymph Node Ratio) 15, neural invasion, N stage, surgical radicality, and differentiation grade emerged as significant predictors of disease-free survival (DFS). The model showed the best discrimination capacity with a C-index value of 0.8 (CI 95 %, 0.77%-0.86 %) and highlighted LNR15 as the most influential factor. Internal and external validations showed the model's robustness and discriminative ability with an Area Under the Curve of 92.4 % (95 % CI, 88.2%-94.4 %) and 91.5 % (95 % CI, 88.4%-93.5 %), respectively. The predictive model is available at https://imim.shinyapps.io/LassoCholangioca/.

CONCLUSIONS

This study pioneers the integration of machine learning into prognostic modeling for dCCA, yielding a robust predictive model for DFS following PD. The tool can provide information to both patients and healthcare providers, enhancing tailored treatments and follow-up.

Prediction model protocols indicate better adherence to recommended guidelines for study conduct and reporting

BACKGROUND AND OBJECTIVE

Protocols are invaluable documents for any research study, especially for prediction model studies. However, the mere existence of a protocol is insufficient if key details are omitted. We reviewed the reporting content and details of the proposed design and methods reported in published protocols for prediction model research.

METHODS

We searched MEDLINE, Embase, and the Web of Science Core Collection for protocols for studies developing or validating a diagnostic or prognostic model using any modeling approach in any clinical area. We screened protocols published between Jan 1, 2022 and June 30, 2022. We used the abstract, introduction, methods, and discussion sections of The Transparent Reporting of a multivariable prediction model of Individual Prognosis Or Diagnosis (TRIPOD) statement to inform data extraction.

RESULTS

We identified 30 protocols, of which 28 were describing plans for model development and six for model validation. All protocols were open access, including a preprint. 15 protocols reported prospectively collecting data. 21 protocols planned to use clustered data, of which one-third planned methods to account for it. A planned sample size was reported for 93% development and 67% validation analyses. 16 protocols reported details of study registration, but all protocols reported a statement on ethics approval. Plans for data sharing were reported in 13 protocols.

CONCLUSION

Protocols for prediction model studies are uncommon, and few are made publicly available. Those that are available were reasonably well-reported and often described their methods following current prediction model research recommendations, likely leading to better reporting and methods in the actual study.

Machine learning algorithms to predict outcomes in children and adolescents with COVID-19: A systematic review

BACKGROUND AND OBJECTIVES

We aimed to analyze the study designs, modeling approaches, and performance evaluation metrics in studies using machine learning techniques to develop clinical prediction models for children and adolescents with COVID-19.

METHODS

We searched four databases for articles published between 01/01/2020 and 10/25/2023, describing the development of multivariable prediction models using any machine learning technique for predicting several outcomes in children and adolescents who had COVID-19.

RESULTS

We included ten articles, six (60 % [95 % confidence interval (CI) 0.31 - 0.83]) were predictive diagnostic models and four (40% [95 % CI 0.170.69]) were prognostic models. All models were developed to predict a binary outcome (n= 10/10, 100 % [95 % CI 0.72-1]). The most frequently predicted outcome was disease detection (n=3/10, 30% [95 % CI 0.11-0.60]). The most commonly used machine learning models in the studies were tree-based (n=12/33, 36.3% [95 % CI 0.17-0.47]) and neural networks (n=9/27, 33.2% [95% CI 0.15-0.44]).

CONCLUSION

Our review revealed that attention is required to address problems including small sample sizes, inconsistent reporting practices on data preparation, biases in data sources, lack of reporting metrics such as calibration and discrimination, hyperparameters and other aspects that allow reproducibility by other researchers and might improve the methodology.

The reporting completeness and transparency of systematic reviews of prognostic prediction models for COVID-19 was poor: a methodological overview of systematic reviews

OBJECTIVES

To conduct a methodological overview of reviews to evaluate the reporting completeness and transparency of systematic reviews (SRs) of prognostic prediction models (PPMs) for COVID-19.

STUDY DESIGN AND SETTING

MEDLINE, Scopus, Cochrane Database of Systematic Reviews, and Epistemonikos (epistemonikos.org) were searched for SRs of PPMs for COVID-19 until December 31, 2022. The risk of bias in systematic reviews tool was used to assess the risk of bias. The protocol for this overview was uploaded in the Open Science Framework (https://osf.io/7y94c).

RESULTS

Ten SRs were retrieved; none of them synthesized the results in a meta-analysis. For most of the studies, there was absence of a predefined protocol and missing information on study selection, data collection process, and reporting of primary studies and models included, while only one SR had its data publicly available. In addition, for the majority of the SRs, the overall risk of bias was judged as being high. The overall corrected covered area was 6.3% showing a small amount of overlapping among the SRs.

CONCLUSION

The reporting completeness and transparency of SRs of PPMs for COVID-19 was poor. Guidance is urgently required, with increased awareness and education of minimum reporting standards and quality criteria. Specific focus is needed in predefined protocol, information on study selection and data collection process, and in the reporting of findings to improve the quality of SRs of PPMs for COVID-19.

Trust and acceptability of data-driven clinical recommendations in everyday practice: A scoping review

BACKGROUND

Increasing attention is being given to the analysis of large health datasets to derive new clinical decision support systems (CDSS). However, few data-driven CDSS are being adopted into clinical practice. Trust in these tools is believed to be fundamental for acceptance and uptake but to date little attention has been given to defining or evaluating trust in clinical settings.

OBJECTIVES

A scoping review was conducted to explore how and where acceptability and trustworthiness of data-driven CDSS have been assessed from the health professional's perspective.

METHODS

Medline, Embase, PsycInfo, Web of Science, Scopus, ACM Digital, IEEE Xplore and Google Scholar were searched in March 2022 using terms expanded from: "data-driven" AND "clinical decision support" AND "acceptability". Included studies focused on healthcare practitioner-facing data-driven CDSS, relating directly to clinical care. They included trust or a proxy as an outcome, or in the discussion. The preferred reporting items for systematic reviews and meta-analyses extension for scoping reviews (PRISMA-ScR) is followed in the reporting of this review.

RESULTS

3291 papers were screened, with 85 primary research studies eligible for inclusion. Studies covered a diverse range of clinical specialisms and intended contexts, but hypothetical systems (24) outnumbered those in clinical use (18). Twenty-five studies measured trust, via a wide variety of quantitative, qualitative and mixed methods. A further 24 discussed themes of trust without it being explicitly evaluated, and from these, themes of transparency, explainability, and supporting evidence were identified as factors influencing healthcare practitioner trust in data-driven CDSS.

CONCLUSION

There is a growing body of research on data-driven CDSS, but few studies have explored stakeholder perceptions in depth, with limited focused research on trustworthiness. Further research on healthcare practitioner acceptance, including requirements for transparency and explainability, should inform clinical implementation.

ADNEX risk prediction model for diagnosis of ovarian cancer: systematic review and meta-analysis of external validation studies

Objectives

To conduct a systematic review of studies externally validating the ADNEX (Assessment of Different Neoplasias in the adnexa) model for diagnosis of ovarian cancer and to present a meta-analysis of its performance.

Design

Systematic review and meta-analysis of external validation studies.

Data sources

Medline, Embase, Web of Science, Scopus, and Europe PMC, from 15 October 2014 to 15 May 2023.

Eligibility criteria for selecting studies

All external validation studies of the performance of ADNEX, with any study design and any study population of patients with an adnexal mass. Two independent reviewers extracted the data. Disagreements were resolved by discussion. Reporting quality of the studies was scored with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) reporting guideline, and methodological conduct and risk of bias with PROBAST (Prediction model Risk Of Bias Assessment Tool). Random effects meta-analysis of the area under the receiver operating characteristic curve (AUC), sensitivity and specificity at the 10% risk of malignancy threshold, and net benefit and relative utility at the 10% risk of malignancy threshold were performed.

Results

47 studies (17 007 tumours) were included, with a median study sample size of 261 (range 24-4905). On average, 61% of TRIPOD items were reported. Handling of missing data, justification of sample size, and model calibration were rarely described. 91% of validations were at high risk of bias, mainly because of the unexplained exclusion of incomplete cases, small sample size, or no assessment of calibration. The summary AUC to distinguish benign from malignant tumours in patients who underwent surgery was 0.93 (95% confidence interval 0.92 to 0.94, 95% prediction interval 0.85 to 0.98) for ADNEX with the serum biomarker, cancer antigen 125 (CA125), as a predictor (9202 tumours, 43 centres, 18 countries, and 21 studies) and 0.93 (95% confidence interval 0.91 to 0.94, 95% prediction interval 0.85 to 0.98) for ADNEX without CA125 (6309 tumours, 31 centres, 13 countries, and 12 studies). The estimated probability that the model has use clinically in a new centre was 95% (with CA125) and 91% (without CA125). When restricting analysis to studies with a low risk of bias, summary AUC values were 0.93 (with CA125) and 0.91 (without CA125), and estimated probabilities that the model has use clinically were 89% (with CA125) and 87% (without CA125).

Conclusions

The results of the meta-analysis indicated that ADNEX performed well in distinguishing between benign and malignant tumours in populations from different countries and settings, regardless of whether the serum biomarker, CA125, was used as a predictor. A key limitation was that calibration was rarely assessed.

Systematic review registration

PROSPERO CRD42022373182.

The application and use of artificial intelligence in cancer nursing: A systematic review

PURPOSE

Artificial Intelligence is being applied in oncology to improve patient and service outcomes. Yet, there is a limited understanding of how these advanced computational techniques are employed in cancer nursing to inform clinical practice. This review aimed to identify and synthesise evidence on artificial intelligence in cancer nursing.

METHODS

CINAHL, MEDLINE, PsycINFO, and PubMed were searched using key terms between January 2010 and December 2022. Titles, abstracts, and then full texts were screened against eligibility criteria, resulting in twenty studies being included. Critical appraisal was undertaken, and relevant data extracted and analysed. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines were followed.

RESULTS

Artificial intelligence was used in numerous areas including breast, colorectal, liver, and ovarian cancer care among others. Algorithms were trained and tested on primary and secondary datasets to build predictive models of health problems related to cancer. Studies reported this led to improvements in the accuracy of predicting health outcomes or identifying variables that improved outcome prediction. While nurses led most studies, few deployed an artificial intelligence based digital tool with cancer nurses in a real-world setting as studies largely focused on developing and validating predictive models.

CONCLUSION

Electronic cancer nursing datasets should be established to enable artificial intelligence techniques to be tested and if effective implemented in digital prediction and other AI-based tools. Cancer nurses need more education on machine learning and natural language processing, so they can lead and contribute to artificial intelligence developments in oncology.

Machine learning prediction model of major adverse outcomes after pediatric congenital heart surgery-a retrospective cohort study

BACKGROUND

Major adverse postoperative outcomes (APOs) can greatly affect mortality, hospital stay, care management and planning, and quality of life. This study aimed to evaluate the performance of five machine learning (ML) algorithms for predicting four major APOs after pediatric congenital heart surgery and their clinically meaningful model interpretations.

METHODS

Between August 2014 and December 2021, 23,000 consecutive pediatric patients receiving congenital heart surgery were enrolled. Based on the split date of 1 January 2019, we selected 13,927 participants for the training cohort, and 9,073 participants for the testing cohort. Four predefined major APOs including low cardiac output syndrome (LCOS), pneumonia, renal failure, and deep venous thrombosis (DVT) were investigated. 39 clinical and laboratory features were inputted in five ML models: light gradient boosting machine (LightGBM), logistic regression (LR), support vector machine, random forest, and CatBoost. The performance and interpretations of ML models were evaluated using the area under the receiver operating characteristic curve (AUC) and Shapley Additive Explanations (SHAP).

RESULTS

In the training cohort, CatBoost algorithms outperformed others with the mean AUCs of 0.908 for LCOS and 0.957 for renal failure, while LightGBM and LR achieved the best mean AUCs of 0.886 for pneumonia and 0.942 for DVT, respectively. In the testing cohort, the best-performing ML model for each major APOs with the following mean AUCs: LCOS (LightGBM), 0.893 (95% confidence interval (CI), 0.884-0.895); pneumonia (LR), 0.929 (95% CI, 0.926-0.931); renal failure (LightGBM), 0.963 (95% CI, 0.947-0.979), and DVT (LightGBM), 0.970 (95% CI, 0.953-0.982). The performance of ML models using only clinical variables was slightly lower than those using combined data, with the mean AUCs of 0.873 for LCOS, 0.894 for pneumonia, 0.953 for renal failure, and 0.933 for DVT. The SHAP showed that mechanical ventilation time was the most important contributor of four major APOs.

CONCLUSIONS

In pediatric congenital heart surgery, the established ML model can accurately predict the risk of four major APOs, providing reliable interpretations for high-risk contributor identification and informed clinical decisions making.

Open science practices need substantial improvement in prognostic model studies in oncology using machine learning

OBJECTIVE

To describe the frequency of open science practices in a contemporary sample of studies developing prognostic models using machine learning methods in the field of oncology.

STUDY DESIGN AND SETTING

We conducted a systematic review, searching the MEDLINE database between December 1, 2022, and December 31, 2022, for studies developing a multivariable prognostic model using machine learning methods (as defined by the authors) in oncology. Two authors independently screened records and extracted open science practices.

RESULTS

We identified 46 publications describing the development of a multivariable prognostic model. The adoption of open science principles was poor. Only one study reported availability of a study protocol, and only one study was registered. Funding statements and conflicts of interest statements were common. Thirty-five studies (76%) provided data sharing statements, with 21 (46%) indicating data were available on request to the authors and seven declaring data sharing was not applicable. Two studies (4%) shared data. Only 12 studies (26%) provided code sharing statements, including 2 (4%) that indicated the code was available on request to the authors. Only 11 studies (24%) provided sufficient information to allow their model to be used in practice. The use of reporting guidelines was rare: eight studies (18%) mentioning using a reporting guideline, with 4 (10%) using the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis Or Diagnosis statement, 1 (2%) using Minimum Information About Clinical Artificial Intelligence Modeling and Consolidated Standards Of Reporting Trials-Artificial Intelligence, 1 (2%) using Strengthening The Reporting Of Observational Studies In Epidemiology, 1 (2%) using Standards for Reporting Diagnostic Accuracy Studies, and 1 (2%) using Transparent Reporting of Evaluations with Nonrandomized Designs.

CONCLUSION

The adoption of open science principles in oncology studies developing prognostic models using machine learning methods is poor. Guidance and an increased awareness of benefits and best practices of open science are needed for prediction research in oncology.

Making the Improbable Possible: Generalizing Models Designed for a Syndrome-Based, Heterogeneous Patient Landscape

Syndromic conditions, such as sepsis, are commonly encountered in the intensive care unit. Although these conditions are easy for clinicians to grasp, these conditions may limit the performance of machine-learning algorithms. Individual hospital practice patterns may limit external generalizability. Data missingness is another barrier to optimal algorithm performance and various strategies exist to mitigate this. Recent advances in data science, such as transfer learning, conformal prediction, and continual learning, may improve generalizability of machine-learning algorithms in critically ill patients. Randomized trials with these approaches are indicated to demonstrate improvements in patient-centered outcomes at this point.

Prognostic models for predicting clinical disease progression, worsening and activity in people with multiple sclerosis

BACKGROUND

Multiple sclerosis (MS) is a chronic inflammatory disease of the central nervous system that affects millions of people worldwide. The disease course varies greatly across individuals and many disease-modifying treatments with different safety and efficacy profiles have been developed recently. Prognostic models evaluated and shown to be valid in different settings have the potential to support people with MS and their physicians during the decision-making process for treatment or disease/life management, allow stratified and more precise interpretation of interventional trials, and provide insights into disease mechanisms. Many researchers have turned to prognostic models to help predict clinical outcomes in people with MS; however, to our knowledge, no widely accepted prognostic model for MS is being used in clinical practice yet.

OBJECTIVES

To identify and summarise multivariable prognostic models, and their validation studies for quantifying the risk of clinical disease progression, worsening, and activity in adults with MS.

SEARCH METHODS

We searched MEDLINE, Embase, and the Cochrane Database of Systematic Reviews from January 1996 until July 2021. We also screened the reference lists of included studies and relevant reviews, and references citing the included studies.

SELECTION CRITERIA

We included all statistically developed multivariable prognostic models aiming to predict clinical disease progression, worsening, and activity, as measured by disability, relapse, conversion to definite MS, conversion to progressive MS, or a composite of these in adult individuals with MS. We also included any studies evaluating the performance of (i.e. validating) these models. There were no restrictions based on language, data source, timing of prognostication, or timing of outcome.

DATA COLLECTION AND ANALYSIS

Pairs of review authors independently screened titles/abstracts and full texts, extracted data using a piloted form based on the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS), assessed risk of bias using the Prediction Model Risk Of Bias Assessment Tool (PROBAST), and assessed reporting deficiencies based on the checklist items in Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD). The characteristics of the included models and their validations are described narratively. We planned to meta-analyse the discrimination and calibration of models with at least three external validations outside the model development study but no model met this criterion. We summarised between-study heterogeneity narratively but again could not perform the planned meta-regression.

MAIN RESULTS

We included 57 studies, from which we identified 75 model developments, 15 external validations corresponding to only 12 (16%) of the models, and six author-reported validations. Only two models were externally validated multiple times. None of the identified external validations were performed by researchers independent of those that developed the model. The outcome was related to disease progression in 39 (41%), relapses in 8 (8%), conversion to definite MS in 17 (18%), and conversion to progressive MS in 27 (28%) of the 96 models or validations. The disease and treatment-related characteristics of included participants, and definitions of considered predictors and outcome, were highly heterogeneous amongst the studies. Based on the publication year, we observed an increase in the percent of participants on treatment, diversification of the diagnostic criteria used, an increase in consideration of biomarkers or treatment as predictors, and increased use of machine learning methods over time. Usability and reproducibility All identified models contained at least one predictor requiring the skills of a medical specialist for measurement or assessment. Most of the models (44; 59%) contained predictors that require specialist equipment likely to be absent from primary care or standard hospital settings. Over half (52%) of the developed models were not accompanied by model coefficients, tools, or instructions, which hinders their application, independent validation or reproduction. The data used in model developments were made publicly available or reported to be available on request only in a few studies (two and six, respectively). Risk of bias We rated all but one of the model developments or validations as having high overall risk of bias. The main reason for this was the statistical methods used for the development or evaluation of prognostic models; we rated all but two of the included model developments or validations as having high risk of bias in the analysis domain. None of the model developments that were externally validated or these models' external validations had low risk of bias. There were concerns related to applicability of the models to our research question in over one-third (38%) of the models or their validations. Reporting deficiencies Reporting was poor overall and there was no observable increase in the quality of reporting over time. The items that were unclearly reported or not reported at all for most of the included models or validations were related to sample size justification, blinding of outcome assessors, details of the full model or how to obtain predictions from it, amount of missing data, and treatments received by the participants. Reporting of preferred model performance measures of discrimination and calibration was suboptimal.

AUTHORS' CONCLUSIONS

The current evidence is not sufficient for recommending the use of any of the published prognostic prediction models for people with MS in clinical routine today due to lack of independent external validations. The MS prognostic research community should adhere to the current reporting and methodological guidelines and conduct many more state-of-the-art external validation studies for the existing or newly developed models.

Sample size requirements are not being considered in studies developing prediction models for binary outcomes: a systematic review

BACKGROUND

Having an appropriate sample size is important when developing a clinical prediction model. We aimed to review how sample size is considered in studies developing a prediction model for a binary outcome.

METHODS

We searched PubMed for studies published between 01/07/2020 and 30/07/2020 and reviewed the sample size calculations used to develop the prediction models. Using the available information, we calculated the minimum sample size that would be needed to estimate overall risk and minimise overfitting in each study and summarised the difference between the calculated and used sample size.

RESULTS

A total of 119 studies were included, of which nine studies provided sample size justification (8%). The recommended minimum sample size could be calculated for 94 studies: 73% (95% CI: 63-82%) used sample sizes lower than required to estimate overall risk and minimise overfitting including 26% studies that used sample sizes lower than required to estimate overall risk only. A similar number of studies did not meet the ≥ 10EPV criteria (75%, 95% CI: 66-84%). The median deficit of the number of events used to develop a model was 75 [IQR: 234 lower to 7 higher]) which reduced to 63 if the total available data (before any data splitting) was used [IQR:225 lower to 7 higher]. Studies that met the minimum required sample size had a median c-statistic of 0.84 (IQR:0.80 to 0.9) and studies where the minimum sample size was not met had a median c-statistic of 0.83 (IQR: 0.75 to 0.9). Studies that met the ≥ 10 EPP criteria had a median c-statistic of 0.80 (IQR: 0.73 to 0.84).

CONCLUSIONS

Prediction models are often developed with no sample size calculation, as a consequence many are too small to precisely estimate the overall risk. We encourage researchers to justify, perform and report sample size calculations when developing a prediction model.

Artificial intelligence in lung cancer diagnostic imaging: a review of the reporting and conduct of research published 2018-2019

Objective

This study aimed to describe the methodologies used to develop and evaluate models that use artificial intelligence (AI) to analyse lung images in order to detect, segment (outline borders of), or classify pulmonary nodules as benign or malignant.

Methods

In October 2019, we systematically searched the literature for original studies published between 2018 and 2019 that described prediction models using AI to evaluate human pulmonary nodules on diagnostic chest images. Two evaluators independently extracted information from studies, such as study aims, sample size, AI type, patient characteristics, and performance. We summarised data descriptively.

Results

The review included 153 studies: 136 (89%) development-only studies, 12 (8%) development and validation, and 5 (3%) validation-only. CT scans were the most common type of image type used (83%), often acquired from public databases (58%). Eight studies (5%) compared model outputs with biopsy results. 41 studies (26.8%) reported patient characteristics. The models were based on different units of analysis, such as patients, images, nodules, or image slices or patches.

Conclusion

The methods used to develop and evaluate prediction models using AI to detect, segment, or classify pulmonary nodules in medical imaging vary, are poorly reported, and therefore difficult to evaluate. Transparent and complete reporting of methods, results and code would fill the gaps in information we observed in the study publications.

Advances in knowledge

We reviewed the methodology of AI models detecting nodules on lung images and found that the models were poorly reported and had no description of patient characteristics, with just a few comparing models' outputs with biopsies results. When lung biopsy is not available, lung-RADS could help standardise the comparisons between the human radiologist and the machine. The field of radiology should not give up principles from the diagnostic accuracy studies, such as the choice for the correct ground truth, just because AI is used. Clear and complete reporting of the reference standard used would help radiologists trust in the performance that AI models claim to have. This review presents clear recommendations about the essential methodological aspects of diagnostic models that should be incorporated in studies using AI to help detect or segmentate lung nodules. The manuscript also reinforces the need for more complete and transparent reporting, which can be helped using the recommended reporting guidelines.

Development and internal-external validation of statistical and machine learning models for breast cancer prognostication: cohort study

OBJECTIVE

To develop a clinically useful model that estimates the 10 year risk of breast cancer related mortality in women (self-reported female sex) with breast cancer of any stage, comparing results from regression and machine learning approaches.

DESIGN

Population based cohort study.

SETTING

QResearch primary care database in England, with individual level linkage to the national cancer registry, Hospital Episodes Statistics, and national mortality registers.

PARTICIPANTS

141 765 women aged 20 years and older with a diagnosis of invasive breast cancer between 1 January 2000 and 31 December 2020.

MAIN OUTCOME MEASURES

Four model building strategies comprising two regression (Cox proportional hazards and competing risks regression) and two machine learning (XGBoost and an artificial neural network) approaches. Internal-external cross validation was used for model evaluation. Random effects meta-analysis that pooled estimates of discrimination and calibration metrics, calibration plots, and decision curve analysis were used to assess model performance, transportability, and clinical utility.

RESULTS

During a median 4.16 years (interquartile range 1.76-8.26) of follow-up, 21 688 breast cancer related deaths and 11 454 deaths from other causes occurred. Restricting to 10 years maximum follow-up from breast cancer diagnosis, 20 367 breast cancer related deaths occurred during a total of 688 564.81 person years. The crude breast cancer mortality rate was 295.79 per 10 000 person years (95% confidence interval 291.75 to 299.88). Predictors varied for each regression model, but both Cox and competing risks models included age at diagnosis, body mass index, smoking status, route to diagnosis, hormone receptor status, cancer stage, and grade of breast cancer. The Cox model's random effects meta-analysis pooled estimate for Harrell's C index was the highest of any model at 0.858 (95% confidence interval 0.853 to 0.864, and 95% prediction interval 0.843 to 0.873). It appeared acceptably calibrated on calibration plots. The competing risks regression model had good discrimination: pooled Harrell's C index 0.849 (0.839 to 0.859, and 0.821 to 0.876, and evidence of systematic miscalibration on summary metrics was lacking. The machine learning models had acceptable discrimination overall (Harrell's C index: XGBoost 0.821 (0.813 to 0.828, and 0.805 to 0.837); neural network 0.847 (0.835 to 0.858, and 0.816 to 0.878)), but had more complex patterns of miscalibration and more variable regional and stage specific performance. Decision curve analysis suggested that the Cox and competing risks regression models tested may have higher clinical utility than the two machine learning approaches.

CONCLUSION

In women with breast cancer of any stage, using the predictors available in this dataset, regression based methods had better and more consistent performance compared with machine learning approaches and may be worthy of further evaluation for potential clinical use, such as for stratified follow-up.

Overinterpretation of findings in machine learning prediction model studies in oncology: a systematic review

OBJECTIVES

In biomedical research, spin is the overinterpretation of findings, and it is a growing concern. To date, the presence of spin has not been evaluated in prognostic model research in oncology, including studies developing and validating models for individualized risk prediction.

STUDY DESIGN AND SETTING

We conducted a systematic review, searching MEDLINE and EMBASE for oncology-related studies that developed and validated a prognostic model using machine learning published between 1st January, 2019, and 5th September, 2019. We used existing spin frameworks and described areas of highly suggestive spin practices.

RESULTS

We included 62 publications (including 152 developed models; 37 validated models). Reporting was inconsistent between methods and the results in 27% of studies due to additional analysis and selective reporting. Thirty-two studies (out of 36 applicable studies) reported comparisons between developed models in their discussion and predominantly used discrimination measures to support their claims (78%). Thirty-five studies (56%) used an overly strong or leading word in their title, abstract, results, discussion, or conclusion.

CONCLUSION

The potential for spin needs to be considered when reading, interpreting, and using studies that developed and validated prognostic models in oncology. Researchers should carefully report their prognostic model research using words that reflect their actual results and strength of evidence.

Machine learning does not outperform traditional statistical modelling for kidney allograft failure prediction

Machine learning (ML) models have recently shown potential for predicting kidney allograft outcomes. However, their ability to outperform traditional approaches remains poorly investigated. Therefore, using large cohorts of kidney transplant recipients from 14 centers worldwide, we developed ML-based prediction models for kidney allograft survival and compared their prediction performances to those achieved by a validated Cox-Based Prognostication System (CBPS). In a French derivation cohort of 4000 patients, candidate determinants of allograft failure including donor, recipient and transplant-related parameters were used as predictors to develop tree-based models (RSF, RSF-ERT, CIF), Support Vector Machine models (LK-SVM, AK-SVM) and a gradient boosting model (XGBoost). Models were externally validated with cohorts of 2214 patients from Europe, 1537 from North America, and 671 from South America. Among these 8422 kidney transplant recipients, 1081 (12.84%) lost their grafts after a median post-transplant follow-up time of 6.25 years (Inter Quartile Range 4.33-8.73). At seven years post-risk evaluation, the ML models achieved a C-index of 0.788 (95% bootstrap percentile confidence interval 0.736-0.833), 0.779 (0.724-0.825), 0.786 (0.735-0.832), 0.527 (0.456-0.602), 0.704 (0.648-0.759) and 0.767 (0.711-0.815) for RSF, RSF-ERT, CIF, LK-SVM, AK-SVM and XGBoost respectively, compared with 0.808 (0.792-0.829) for the CBPS. In validation cohorts, ML models' discrimination performances were in a similar range of those of the CBPS. Calibrations of the ML models were similar or less accurate than those of the CBPS. Thus, when using a transparent methodological pipeline in validated international cohorts, ML models, despite overall good performances, do not outperform a traditional CBPS in predicting kidney allograft failure. Hence, our current study supports the continued use of traditional statistical approaches for kidney graft prognostication.

Evaluation of Risk of Bias in Neuroimaging-Based Artificial Intelligence Models for Psychiatric Diagnosis: A Systematic Review

IMPORTANCE

Neuroimaging-based artificial intelligence (AI) diagnostic models have proliferated in psychiatry. However, their clinical applicability and reporting quality (ie, feasibility) for clinical practice have not been systematically evaluated.

OBJECTIVE

To systematically assess the risk of bias (ROB) and reporting quality of neuroimaging-based AI models for psychiatric diagnosis.

EVIDENCE REVIEW

PubMed was searched for peer-reviewed, full-length articles published between January 1, 1990, and March 16, 2022. Studies aimed at developing or validating neuroimaging-based AI models for clinical diagnosis of psychiatric disorders were included. Reference lists were further searched for suitable original studies. Data extraction followed the CHARMS (Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies) and PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-analyses) guidelines. A closed-loop cross-sequential design was used for quality control. The PROBAST (Prediction Model Risk of Bias Assessment Tool) and modified CLEAR (Checklist for Evaluation of Image-Based Artificial Intelligence Reports) benchmarks were used to systematically evaluate ROB and reporting quality.

FINDINGS

A total of 517 studies presenting 555 AI models were included and evaluated. Of these models, 461 (83.1%; 95% CI, 80.0%-86.2%) were rated as having a high overall ROB based on the PROBAST. The ROB was particular high in the analysis domain, including inadequate sample size (398 of 555 models [71.7%; 95% CI, 68.0%-75.6%]), poor model performance examination (with 100% of models lacking calibration examination), and lack of handling data complexity (550 of 555 models [99.1%; 95% CI, 98.3%-99.9%]). None of the AI models was perceived to be applicable to clinical practices. Overall reporting completeness (ie, number of reported items/number of total items) for the AI models was 61.2% (95% CI, 60.6%-61.8%), and the completeness was poorest for the technical assessment domain with 39.9% (95% CI, 38.8%-41.1%).

CONCLUSIONS AND RELEVANCE

This systematic review found that the clinical applicability and feasibility of neuroimaging-based AI models for psychiatric diagnosis were challenged by a high ROB and poor reporting quality. Particularly in the analysis domain, ROB in AI diagnostic models should be addressed before clinical application.

Statistical models versus machine learning for competing risks: development and validation of prognostic models

BACKGROUND

In health research, several chronic diseases are susceptible to competing risks (CRs). Initially, statistical models (SM) were developed to estimate the cumulative incidence of an event in the presence of CRs. As recently there is a growing interest in applying machine learning (ML) for clinical prediction, these techniques have also been extended to model CRs but literature is limited. Here, our aim is to investigate the potential role of ML versus SM for CRs within non-complex data (small/medium sample size, low dimensional setting).

METHODS

A dataset with 3826 retrospectively collected patients with extremity soft-tissue sarcoma (eSTS) and nine predictors is used to evaluate model-predictive performance in terms of discrimination and calibration. Two SM (cause-specific Cox, Fine-Gray) and three ML techniques are compared for CRs in a simple clinical setting. ML models include an original partial logistic artificial neural network for CRs (PLANNCR original), a PLANNCR with novel specifications in terms of architecture (PLANNCR extended), and a random survival forest for CRs (RSFCR). The clinical endpoint is the time in years between surgery and disease progression (event of interest) or death (competing event). Time points of interest are 2, 5, and 10 years.

RESULTS

Based on the original eSTS data, 100 bootstrapped training datasets are drawn. Performance of the final models is assessed on validation data (left out samples) by employing as measures the Brier score and the Area Under the Curve (AUC) with CRs. Miscalibration (absolute accuracy error) is also estimated. Results show that the ML models are able to reach a comparable performance versus the SM at 2, 5, and 10 years regarding both Brier score and AUC (95% confidence intervals overlapped). However, the SM are frequently better calibrated.

CONCLUSIONS

Overall, ML techniques are less practical as they require substantial implementation time (data preprocessing, hyperparameter tuning, computational intensity), whereas regression methods can perform well without the additional workload of model training. As such, for non-complex real life survival data, these techniques should only be applied complementary to SM as exploratory tools of model's performance. More attention to model calibration is urgently needed.

Prediction and diagnosis of chronic kidney disease development and progression using machine-learning: Protocol for a systematic review and meta-analysis of reporting standards and model performance

Chronic Kidney disease (CKD) is an important yet under-recognized contributor to morbidity and mortality globally. Machine-learning (ML) based decision support tools have been developed across many aspects of CKD care. Notably, algorithms developed in the prediction and diagnosis of CKD development and progression may help to facilitate early disease prevention, assist with early planning of renal replacement therapy, and offer potential clinical and economic benefits to patients and health systems. Clinical implementation can be affected by the uncertainty surrounding the methodological rigor and performance of ML-based models. This systematic review aims to evaluate the application of prognostic and diagnostic ML tools in CKD development and progression. The protocol has been prepared using the Preferred Items for Systematic Review and Meta-analysis Protocols (PRISMA-P) guidelines. The systematic review protocol for CKD prediction and diagnosis have been registered with the International Prospective Register of Systematic Reviews (PROSPERO) (CRD42022356704, CRD42022372378). A systematic search will be undertaken of PubMed, Embase, the Cochrane Central Register of Controlled Trials (CENTRAL), the Web of Science, and the IEEE Xplore digital library. Studies in which ML has been applied to predict and diagnose CKD development and progression will be included. The primary outcome will be the comparison of the performance of ML-based models with non-ML-based models. Secondary analysis will consist of model use cases, model construct, and model reporting quality. This systematic review will offer valuable insight into the performance and reporting quality of ML-based models in CKD diagnosis and prediction. This will inform clinicians and technical specialists of the current development of ML in CKD care, as well as direct future model development and standardization.

Systematic review identifies the design and methodological conduct of studies on machine learning-based prediction models

BACKGROUND AND OBJECTIVES

We sought to summarize the study design, modelling strategies, and performance measures reported in studies on clinical prediction models developed using machine learning techniques.

METHODS

We search PubMed for articles published between 01/01/2018 and 31/12/2019, describing the development or the development with external validation of a multivariable prediction model using any supervised machine learning technique. No restrictions were made based on study design, data source, or predicted patient-related health outcomes.

RESULTS

We included 152 studies, 58 (38.2% [95% CI 30.8-46.1]) were diagnostic and 94 (61.8% [95% CI 53.9-69.2]) prognostic studies. Most studies reported only the development of prediction models (n = 133, 87.5% [95% CI 81.3-91.8]), focused on binary outcomes (n = 131, 86.2% [95% CI 79.8-90.8), and did not report a sample size calculation (n = 125, 82.2% [95% CI 75.4-87.5]). The most common algorithms used were support vector machine (n = 86/522, 16.5% [95% CI 13.5-19.9]) and random forest (n = 73/522, 14% [95% CI 11.3-17.2]). Values for area under the Receiver Operating Characteristic curve ranged from 0.45 to 1.00. Calibration metrics were often missed (n = 494/522, 94.6% [95% CI 92.4-96.3]).

CONCLUSION

Our review revealed that focus is required on handling of missing values, methods for internal validation, and reporting of calibration to improve the methodological conduct of studies on machine learning-based prediction models.

SYSTEMATIC REVIEW REGISTRATION

PROSPERO, CRD42019161764.

Methodological guidance for the evaluation and updating of clinical prediction models: a systematic review

BACKGROUND

Clinical prediction models are often not evaluated properly in specific settings or updated, for instance, with information from new markers. These key steps are needed such that models are fit for purpose and remain relevant in the long-term. We aimed to present an overview of methodological guidance for the evaluation (i.e., validation and impact assessment) and updating of clinical prediction models.

METHODS

We systematically searched nine databases from January 2000 to January 2022 for articles in English with methodological recommendations for the post-derivation stages of interest. Qualitative analysis was used to summarize the 70 selected guidance papers.

RESULTS

Key aspects for validation are the assessment of statistical performance using measures for discrimination (e.g., C-statistic) and calibration (e.g., calibration-in-the-large and calibration slope). For assessing impact or usefulness in clinical decision-making, recent papers advise using decision-analytic measures (e.g., the Net Benefit) over simplistic classification measures that ignore clinical consequences (e.g., accuracy, overall Net Reclassification Index). Commonly recommended methods for model updating are recalibration (i.e., adjustment of intercept or baseline hazard and/or slope), revision (i.e., re-estimation of individual predictor effects), and extension (i.e., addition of new markers). Additional methodological guidance is needed for newer types of updating (e.g., meta-model and dynamic updating) and machine learning-based models.

CONCLUSION

Substantial guidance was found for model evaluation and more conventional updating of regression-based models. An important development in model evaluation is the introduction of a decision-analytic framework for assessing clinical usefulness. Consensus is emerging on methods for model updating.

Reporting and risk of bias of prediction models based on machine learning methods in preterm birth: A systematic review

INTRODUCTION

There was limited evidence on the quality of reporting and methodological quality of prediction models using machine learning methods in preterm birth. This systematic review aimed to assess the reporting quality and risk of bias of a machine learning-based prediction model in preterm birth.

MATERIAL AND METHODS

We conducted a systematic review, searching the PubMed, Embase, the Cochrane Library, China National Knowledge Infrastructure, China Biology Medicine disk, VIP Database, and WanFang Data from inception to September 27, 2021. Studies that developed (validated) a prediction model using machine learning methods in preterm birth were included. We used the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement and Prediction model Risk of Bias Assessment Tool (PROBAST) to evaluate the reporting quality and the risk of bias of included studies, respectively. Findings were summarized using descriptive statistics and visual plots. The protocol was registered in PROSPERO (no. CRD 42022301623).

RESULTS

Twenty-nine studies met the inclusion criteria, with 24 development-only studies and 5 development-with-validation studies. Overall, TRIPOD adherence per study ranged from 17% to 79%, with a median adherence of 49%. The reporting of title, abstract, blinding of predictors, sample size justification, explanation of model, and model performance were mostly poor, with TRIPOD adherence ranging from 4% to 17%. For all included studies, 79% had a high overall risk of bias, and 21% had an unclear overall risk of bias. The analysis domain was most commonly rated as high risk of bias in included studies, mainly as a result of small effective sample size, selection of predictors based on univariable analysis, and lack of calibration evaluation.

CONCLUSIONS

Reporting and methodological quality of machine learning-based prediction models in preterm birth were poor. It is urgent to improve the design, conduct, and reporting of such studies to boost the application of machine learning-based prediction models in preterm birth in clinical practice.

Open Source Repository and Online Calculator of Prediction Models for Diagnosis and Prognosis in Oncology

(1) Background: The main aim was to develop a prototype application that would serve as an open-source repository for a curated subset of predictive and prognostic models regarding oncology, and provide a user-friendly interface for the included models to allow online calculation. The focus of the application is on providing physicians and health professionals with patient-specific information regarding treatment plans, survival rates, and side effects for different expected treatments. (2) Methods: The primarily used models were the ones developed by our research group in the past. This selection was completed by a number of models, addressing the same cancer types but focusing on other outcomes that were selected based on a literature search in PubMed and Medline databases. All selected models were publicly available and had been validated TRIPOD (Transparent Reporting of studies on prediction models for Individual Prognosis Or Diagnosis) type 3 or 2b. (3) Results: The open source repository currently incorporates 18 models from different research groups, evaluated on datasets from different countries. Model types included logistic regression, Cox regression, and recursive partition analysis (decision trees). (4) Conclusions: An application was developed to enable physicians to complement their clinical judgment with user-friendly patient-specific predictions using models that have received internal/external validation. Additionally, this platform enables researchers to display their work, enhancing the use and exposure of their models.

Just How Confident Can We Be in Predicting Sports Injuries? A Systematic Review of the Methodological Conduct and Performance of Existing Musculoskeletal Injury Prediction Models in Sport

BACKGROUND

An increasing number of musculoskeletal injury prediction models are being developed and implemented in sports medicine. Prediction model quality needs to be evaluated so clinicians can be informed of their potential usefulness.

OBJECTIVE

To evaluate the methodological conduct and completeness of reporting of musculoskeletal injury prediction models in sport.

METHODS

A systematic review was performed from inception to June 2021. Studies were included if they: (1) predicted sport injury; (2) used regression, machine learning, or deep learning models; (3) were written in English; (4) were peer reviewed.

RESULTS

Thirty studies (204 models) were included; 60% of studies utilized only regression methods, 13% only machine learning, and 27% both regression and machine learning approaches. All studies developed a prediction model and no studies externally validated a prediction model. Two percent of models (7% of studies) were low risk of bias and 98% of models (93% of studies) were high or unclear risk of bias. Three studies (10%) performed an a priori sample size calculation; 14 (47%) performed internal validation. Nineteen studies (63%) reported discrimination and two (7%) reported calibration. Four studies (13%) reported model equations for statistical predictions and no machine learning studies reported code or hyperparameters.

CONCLUSION

Existing sport musculoskeletal injury prediction models were poorly developed and have a high risk of bias. No models could be recommended for use in practice. The majority of models were developed with small sample sizes, had inadequate assessment of model performance, and were poorly reported. To create clinically useful sports musculoskeletal injury prediction models, considerable improvements in methodology and reporting are urgently required.

Neural Networks for Survival Prediction in Medicine Using Prognostic Factors: A Review and Critical Appraisal

Survival analysis deals with the expected duration of time until one or more events of interest occur. Time to the event of interest may be unobserved, a phenomenon commonly known as right censoring, which renders the analysis of these data challenging. Over the years, machine learning algorithms have been developed and adapted to right-censored data. Neural networks have been repeatedly employed to build clinical prediction models in healthcare with a focus on cancer and cardiology. We present the first ever attempt at a large-scale review of survival neural networks (SNNs) with prognostic factors for clinical prediction in medicine. This work provides a comprehensive understanding of the literature (24 studies from 1990 to August 2021, global search in PubMed). Relevant manuscripts are classified as methodological/technical (novel methodology or new theoretical model; 13 studies) or applications (11 studies). We investigate how researchers have used neural networks to fit survival data for prediction. There are two methodological trends: either time is added as part of the input features and a single output node is specified, or multiple output nodes are defined for each time interval. A critical appraisal of model aspects that should be designed and reported more carefully is performed. We identify key characteristics of prediction models (i.e., number of patients/predictors, evaluation measures, calibration), and compare ANN's predictive performance to the Cox proportional hazards model. The median sample size is 920 patients, and the median number of predictors is 7. Major findings include poor reporting (e.g., regarding missing data, hyperparameters) as well as inaccurate model development/validation. Calibration is neglected in more than half of the studies. Cox models are not developed to their full potential and claims for the performance of SNNs are exaggerated. Light is shed on the current state of art of SNNs in medicine with prognostic factors. Recommendations are made for the reporting of clinical prediction models. Limitations are discussed, and future directions are proposed for researchers who seek to develop existing methodology.

Integration of Machine Learning Algorithms and Discrete-Event Simulation for the Cost of Healthcare Resources

A healthcare resource allocation generally plays a vital role in the number of patients treated (p_nt) and the patient waiting time (w_t) in healthcare institutions. This study aimed to estimate p_nt and w_t as output variables by considering the number of healthcare resources employed and analyze the cost of health resources to the hospital depending on the cost coefficient (δ_i) in an emergency department (ED). The integration of the discrete-event simulation (DES) model and machine learning (ML) algorithms, namely random forest (RF), gradient boosting (GB), and AdaBoost (AB), was used to calculate the estimation of the output variables depending on the δ_i of resources cost. The AB algorithm performed best in almost all scenarios based on the results of the analysis. According to the AB algorithm based on the δ_0.0, δ_0.1, δ_0.2, and δ_0.3, the accuracy data were calculated as 0.9838, 0.9843, 0.9838, and 0.9846 for p_nt; 0.9514, 0.9517, 0.9514, and 0.9514 for w_t, respectively in the training stage. The GB algorithm had the best performance value, except for the results of the δ_0.2 (AB had a better accuracy at 0.8709 based on the value of δ_0.2 for p_nt) in the test stage. According to the AB algorithm based on the δ_0.0, δ_0.1, δ_0.2, and δ_0.3, the accuracy data were calculated as 0.7956, 0.9298, 0.8288, and 0.7394 for p_nt; 0.8820, 0.8821, 0.8819, and 0.8818 for w_t in the training phase, respectively. All scenarios created by the δ_i coefficient should be preferred for ED since the income provided by the p_nt value to the hospital was more than the cost of healthcare resources. On the contrary, the w_t estimation results of ML algorithms based on the δ_i coefficient differed. Although w_t values in all ML algorithms with δ_0.0 and δ_0.1 coefficients reduced the cost of the hospital, w_t values based on δ_0.2 and δ_0.3 increased the cost of the hospital.

Lessons learned from developing a COVID-19 algorithm governance framework in Aotearoa New Zealand

Aotearoa New Zealand's response to the COVID-19 pandemic has included the use of algorithms that could aid decision making. Te Pokapū Hātepe o Aotearoa, the New Zealand Algorithm Hub, was established to evaluate and host COVID-19 related models and algorithms, and provide a central and secure infrastructure to support the country's pandemic response. A critical aspect of the Hub was the formation of an appropriate governance group to ensure that algorithms being deployed underwent cross-disciplinary scrutiny prior to being made available for quick and safe implementation. This framework necessarily canvassed a broad range of perspectives, including from data science, clinical, Māori, consumer, ethical, public health, privacy, legal and governmental perspectives. To our knowledge, this is the first implementation of national algorithm governance of this type, building upon broad local and global discussion of guidelines in recent years. This paper describes the experiences and lessons learned through this process from the perspective of governance group members, emphasising the role of robust governance processes in building a high-trust platform that enables rapid translation of algorithms from research to practice.

Joint models for dynamic prediction in localised prostate cancer: a literature review

BACKGROUND

Prostate cancer is a very prevalent disease in men. Patients are monitored regularly during and after treatment with repeated assessment of prostate-specific antigen (PSA) levels. Prognosis of localised prostate cancer is generally good after treatment, and the risk of having a recurrence is usually estimated based on factors measured at diagnosis. Incorporating PSA measurements over time in a dynamic prediction joint model enables updates of patients' risk as new information becomes available. We review joint model strategies that have been applied to model time-dependent PSA trajectories to predict time-to-event outcomes in localised prostate cancer.

METHODS

We identify articles that developed joint models for prediction of localised prostate cancer recurrence over the last two decades. We report, compare, and summarise the methodological approaches and applications that use joint modelling accounting for two processes: the longitudinal model (PSA), and the time-to-event process (clinical failure). The methods explored differ in how they specify the association between these two processes.

RESULTS

Twelve relevant articles were identified. A range of methodological frameworks were found, and we describe in detail shared-parameter joint models (9 of 12, 75%) and joint latent class models (3 of 12, 25%). Within each framework, these articles presented model development, estimation of dynamic predictions and model validations.

CONCLUSIONS

Each framework has its unique principles with corresponding advantages and differing interpretations. Regardless of the framework used, dynamic prediction models enable real-time prediction of individual patient prognosis. They utilise all available longitudinal information, in addition to baseline prognostic risk factors, and are superior to traditional baseline-only prediction models.

Critical appraisal of artificial intelligence-based prediction models for cardiovascular disease

The medical field has seen a rapid increase in the development of artificial intelligence (AI)-based prediction models. With the introduction of such AI-based prediction model tools and software in cardiovascular patient care, the cardiovascular researcher and healthcare professional are challenged to understand the opportunities as well as the limitations of the AI-based predictions. In this article, we present 12 critical questions for cardiovascular health professionals to ask when confronted with an AI-based prediction model. We aim to support medical professionals to distinguish the AI-based prediction models that can add value to patient care from the AI that does not.