Calibration drift in regression and machine learning models for acute kidney injury

Predictive modelling for postoperative acute kidney injury: big data enhancing quality or the Emperor's new clothes?

The increased availability of large clinical datasets together with increasingly sophisticated computing power has facilitated development of numerous risk prediction models for various adverse perioperative outcomes, including acute kidney injury (AKI). The rationale for developing such models is straightforward. However, despite numerous purported benefits, the uptake of preoperative prediction models into clinical practice has been limited. Barriers to implementation of predictive models, including limitations in their discrimination and accuracy, as well as their ability to meaningfully impact clinical practice and patient outcomes, are increasingly recognised. Some of the purported benefits of predictive modelling, particularly when applied to postoperative AKI, might not fare well under detailed scrutiny. Future research should address existing limitations and seek to demonstrate both benefit to patients and value to healthcare systems from implementation of these models in clinical practice.

Preoperative risk prediction models for acute kidney injury after noncardiac surgery: an independent external validation cohort study

BACKGROUND

Numerous models have been developed to predict acute kidney injury (AKI) after noncardiac surgery, yet there is a lack of independent validation and comparison among them.

METHODS

We conducted a systematic literature search to review published risk prediction models for AKI after noncardiac surgery. An independent external validation was performed using a retrospective surgical cohort at a large Chinese hospital from January 2019 to October 2022. The cohort included patients undergoing a wide range of noncardiac surgeries with perioperative creatinine measurements. Postoperative AKI was defined according to the Kidney Disease Improving Global Outcomes creatinine criteria. Model performance was assessed in terms of discrimination (area under the receiver operating characteristic curve, AUROC), calibration (calibration plot), and clinical utility (net benefit), before and after model recalibration through intercept and slope updates. A sensitivity analysis was conducted by including patients without postoperative creatinine measurements in the validation cohort and categorising them as non-AKI cases.

RESULTS

Nine prediction models were evaluated, each with varying clinical and methodological characteristics, including the types of surgical cohorts used for model development, AKI definitions, and predictors. In the validation cohort involving 13,186 patients, 650 (4.9%) developed AKI. Three models demonstrated fair discrimination (AUROC between 0.71 and 0.75); other models had poor or failed discrimination. All models exhibited some miscalibration; five of the nine models were well-calibrated after intercept and slope updates. Decision curve analysis indicated that the three models with fair discrimination consistently provided a positive net benefit after recalibration. The results were confirmed in the sensitivity analysis.

CONCLUSIONS

We identified three models with fair discrimination and potential clinical utility after recalibration for assessing the risk of acute kidney injury after noncardiac surgery.

Artificial Intelligence in Perioperative Care: Opportunities and Challenges

Artificial intelligence (AI) applications have great potential to enhance perioperative care. This paper explores promising areas for AI in anesthesiology; expertise, stakeholders, and infrastructure for development; and barriers and challenges to implementation.

How Artificial Intelligence Will Transform Clinical Care, Research, and Trials for Inflammatory Bowel Disease

Artificial intelligence (AI) refers to computer-based methodologies that use data to teach a computer to solve pre-defined tasks; these methods can be applied to identify patterns in large multi-modal data sources. AI applications in inflammatory bowel disease (IBD) includes predicting response to therapy, disease activity scoring of endoscopy, drug discovery, and identifying bowel damage in images. As a complex disease with entangled relationships between genomics, metabolomics, microbiome, and the environment, IBD stands to benefit greatly from methodologies that can handle this complexity. We describe current applications, critical challenges, and propose future directions of AI in IBD.

Assessing calibration and bias of a deployed machine learning malnutrition prediction model within a large healthcare system

Malnutrition is a frequently underdiagnosed condition leading to increased morbidity, mortality, and healthcare costs. The Mount Sinai Health System (MSHS) deployed a machine learning model (MUST-Plus) to detect malnutrition upon hospital admission. However, in diverse patient groups, a poorly calibrated model may lead to misdiagnosis, exacerbating health care disparities. We explored the model's calibration across different variables and methods to improve calibration. Data from adult patients admitted to five MSHS hospitals from January 1, 2021 - December 31, 2022, were analyzed. We compared MUST-Plus prediction to the registered dietitian's formal assessment. Hierarchical calibration was assessed and compared between the recalibration sample (N = 49,562) of patients admitted between January 1, 2021 - December 31, 2022, and the hold-out sample (N = 17,278) of patients admitted between January 1, 2023 - September 30, 2023. Statistical differences in calibration metrics were tested using bootstrapping with replacement. Before recalibration, the overall model calibration intercept was -1.17 (95% CI: -1.20, -1.14), slope was 1.37 (95% CI: 1.34, 1.40), and Brier score was 0.26 (95% CI: 0.25, 0.26). Both weak and moderate measures of calibration were significantly different between White and Black patients and between male and female patients. Logistic recalibration significantly improved calibration of the model across race and gender in the hold-out sample. The original MUST-Plus model showed significant differences in calibration between White vs. Black patients. It also overestimated malnutrition in females compared to males. Logistic recalibration effectively reduced miscalibration across all patient subgroups. Continual monitoring and timely recalibration can improve model accuracy.

External validation of a paediatric Smart triage model for use in resource limited facilities

Models for digital triage of sick children at emergency departments of hospitals in resource poor settings have been developed. However, prior to their adoption, external validation should be performed to ensure their generalizability. We externally validated a previously published nine-predictor paediatric triage model (Smart Triage) developed in Uganda using data from two hospitals in Kenya. Both discrimination and calibration were assessed, and recalibration was performed by optimizing the intercept for classifying patients into emergency, priority, or non-urgent categories based on low-risk and high-risk thresholds. A total of 2539 patients were eligible at Hospital 1 and 2464 at Hospital 2, and 5003 for both hospitals combined; admission rates were 8.9%, 4.5%, and 6.8%, respectively. The model showed good discrimination, with area under the receiver-operator curve (AUC) of 0.826, 0.784 and 0.821, respectively. The pre-calibrated model at a low-risk threshold of 8% achieved a sensitivity of 93% (95% confidence interval, (CI):89%-96%), 81% (CI:74%-88%), and 89% (CI:85%-92%), respectively, and at a high-risk threshold of 40%, the model achieved a specificity of 86% (CI:84%-87%), 96% (CI:95%-97%), and 91% (CI:90%-92%), respectively. Recalibration improved the graphical fit, but new risk thresholds were required to optimize sensitivity and specificity.The Smart Triage model showed good discrimination on external validation but required recalibration to improve the graphical fit of the calibration plot. There was no change in the order of prioritization of patients following recalibration in the respective triage categories. Recalibration required new site-specific risk thresholds that may not be needed if prioritization based on rank is all that is required. The Smart Triage model shows promise for wider application for use in triage for sick children in different settings.

Pathophysiological Features in Electronic Medical Records Sustain Model Performance under Temporal Dataset Shift

Access to real-world data streams like electronic medical records (EMRs) has accelerated the development of supervised machine learning (ML) models for clinical applications. However, few studies investigate the differential impact of particular features in the EMR on model performance under temporal dataset shift. To explain how features in the EMR impact models over time, this study aggregates features into feature groups by their source (e.g. medication orders, diagnosis codes and lab results) and feature categories based on their reflection of patient pathophysiology or healthcare processes. We adapt Shapley values to explain feature groups' and feature categories' marginal contribution to initial and sustained model performance. We investigate three standard clinical prediction tasks and find that while feature contributions to initial performance differ across tasks, pathophysiological features help mitigate temporal discrimination deterioration. These results provide interpretable insights on how specific feature groups contribute to model performance and robustness to temporal dataset shift.

Intelligent Animal Husbandry: Present and Future

The main priorities in the contemporary breeding of different animal species have been directed toward the use of intelligent approaches for accelerating genetic progress, ensuring animal welfare and environmental protection by reducing the release of manure and gas emissions [...].

Prediction of 30-Day Mortality Following Revision Total Hip and Knee Arthroplasty: Machine Learning Algorithms Outperform CARDE-B, 5-Item, and 6-Item Modified Frailty Index Risk Scores

BACKGROUND

Although risk calculators are used to prognosticate postoperative outcomes following revision total hip and knee arthroplasty (total joint arthroplasty [TJA]), machine learning (ML) based predictive tools have emerged as a promising alternative for improved risk stratification. This study aimed to compare the predictive ability of ML models for 30-day mortality following revision TJA to that of traditional risk-assessment indices such as the CARDE-B score (congestive heart failure, albumin (< 3.5 mg/dL), renal failure on dialysis, dependence for daily living, elderly (> 65 years of age), and body mass index (BMI) of < 25 kg/m2), 5-item modified frailty index (5MFI), and 6MFI.

METHODS

Adult patients undergoing revision TJA between 2013 and 2020 were selected from the American College of Surgeons National Surgical Quality Improvement Program database and randomly split 80:20 to compose the training and validation cohorts. There were 3 ML models - extreme gradient boosting, random forest, and elastic-net penalized logistic regression (NEPLR) - that were developed and evaluated using discrimination, calibration metrics, and accuracy. The discrimination of CARDE-B, 5MFI, and 6MFI scores was assessed individually and compared to that of ML models.

RESULTS

All models were equally accurate (Brier score = 0.005) and demonstrated outstanding discrimination with similar areas under the receiver operating characteristic curve (AUCs, extreme gradient boosting = 0.94, random forest = NEPLR = 0.93). The NEPLR was the best-calibrated model overall (slope = 0.54, intercept = -0.004). The CARDE-B had the highest discrimination among the scores (AUC = 0.89), followed by 6MFI (AUC = 0.80), and 5MFI (AUC = 0.68). Albumin < 3.5 mg/dL and BMI (< 30.15) were the most important predictors of 30-day mortality following revision TJA.

CONCLUSIONS

The ML models outperform traditional risk-assessment indices in predicting postoperative 30-day mortality after revision TJA. Our findings highlight the utility of ML for risk stratification in a clinical setting. The identification of hypoalbuminemia and BMI as prognostic markers may allow patient-specific perioperative optimization strategies to improve outcomes following revision TJA.

Applying models of care for total hip and knee arthroplasty: External validation of a published predictive model to identify extended stay risk prior to lower-limb arthroplasty

OBJECTIVE

This study aimed to externally validate a reported model for identifying patients requiring extended stay following lower limb arthroplasty in a new setting.

DESIGN

External validation of a previously reported prognostic model, using retrospective data.

SETTING

Medium-sized hospital orthopaedic department, Australia.

PARTICIPANTS

Electronic medical records were accessed for data collection between Sep-2019 and Feb-2020 and retrospective data extracted from 200 randomly selected total hip or knee arthroplasty patients.

INTERVENTION

Participants received total hip or knee replacement between 2-Feb-16 and 4-Apr-19. This study was a non-interventional retrospective study.

MAIN MEASURES

Model validation was assessed with discrimination, calibration on both original and adjusted forms of the candidate model. Decision curve analysis was conducted on the outputs of the adjusted model to determine net benefit at a predetermined decision threshold (0.5).

RESULTS

The original model performed poorly, grossly overestimating length of stay with mean calibration of -3.6 (95% confidence interval -3.9 to -3.2) and calibration slope of 0.52. Performance improved following adjustment of the model intercept and model coefficients (mean calibration 0.48, 95% confidence interval 0.16 to 0.80 and slope of 1.0), but remained poorly calibrated at low and medium risk threshold and net benefit was modest (three additional patients per hundred identified as at-risk) at the a-priori risk threshold.

CONCLUSIONS

External validation demonstrated poor performance when applied to a new patient population and would provide limited benefit for our institution. Implementation of predictive models for arthroplasty should include practical assessment of discrimination, calibration and net benefit at a clinically acceptable threshold.

Applying Machine Learning Techniques to Implementation Science

Machine learning (ML) approaches could expand the usefulness and application of implementation science methods in clinical medicine and public health settings. The aim of this viewpoint is to introduce a roadmap for applying ML techniques to address implementation science questions, such as predicting what will work best, for whom, under what circumstances, and with what predicted level of support, and what and when adaptation or deimplementation are needed. We describe how ML approaches could be used and discuss challenges that implementation scientists and methodologists will need to consider when using ML throughout the stages of implementation.

Analytical Performance Specifications for Input Variables: Investigation of the Model of End-Stage Liver Disease

BACKGROUND

Artificial intelligence models constitute specific uses of analysis results and, therefore, necessitate evaluation of analytical performance specifications (APS) for this context specifically. The Model of End-stage Liver Disease (MELD) is a clinical prediction model based on measurements of bilirubin, creatinine, and the international normalized ratio (INR). This study evaluates the propagation of error through the MELD, to inform choice of APS for the MELD input variables.

METHODS

A total of 6093 consecutive MELD scores and underlying analysis results were retrospectively collected. "Desirable analytical variation" based on biological variation as well as current local analytical variation was simulated onto the data set as well as onto a constructed data set, representing a worst-case scenario. Resulting changes in MELD score and risk classification were calculated.

RESULTS

Biological variation-based APS in the worst-case scenario resulted in 3.26% of scores changing by ≥1 MELD point. In the patient-derived data set, the same variation resulted in 0.92% of samples changing by ≥1 MELD point, and 5.5% of samples changing risk category. Local analytical performance resulted in lower reclassification rates.

CONCLUSIONS

Error propagation through MELD is complex and includes population-dependent mechanisms. Biological variation-derived APS were acceptable for all uses of the MELD score. Other combinations of APS can yield equally acceptable results. This analysis exemplifies how error propagation through artificial intelligence models can become highly complex. This complexity will necessitate that both model suppliers and clinical laboratories address analytical performance specifications for the specific use case, as these may differ from performance specifications for traditional use of the analyses.

Preoperative Prediction of Postoperative Infections Using Machine Learning and Electronic Health Record Data

OBJECTIVE

To estimate preoperative risk of postoperative infections using structured electronic health record (EHR) data.

BACKGROUND

Surveillance and reporting of postoperative infections is primarily done through costly, labor-intensive manual chart reviews on a small sample of patients. Automated methods using statistical models applied to postoperative EHR data have shown promise to augment manual review as they can cover all operations in a timely manner. However, there are no specific models for risk-adjusting infectious complication rates using EHR data.

METHODS

Preoperative EHR data from 30,639 patients (2013-2019) were linked to the American College of Surgeons National Surgical Quality Improvement Program preoperative data and postoperative infection outcomes data from 5 hospitals in the University of Colorado Health System. EHR data included diagnoses, procedures, operative variables, patient characteristics, and medications. Lasso and the knockoff filter were used to perform controlled variable selection. Outcomes included surgical site infection, urinary tract infection, sepsis/septic shock, and pneumonia up to 30 days postoperatively.

RESULTS

Among >15,000 candidate predictors, 7 were chosen for the surgical site infection model and 6 for each of the urinary tract infection, sepsis, and pneumonia models. Important variables included preoperative presence of the specific outcome, wound classification, comorbidities, and American Society of Anesthesiologists physical status classification. The area under the receiver operating characteristic curve for each model ranged from 0.73 to 0.89.

CONCLUSIONS

Parsimonious preoperative models for predicting postoperative infection risk using EHR data were developed and showed comparable performance to existing American College of Surgeons National Surgical Quality Improvement Program risk models that use manual chart review. These models can be used to estimate risk-adjusted postoperative infection rates applied to large volumes of EHR data in a timely manner.

[Artificial intelligence and acute kidney injury]

Digitalization is increasingly finding its way into intensive care units and with it artificial intelligence (AI) for critically ill patients. One promising area for the use of AI is in the field of acute kidney injury (AKI). The use of AI is primarily focused on the prediction of AKI, but further approaches are also being used to classify existing AKI into different phenotypes. Different AI models are used for prediction. The area under the receiver operating characteristic curve values (AUROC) achieved with these models vary and are influenced by several factors, such as the prediction time and the definition of AKI. Most models have an AUROC between 0.650 and 0.900, with lower values for predictions further into the future and when applying Acute Kidney Injury Network (AKIN) instead of KDIGO criteria. Classification into phenotypes already makes it possible to categorize patients into groups with different risks of mortality or requirement of renal replacement therapy (RRT), but the etiologies or therapeutic consequences derived from this are still lacking. However, all the models suffer from AI-specific shortcomings. The use of large databases does not make it possible to promptly include recent changes in therapy and the implementation of new biomarkers in a relevant proportion. For this reason, serum creatinine and urinary output, with their known limitations, dominate current AI models for prediction impairing the performance of the current models. On the other hand, the increasingly complex models no longer allow physicians to understand the basis on which the warning of a threatening AKI is calculated and subsequent initiation of therapy should take place. The successful use of AIs in routine clinical practice will be highly determined by the trust of the physicians in the systems and overcoming the aforementioned weaknesses. However, the clinician will remain irreplaceable as the decisive authority for critically ill patients by combining measurable and nonmeasurable parameters.

Development and External Validation of a Prediction Model for Colorectal Cancer Among Patients Awaiting Surveillance Colonoscopy Following Polypectomy

Background and Aims

Demand for surveillance colonoscopy can sometimes exceed capacity, such as during and following the coronavirus disease 2019 pandemic, yet no tools exist to prioritize the patients most likely to be diagnosed with colorectal cancer (CRC) among those awaiting surveillance colonoscopy. We developed a multivariable prediction model for CRC at surveillance comparing performance to a model that assigned patients as low or high risk based solely on polyp characteristics (guideline-based model).

Methods

Logistic regression was used for model development among patients receiving surveillance colonoscopy in 2014-2019. Candidate predictors included index colonoscopy indication, findings, and endoscopist adenoma detection rate, and patient and clinical characteristics at surveillance. Patients were randomly divided into model development (n = 36,994) and internal validation cohorts (n = 15,854). External validation was performed on 30,015 patients receiving surveillance colonoscopy in 2020-2022, and the multivariable model was then updated and retested.

Results

One hundred fourteen, 43, and 71 CRCs were detected at surveillance in the 3 cohorts, respectively. Polyp size ≥10 mm, adenoma detection rate <32.5% or missing, patient age, and ever smoked tobacco were significant CRC predictors; this multivariable model outperformed the guideline-based model (internal validation cohort area under the receiver-operating characteristic curve: 0.73, 95% confidence interval (CI): 0.66-0.81 vs 0.52, 95% CI: 0.45-0.60). Performance declined at external validation but recovered with model updating (operating characteristic curve: 0.72 95% CI: 0.66-0.77).

Conclusion

When surveillance colonoscopy demand exceeds capacity, a prediction model featuring common clinical predictors may help prioritize patients at highest risk for CRC among those awaiting surveillance. Also, regular model updates can address model performance drift.

Why do probabilistic clinical models fail to transport between sites

The rising popularity of artificial intelligence in healthcare is highlighting the problem that a computational model achieving super-human clinical performance at its training sites may perform substantially worse at new sites. In this perspective, we argue that we should typically expect this failure to transport, and we present common sources for it, divided into those under the control of the experimenter and those inherent to the clinical data-generating process. Of the inherent sources we look a little deeper into site-specific clinical practices that can affect the data distribution, and propose a potential solution intended to isolate the imprint of those practices on the data from the patterns of disease cause and effect that are the usual target of probabilistic clinical models.

Monitoring performance of clinical artificial intelligence: a scoping review protocol

OBJECTIVE

The objective of this scoping review is to describe the scope and nature of research on the monitoring of clinical artificial intelligence (AI) systems. The review will identify the various methodologies used to monitor clinical AI, while also mapping the factors that influence the selection of monitoring approaches.

INTRODUCTION

AI is being used in clinical decision-making at an increasing rate. While much attention has been directed toward the development and validation of AI for clinical applications, the practical implementation aspects, notably the establishment of rational monitoring/quality assurance systems, has received comparatively limited scientific interest. Given the scarcity of evidence and the heterogeneity of methodologies used in this domain, there is a compelling rationale for conducting a scoping review on this subject.

INCLUSION CRITERIA

This scoping review will include any publications that describe systematic, continuous, or repeated initiatives that evaluate or predict clinical performance of AI models with direct implications for the management of patients in any segment of the health care system.

METHODS

Publications will be identified through searches of the MEDLINE (Ovid), Embase (Ovid), and Scopus databases. Additionally, backward and forward citation searches, as well as a thorough investigation of gray literature, will be conducted. Title and abstract screening, full-text evaluation, and data extraction will be performed by 2 or more independent reviewers. Data will be extracted using a tool developed by the authors. The results will be presented graphically and narratively.

REVIEW REGISTRATION

Open Science Framework https://osf.io/afkrn.

National Multi-Institutional Validation of a Surgical Transfusion Risk Prediction Model

BACKGROUND

Accurate estimation of surgical transfusion risk is important for many aspects of surgical planning, yet few methods for estimating are available for estimating such risk. There is a need for reliable validated methods for transfusion risk stratification to support effective perioperative planning and resource stewardship.

STUDY DESIGN

This study was conducted using the American College of Surgeons NSQIP datafile from 2019. S-PATH performance was evaluated at each contributing hospital, with and without hospital-specific model tuning. Linear regression was used to assess the relationship between hospital characteristics and area under the receiver operating characteristic (AUROC) curve.

RESULTS

A total of 1,000,927 surgical cases from 414 hospitals were evaluated. Aggregate AUROC was 0.910 (95% CI 0.904 to 0.916) without model tuning and 0.925 (95% CI 0.919 to 0.931) with model tuning. AUROC varied across individual hospitals (median 0.900, interquartile range 0.849 to 0.944), but no statistically significant relationships were found between hospital-level characteristics studied and model AUROC.

CONCLUSIONS

S-PATH demonstrated excellent discriminative performance, although there was variation across hospitals that was not well-explained by hospital-level characteristics. These results highlight the S-PATH's viability as a generalizable surgical transfusion risk prediction tool.

Use of Electronic Medical Records (EMR) in Gerontology: Benefits, Considerations and a Promising Future

Electronic medical records (EMRs) have many benefits in clinical research in gerontology, enabling data analysis, development of prognostic tools and disease risk prediction. EMRs also offer a range of advantages in clinical practice, such as comprehensive medical records, streamlined communication with healthcare providers, remote data access, and rapid retrieval of test results, ultimately leading to increased efficiency, enhanced patient safety, and improved quality of care in gerontology, which includes benefits like reduced medication use and better patient history taking and physical examination assessments. The use of artificial intelligence (AI) and machine learning (ML) approaches on EMRs can further improve disease diagnosis, symptom classification, and support clinical decision-making. However, there are also challenges related to data quality, data entry errors, as well as the ethics and safety of using AI in healthcare. This article discusses the future of EMRs in gerontology and the application of AI and ML in clinical research. Ethical and legal issues surrounding data sharing and the need for healthcare professionals to critically evaluate and integrate these technologies are also emphasized. The article concludes by discussing the challenges related to the use of EMRs in research as well as in their primary intended use, the daily clinical practice.

Evaluation of crowdsourced mortality prediction models as a framework for assessing artificial intelligence in medicine

OBJECTIVE

Applications of machine learning in healthcare are of high interest and have the potential to improve patient care. Yet, the real-world accuracy of these models in clinical practice and on different patient subpopulations remains unclear. To address these important questions, we hosted a community challenge to evaluate methods that predict healthcare outcomes. We focused on the prediction of all-cause mortality as the community challenge question.

MATERIALS AND METHODS

Using a Model-to-Data framework, 345 registered participants, coalescing into 25 independent teams, spread over 3 continents and 10 countries, generated 25 accurate models all trained on a dataset of over 1.1 million patients and evaluated on patients prospectively collected over a 1-year observation of a large health system.

RESULTS

The top performing team achieved a final area under the receiver operator curve of 0.947 (95% CI, 0.942-0.951) and an area under the precision-recall curve of 0.487 (95% CI, 0.458-0.499) on a prospectively collected patient cohort.

DISCUSSION

Post hoc analysis after the challenge revealed that models differ in accuracy on subpopulations, delineated by race or gender, even when they are trained on the same data.

CONCLUSION

This is the largest community challenge focused on the evaluation of state-of-the-art machine learning methods in a healthcare system performed to date, revealing both opportunities and pitfalls of clinical AI.

An external validation of the Kidney Donor Risk Index in the UK transplant population in the presence of semi-competing events

BACKGROUND

Transplantation represents the optimal treatment for many patients with end-stage kidney disease. When a donor kidney is available to a waitlisted patient, clinicians responsible for the care of the potential recipient must make the decision to accept or decline the offer based upon complex and variable information about the donor, the recipient and the transplant process. A clinical prediction model may be able to support clinicians in their decision-making. The Kidney Donor Risk Index (KDRI) was developed in the United States to predict graft failure following kidney transplantation. The survival process following transplantation consists of semi-competing events where death precludes graft failure, but not vice-versa.

METHODS

We externally validated the KDRI in the UK kidney transplant population and assessed whether validation under a semi-competing risks framework impacted predictive performance. Additionally, we explored whether the KDRI requires updating. We included 20,035 adult recipients of first, deceased donor, single, kidney-only transplants between January 1, 2004, and December 31, 2018, collected by the UK Transplant Registry and held by NHS Blood and Transplant. The outcomes of interest were 1- and 5-year graft failure following transplantation. In light of the semi-competing events, recipient death was handled in two ways: censoring patients at the time of death and modelling death as a competing event. Cox proportional hazard models were used to validate the KDRI when censoring graft failure by death, and cause-specific Cox models were used to account for death as a competing event.

RESULTS

The KDRI underestimated event probabilities for those at higher risk of graft failure. For 5-year graft failure, discrimination was poorer in the semi-competing risks model (0.625, 95% CI 0.611 to 0.640;0.611, 95% CI 0.597 to 0.625), but predictions were more accurate (Brier score 0.117, 95% CI 0.112 to 0.121; 0.114, 95% CI 0.109 to 0.118). Calibration plots were similar regardless of whether the death was modelled as a competing event or not. Updating the KDRI worsened calibration, but marginally improved discrimination.

CONCLUSIONS

Predictive performance for 1-year graft failure was similar between death-censored and competing event graft failure, but differences appeared when predicting 5-year graft failure. The updated index did not have superior performance and we conclude that updating the KDRI in the present form is not required.

The Trade Secret Taboo: Open Science Methods are Required to Improve Prediction Models in Sports Medicine and Performance

Clinical prediction models in sports medicine that utilize regression or machine learning techniques have become more widely published, used, and disseminated. However, these models are typically characterized by poor methodology and incomplete reporting, and an inadequate evaluation of performance, leading to unreliable predictions and weak clinical utility within their intended sport population. Before implementation in practice, models require a thorough evaluation. Strong replicable methods and transparency reporting allow practitioners and researchers to make independent judgments as to the model's validity, performance, clinical usefulness, and confidence it will do no harm. However, this is not reflected in the sports medicine literature. As shown in a recent systematic review of models for predicting sports injury models, most were typically characterized by poor methodology, incomplete reporting, and inadequate performance evaluation. Because of constraints imposed by data from individual teams, the development of accurate, reliable, and useful models is highly reliant on external validation. However, a barrier to collaboration is a desire to maintain a competitive advantage; a team's proprietary information is often perceived as high value, and so these 'trade secrets' are frequently guarded. These 'trade secrets' also apply to commercially available models, as developers are unwilling to share proprietary (and potentially profitable) development and validation information. In this Current Opinion, we: (1) argue that open science is essential for improving sport prediction models and (2) critically examine sport prediction models for open science practices.

Evaluation of MRI-based machine learning approaches for computer-aided diagnosis of dementia in a clinical data warehouse

A variety of algorithms have been proposed for computer-aided diagnosis of dementia from anatomical brain MRI. These approaches achieve high accuracy when applied to research data sets but their performance on real-life clinical routine data has not been evaluated yet. The aim of this work was to study the performance of such approaches on clinical routine data, based on a hospital data warehouse, and to compare the results to those obtained on a research data set. The clinical data set was extracted from the hospital data warehouse of the Greater Paris area, which includes 39 different hospitals. The research set was composed of data from the Alzheimer's Disease Neuroimaging Initiative data set. In the clinical set, the population of interest was identified by exploiting the diagnostic codes from the 10th revision of the International Classification of Diseases that are assigned to each patient. We studied how the imbalance of the training sets, in terms of contrast agent injection and image quality, may bias the results. We demonstrated that computer-aided diagnosis performance was strongly biased upwards (over 17 percent points of balanced accuracy) by the confounders of image quality and contrast agent injection, a phenomenon known as the Clever Hans effect or shortcut learning. When these biases were removed, the performance was very poor. In any case, the performance was considerably lower than on the research data set. Our study highlights that there are still considerable challenges for translating dementia computer-aided diagnosis systems to clinical routine.

Implications of the Use of Artificial Intelligence Predictive Models in Health Care Settings : A Simulation Study

BACKGROUND

Substantial effort has been directed toward demonstrating uses of predictive models in health care. However, implementation of these models into clinical practice may influence patient outcomes, which in turn are captured in electronic health record data. As a result, deployed models may affect the predictive ability of current and future models.

OBJECTIVE

To estimate changes in predictive model performance with use through 3 common scenarios: model retraining, sequentially implementing 1 model after another, and intervening in response to a model when 2 are simultaneously implemented.

DESIGN

Simulation of model implementation and use in critical care settings at various levels of intervention effectiveness and clinician adherence. Models were either trained or retrained after simulated implementation.

SETTING

Admissions to the intensive care unit (ICU) at Mount Sinai Health System (New York, New York) and Beth Israel Deaconess Medical Center (Boston, Massachusetts).

PATIENTS

130 000 critical care admissions across both health systems.

INTERVENTION

Across 3 scenarios, interventions were simulated at varying levels of clinician adherence and effectiveness.

MEASUREMENTS

Statistical measures of performance, including threshold-independent (area under the curve) and threshold-dependent measures.

RESULTS

At fixed 90% sensitivity, in scenario 1 a mortality prediction model lost 9% to 39% specificity after retraining once and in scenario 2 a mortality prediction model lost 8% to 15% specificity when created after the implementation of an acute kidney injury (AKI) prediction model; in scenario 3, models for AKI and mortality prediction implemented simultaneously, each led to reduced effective accuracy of the other by 1% to 28%.

LIMITATIONS

In real-world practice, the effectiveness of and adherence to model-based recommendations are rarely known in advance. Only binary classifiers for tabular ICU admissions data were simulated.

CONCLUSION

In simulated ICU settings, a universally effective model-updating approach for maintaining model performance does not seem to exist. Model use may have to be recorded to maintain viability of predictive modeling.

PRIMARY FUNDING SOURCE

National Center for Advancing Translational Sciences.

Prediction differences and implications of acute kidney injury with and without urine output criteria in adult critically ill patients

BACKGROUND

Due to the convenience of serum creatinine (SCr) monitoring and the relative complexity of urine output (UO) monitoring, most studies have predicted acute kidney injury (AKI) only based on SCr criteria. This study aimed to compare the differences between SCr alone and combined UO criteria in predicting AKI.

METHODS

We applied machine learning methods to evaluate the performance of 13 prediction models composed of different feature categories on 16 risk assessment tasks (half used only SCr criteria, half used both SCr and UO criteria). The area under receiver operator characteristic curve (AUROC), the area under precision recall curve (AUPRC) and calibration were used to assess the prediction performance.

RESULTS

In the first week after ICU admission, the prevalence of any AKI was 29% under SCr criteria alone and increased to 60% when the UO criteria was combined. Adding UO to SCr criteria can significantly identify more AKI patients. The predictive importance of feature types with and without UO was different. Using only laboratory data maintained similar predictive performance to the full feature model under only SCr criteria [e.g. for AKI within the 48-h time window after 1 day of ICU admission, AUROC (95% confidence interval) 0.83 (0.82, 0.84) vs 0.84 (0.83, 0.85)], but it was not sufficient when the UO was added [corresponding AUROC (95% confidence interval) 0.75 (0.74, 0.76) vs 0.84 (0.83, 0.85)].

CONCLUSIONS

This study found that SCr and UO measures should not be regarded as equivalent criteria for AKI staging, and emphasizes the importance and necessity of UO criteria in AKI risk assessment.

Machine learning to develop a predictive model of pressure injury in persons with spinal cord injury

STUDY DESIGN

A 5-year longitudinal, retrospective, cohort study.

OBJECTIVES

Develop a prediction model based on electronic health record (EHR) data to identify veterans with spinal cord injury/diseases (SCI/D) at highest risk for new pressure injuries (PIs).

SETTING

Structured (coded) and text EHR data, for veterans with SCI/D treated in a VHA SCI/D Center between October 1, 2008, and September 30, 2013.

METHODS

A total of 4709 veterans were available for analysis after randomly selecting 175 to act as a validation (gold standard) sample. Machine learning models were created using ten-fold cross validation and three techniques: (1) two-step logistic regression; (2) regression model employing adaptive LASSO; (3) and gradient boosting. Models based on each method were compared using area under the receiver-operating curve (AUC) analysis.

RESULTS

The AUC value for the gradient boosting model was 0.62 (95% CI = 0.54-0.70), for the logistic regression model it was 0.67 (95% CI = 0.59-0.75), and for the adaptive LASSO model it was 0.72 (95% CI = 0.65-80). Based on these results, the adaptive LASSO model was chosen for interpretation. The strongest predictors of new PI cases were having fewer total days in the hospital in the year before the annual exam, higher vs. lower weight and most severe vs. less severe grade of injury based on the American Spinal Cord Injury Association (ASIA) Impairment Scale.

CONCLUSIONS

While the analyses resulted in a potentially useful predictive model, clinical implications were limited because modifiable risk factors were absent in the models.

Assessing the effects of data drift on the performance of machine learning models used in clinical sepsis prediction

BACKGROUND

Data drift can negatively impact the performance of machine learning algorithms (MLAs) that were trained on historical data. As such, MLAs should be continuously monitored and tuned to overcome the systematic changes that occur in the distribution of data. In this paper, we study the extent of data drift and provide insights about its characteristics for sepsis onset prediction. This study will help elucidate the nature of data drift for prediction of sepsis and similar diseases. This may aid with the development of more effective patient monitoring systems that can stratify risk for dynamic disease states in hospitals.

METHODS

We devise a series of simulations that measure the effects of data drift in patients with sepsis, using electronic health records (EHR). We simulate multiple scenarios in which data drift may occur, namely the change in the distribution of the predictor variables (covariate shift), the change in the statistical relationship between the predictors and the target (concept shift), and the occurrence of a major healthcare event (major event) such as the COVID-19 pandemic. We measure the impact of data drift on model performances, identify the circumstances that necessitate model retraining, and compare the effects of different retraining methodologies and model architecture on the outcomes. We present the results for two different MLAs, eXtreme Gradient Boosting (XGB) and Recurrent Neural Network (RNN).

RESULTS

Our results show that the properly retrained XGB models outperform the baseline models in all simulation scenarios, hence signifying the existence of data drift. In the major event scenario, the area under the receiver operating characteristic curve (AUROC) at the end of the simulation period is 0.811 for the baseline XGB model and 0.868 for the retrained XGB model. In the covariate shift scenario, the AUROC at the end of the simulation period for the baseline and retrained XGB models is 0.853 and 0.874 respectively. In the concept shift scenario and under the mixed labeling method, the retrained XGB models perform worse than the baseline model for most simulation steps. However, under the full relabeling method, the AUROC at the end of the simulation period for the baseline and retrained XGB models is 0.852 and 0.877 respectively. The results for the RNN models were mixed, suggesting that retraining based on a fixed network architecture may be inadequate for an RNN. We also present the results in the form of other performance metrics such as the ratio of observed to expected probabilities (calibration) and the normalized rate of positive predictive values (PPV) by prevalence, referred to as lift, at a sensitivity of 0.8.

CONCLUSION

Our simulations reveal that retraining periods of a couple of months or using several thousand patients are likely to be adequate to monitor machine learning models that predict sepsis. This indicates that a machine learning system for sepsis prediction will probably need less infrastructure for performance monitoring and retraining compared to other applications in which data drift is more frequent and continuous. Our results also show that in the event of a concept shift, a full overhaul of the sepsis prediction model may be necessary because it indicates a discrete change in the definition of sepsis labels, and mixing the labels for the sake of incremental training may not produce the desired results.

Estimating individual minimum calibration for deep-learning with predictive performance recovery: An example case of gait surface classification from wearable sensor gait data

Clinical datasets often comprise multiple data points or trials sampled from a single participant. When these datasets are used to train machine learning models, the method used to extract train and test sets must be carefully chosen. Using the standard machine learning approach (random-wise split), different trials from the same participant may appear in both training and test sets. This has led to schemes capable of segregating data points from a same participant into a single set (subject-wise split). Past investigations have demonstrated that models trained in this manner underperform compared to those trained using random-split schemes. Additional training of models via a small subset of trials, known as calibration, bridges the gap in performance across split schemes; however, the amount of calibration trials required to achieve strong model performance is unclear. Thus, this study aims to investigate the relationship between calibration training set size and prediction accuracy on the calibration test set. A database of 30 young, healthy adults performing multiple walking trials across nine different surfaces while fit with inertial measurement unit sensors on the lower limbs was used to develop a deep-learning classifier. For subject-wise trained models, calibration on a single gait cycle per surface yielded a 70% increase in F1-score, the harmonic mean of precision and recall, while 10 gait cycles per surface were sufficient to match the performance of a random-wise trained model. Code to generate calibration curves may be found at (https://github.com/GuillaumeLam/PaCalC).

Impact of the Covid-19 pandemic on the performance of machine learning algorithms for predicting perioperative mortality

BACKGROUND

Machine-learning models are susceptible to external influences which can result in performance deterioration. The aim of our study was to elucidate the impact of a sudden shift in covariates, like the one caused by the Covid-19 pandemic, on model performance.

METHODS

After ethical approval and registration in Clinical Trials (NCT04092933, initial release 17/09/2019), we developed different models for the prediction of perioperative mortality based on preoperative data: one for the pre-pandemic data period until March 2020, one including data before the pandemic and from the first wave until May 2020, and one that covers the complete period before and during the pandemic until October 2021. We applied XGBoost as well as a Deep Learning neural network (DL). Performance metrics of each model during the different pandemic phases were determined, and XGBoost models were analysed for changes in feature importance.

RESULTS

XGBoost and DL provided similar performance on the pre-pandemic data with respect to area under receiver operating characteristic (AUROC, 0.951 vs. 0.942) and area under precision-recall curve (AUPR, 0.144 vs. 0.187). Validation in patient cohorts of the different pandemic waves showed high fluctuations in performance from both AUROC and AUPR for DL, whereas the XGBoost models seemed more stable. Change in variable frequencies with onset of the pandemic were visible in age, ASA score, and the higher proportion of emergency operations, among others. Age consistently showed the highest information gain. Models based on pre-pandemic data performed worse during the first pandemic wave (AUROC 0.914 for XGBoost and DL) whereas models augmented with data from the first wave lacked performance after the first wave (AUROC 0.907 for XGBoost and 0.747 for DL). The deterioration was also visible in AUPR, which worsened by over 50% in both XGBoost and DL in the first phase after re-training.

CONCLUSIONS

A sudden shift in data impacts model performance. Re-training the model with updated data may cause degradation in predictive accuracy if the changes are only transient. Too early re-training should therefore be avoided, and close model surveillance is necessary.

Computed tomography-based radiomics for the differential diagnosis of pneumonitis in stage IV non-small cell lung cancer patients treated with immune checkpoint inhibitors

INTRODUCTION

Immunotherapy-induced pneumonitis (IIP) is a serious side-effect which requires accurate diagnosis and management with high-dose corticosteroids. The differential diagnosis between IIP and other types of pneumonitis (OTP) remains challenging due to similar radiological patterns. This study was aimed to develop a prediction model to differentiate IIP from OTP in patients with stage IV non-small cell lung cancer (NSCLC) who developed pneumonitis during immunotherapy.

METHODS

Consecutive patients with metastatic NSCLC treated with immunotherapy in six centres in the Netherlands and Belgium from 2017 to 2020 were reviewed and cause-specific pneumonitis events were identified. Seven regions of interest (segmented lungs and spheroidal/cubical regions surrounding the inflammation) were examined to extract the most predictive radiomic features from the chest computed tomography images obtained at pneumonitis manifestation. Models were internally tested regarding discrimination, calibration and decisional benefit. To evaluate the clinical application of the models, predicted labels were compared with the separate clinical and radiological judgements.

RESULTS

A total of 556 patients were reviewed; 31 patients (5.6%) developed IIP and 41 patients developed OTP (7.4%). The line of immunotherapy was the only predictive factor in the clinical model (2nd versus 1st odds ratio = 0.08, 95% confidence interval:0.01-0.77). The best radiomic model was achieved using a 75-mm spheroidal region of interest which showed an optimism-corrected area under the receiver operating characteristic curve of 0.83 (95% confidence interval:0.77-0.95) with negative and positive predictive values of 80% and 79%, respectively. Good calibration and net benefits were achieved for the radiomic model across the entire range of probabilities. A correct diagnosis was provided by the radiomic model in 10 out of 12 cases with non-conclusive radiological judgements.

CONCLUSION

Radiomic biomarkers applied to computed tomography imaging may support clinicians making the differential diagnosis of pneumonitis in patients with NSCLC receiving immunotherapy, especially when the radiologic assessment is non-conclusive.

Bibliometric and visual analysis of machine learning-based research in acute kidney injury worldwide

Background

Acute kidney injury (AKI) is a serious clinical complication associated with adverse short-term and long-term outcomes. In recent years, with the rapid popularization of electronic health records and artificial intelligence machine learning technology, the detection rate and treatment of AKI have been greatly improved. At present, there are many studies in this field, and a large number of articles have been published, but we do not know much about the quality of research production in this field, as well as the focus and trend of current research.

Methods

Based on the Web of Science Core Collection, studies reporting machine learning-based AKI research that were published from 2013 to 2022 were retrieved and collected after manual review. VOSviewer and other software were used for bibliometric visualization analysis, including publication trends, geographical distribution characteristics, journal distribution characteristics, author contributions, citations, funding source characteristics, and keyword clustering.

Results

A total of 336 documents were analyzed. Since 2018, publications and citations have increased dramatically, with the United States (143) and China (101) as the main contributors. Regarding authors, Bihorac, A and Ozrazgat-Baslanti, T from the Kansas City Medical Center have published 10 articles. Regarding institutions, the University of California (18) had the most publications. Approximately 1/3 of the publications were published in Q1 and Q2 journals, of which Scientific Reports (19) was the most prolific journal. Tomašev et al.'s study that was published in 2019 has been widely cited by researchers. The results of cluster analysis of co-occurrence keywords suggest that the construction of AKI prediction model related to critical patients and sepsis patients is the research frontier, and XGBoost algorithm is also popular.

Conclusion

This study first provides an updated perspective on machine learning-based AKI research, which may be beneficial for subsequent researchers to choose suitable journals and collaborators and may provide a more convenient and in-depth understanding of the research basis, hotspots and frontiers.

EHR foundation models improve robustness in the presence of temporal distribution shift

Temporal distribution shift negatively impacts the performance of clinical prediction models over time. Pretraining foundation models using self-supervised learning on electronic health records (EHR) may be effective in acquiring informative global patterns that can improve the robustness of task-specific models. The objective was to evaluate the utility of EHR foundation models in improving the in-distribution (ID) and out-of-distribution (OOD) performance of clinical prediction models. Transformer- and gated recurrent unit-based foundation models were pretrained on EHR of up to 1.8 M patients (382 M coded events) collected within pre-determined year groups (e.g., 2009-2012) and were subsequently used to construct patient representations for patients admitted to inpatient units. These representations were used to train logistic regression models to predict hospital mortality, long length of stay, 30-day readmission, and ICU admission. We compared our EHR foundation models with baseline logistic regression models learned on count-based representations (count-LR) in ID and OOD year groups. Performance was measured using area-under-the-receiver-operating-characteristic curve (AUROC), area-under-the-precision-recall curve, and absolute calibration error. Both transformer and recurrent-based foundation models generally showed better ID and OOD discrimination relative to count-LR and often exhibited less decay in tasks where there is observable degradation of discrimination performance (average AUROC decay of 3% for transformer-based foundation model vs. 7% for count-LR after 5-9 years). In addition, the performance and robustness of transformer-based foundation models continued to improve as pretraining set size increased. These results suggest that pretraining EHR foundation models at scale is a useful approach for developing clinical prediction models that perform well in the presence of temporal distribution shift.

There is no such thing as a validated prediction model

BACKGROUND

Clinical prediction models should be validated before implementation in clinical practice. But is favorable performance at internal validation or one external validation sufficient to claim that a prediction model works well in the intended clinical context?

MAIN BODY

We argue to the contrary because (1) patient populations vary, (2) measurement procedures vary, and (3) populations and measurements change over time. Hence, we have to expect heterogeneity in model performance between locations and settings, and across time. It follows that prediction models are never truly validated. This does not imply that validation is not important. Rather, the current focus on developing new models should shift to a focus on more extensive, well-conducted, and well-reported validation studies of promising models.

CONCLUSION

Principled validation strategies are needed to understand and quantify heterogeneity, monitor performance over time, and update prediction models when appropriate. Such strategies will help to ensure that prediction models stay up-to-date and safe to support clinical decision-making.

Performance drift in a mortality prediction algorithm among patients with cancer during the SARS-CoV-2 pandemic

Sudden changes in health care utilization during the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic may have impacted the performance of clinical predictive models that were trained prior to the pandemic. In this study, we evaluated the performance over time of a machine learning, electronic health record-based mortality prediction algorithm currently used in clinical practice to identify patients with cancer who may benefit from early advance care planning conversations. We show that during the pandemic period, algorithm identification of high-risk patients had a substantial and sustained decline. Decreases in laboratory utilization during the peak of the pandemic may have contributed to drift. Calibration and overall discrimination did not markedly decline during the pandemic. This argues for careful attention to the performance and retraining of predictive algorithms that use inputs from the pandemic period.

Optimizing Equity: Working towards Fair Machine Learning Algorithms in Laboratory Medicine

BACKGROUND

Methods of machine learning provide opportunities to use real-world data to solve complex problems. Applications of these methods in laboratory medicine promise to increase diagnostic accuracy and streamline laboratory operations leading to improvement in the quality and efficiency of healthcare delivery. However, machine learning models are vulnerable to learning from undesirable patterns in the data that reflect societal biases. As a result, irresponsible application of machine learning may lead to the perpetuation, or even amplification, of existing disparities in healthcare outcomes.

CONTENT

In this work, we review what it means for a model to be unfair, discuss the various ways that machine learning models become unfair, and present engineering principles emerging from the field of algorithmic fairness. These materials are presented with a focus on the development of machine learning models in laboratory medicine.

SUMMARY

We hope that this work will serve to increase awareness, and stimulate further discussion, of this important issue among laboratorians as the field moves forward with the incorporation of machine learning models into laboratory practice.

Machine learning models for predicting acute kidney injury: a systematic review and critical appraisal

BACKGROUND

The number of studies applying machine learning (ML) to predict acute kidney injury (AKI) has grown steadily over the past decade. We assess and critically appraise the state of the art in ML models for AKI prediction, considering performance, methodological soundness, and applicability.

METHODS

We searched PubMed and ArXiv, extracted data, and critically appraised studies based on the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD), Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS), and Prediction Model Risk of Bias Assessment Tool (PROBAST) guidelines.

RESULTS

Forty-six studies from 3166 titles were included. Thirty-eight studies developed a model, five developed and externally validated one, and three studies externally validated one. Flexible ML methods were used more often than deep learning, although the latter was common with temporal variables and text as predictors. Predictive performance showed an area under receiver operating curves ranging from 0.49 to 0.99. Our critical appraisal identified a high risk of bias in 39 studies. Some studies lacked internal validation, whereas external validation and interpretability of results were rarely considered. Fifteen studies focused on AKI prediction in the intensive care setting, and the US-derived Medical Information Mart for Intensive Care (MIMIC) data set was commonly used. Reproducibility was limited as data and code were usually unavailable.

CONCLUSIONS

Flexible ML methods are popular for the prediction of AKI, although more complex models based on deep learning are emerging. Our critical appraisal identified a high risk of bias in most models: Studies should use calibration measures and external validation more often, improve model interpretability, and share data and code to improve reproducibility.

A hybrid adaptive approach for instance transfer learning with dynamic and imbalanced data

Machine learning has demonstrated success in clinical risk prediction modeling with complex electronic health record data. However, the evolving nature of clinical practices can dynamically change the underlying data distribution over time, leading to model performance drift. Adopting an outdated model is potentially risky and may result in unintentional losses. In this paper, we propose a novel Hybrid Adaptive Boosting approach (HA-Boost) for transfer learning. HA-Boost is characterized by the domain similarity-based and class imbalance-based adaptation mechanisms, which simultaneously address two critical limitations of the classical TrAdaBoost algorithm. We validated HA-Boost in predicting hospital-acquired acute kidney injury using real-world longitudinal electronic health records data. The experiment results demonstrate that HA-Boost stably outperforms the competing baselines in terms of both AUROC and AUPRC across a 7-year time span. This study has confirmed the effectiveness of transfer learning as a superior model updating approach in dynamic environment.

Shifting machine learning for healthcare from development to deployment and from models to data

In the past decade, the application of machine learning (ML) to healthcare has helped drive the automation of physician tasks as well as enhancements in clinical capabilities and access to care. This progress has emphasized that, from model development to model deployment, data play central roles. In this Review, we provide a data-centric view of the innovations and challenges that are defining ML for healthcare. We discuss deep generative models and federated learning as strategies to augment datasets for improved model performance, as well as the use of the more recent transformer models for handling larger datasets and enhancing the modelling of clinical text. We also discuss data-focused problems in the deployment of ML, emphasizing the need to efficiently deliver data to ML models for timely clinical predictions and to account for natural data shifts that can deteriorate model performance.

Developing and Validating Multi-Modal Models for Mortality Prediction in COVID-19 Patients: a Multi-center Retrospective Study

The unprecedented global crisis brought about by the COVID-19 pandemic has sparked numerous efforts to create predictive models for the detection and prognostication of SARS-CoV-2 infections with the goal of helping health systems allocate resources. Machine learning models, in particular, hold promise for their ability to leverage patient clinical information and medical images for prediction. However, most of the published COVID-19 prediction models thus far have little clinical utility due to methodological flaws and lack of appropriate validation. In this paper, we describe our methodology to develop and validate multi-modal models for COVID-19 mortality prediction using multi-center patient data. The models for COVID-19 mortality prediction were developed using retrospective data from Madrid, Spain (N = 2547) and were externally validated in patient cohorts from a community hospital in New Jersey, USA (N = 242) and an academic center in Seoul, Republic of Korea (N = 336). The models we developed performed differently across various clinical settings, underscoring the need for a guided strategy when employing machine learning for clinical decision-making. We demonstrated that using features from both the structured electronic health records and chest X-ray imaging data resulted in better 30-day mortality prediction performance across all three datasets (areas under the receiver operating characteristic curves: 0.85 (95% confidence interval: 0.83-0.87), 0.76 (0.70-0.82), and 0.95 (0.92-0.98)). We discuss the rationale for the decisions made at every step in developing the models and have made our code available to the research community. We employed the best machine learning practices for clinical model development. Our goal is to create a toolkit that would assist investigators and organizations in building multi-modal models for prediction, classification, and/or optimization.

A Transfer Learning Approach to Correct the Temporal Performance Drift of Clinical Prediction Models: Retrospective Cohort Study

Background

Clinical prediction models suffer from performance drift as the patient population shifts over time. There is a great need for model updating approaches or modeling frameworks that can effectively use the old and new data.

Objective

Based on the paradigm of transfer learning, we aimed to develop a novel modeling framework that transfers old knowledge to the new environment for prediction tasks, and contributes to performance drift correction.

Methods

The proposed predictive modeling framework maintains a logistic regression–based stacking ensemble of 2 gradient boosting machine (GBM) models representing old and new knowledge learned from old and new data, respectively (referred to as transfer learning gradient boosting machine [TransferGBM]). The ensemble learning procedure can dynamically balance the old and new knowledge. Using 2010-2017 electronic health record data on a retrospective cohort of 141,696 patients, we validated TransferGBM for hospital-acquired acute kidney injury prediction.

Results

The baseline models (ie, transported models) that were trained on 2010 and 2011 data showed significant performance drift in the temporal validation with 2012-2017 data. Refitting these models using updated samples resulted in performance gains in nearly all cases. The proposed TransferGBM model succeeded in achieving uniformly better performance than the refitted models.

Conclusions

Under the scenario of population shift, incorporating new knowledge while preserving old knowledge is essential for maintaining stable performance. Transfer learning combined with stacking ensemble learning can help achieve a balance of old and new knowledge in a flexible and adaptive way, even in the case of insufficient new data.

A hitchhiker's guide to cancer models

Cancer is a complex and uniquely personal disease. More than 1.7 million people in the United States are diagnosed with cancer every year. As the burden of cancer grows, so does the need for new, more effective therapeutics and for predictive tools to identify optimal, personalized treatment options for every patient. Cancer models that recapitulate various aspects of the disease are fundamental to making advances along the continuum of cancer treatment from benchside discoveries to bedside delivery. In this review, we use a thought experiment as a vehicle to arrive at four broad categories of cancer models and explore the strengths, weaknesses, opportunities, and threats for each category in advancing our understanding of the disease and improving treatment strategies.

Considerations in the reliability and fairness audits of predictive models for advance care planning

Multiple reporting guidelines for artificial intelligence (AI) models in healthcare recommend that models be audited for reliability and fairness. However, there is a gap of operational guidance for performing reliability and fairness audits in practice. Following guideline recommendations, we conducted a reliability audit of two models based on model performance and calibration as well as a fairness audit based on summary statistics, subgroup performance and subgroup calibration. We assessed the Epic End-of-Life (EOL) Index model and an internally developed Stanford Hospital Medicine (HM) Advance Care Planning (ACP) model in 3 practice settings: Primary Care, Inpatient Oncology and Hospital Medicine, using clinicians' answers to the surprise question (“Would you be surprised if [patient X] passed away in [Y years]?”) as a surrogate outcome. For performance, the models had positive predictive value (PPV) at or above 0.76 in all settings. In Hospital Medicine and Inpatient Oncology, the Stanford HM ACP model had higher sensitivity (0.69, 0.89 respectively) than the EOL model (0.20, 0.27), and better calibration (O/E 1.5, 1.7) than the EOL model (O/E 2.5, 3.0). The Epic EOL model flagged fewer patients (11%, 21% respectively) than the Stanford HM ACP model (38%, 75%). There were no differences in performance and calibration by sex. Both models had lower sensitivity in Hispanic/Latino male patients with Race listed as “Other.” 10 clinicians were surveyed after a presentation summarizing the audit. 10/10 reported that summary statistics, overall performance, and subgroup performance would affect their decision to use the model to guide care; 9/10 said the same for overall and subgroup calibration. The most commonly identified barriers for routinely conducting such reliability and fairness audits were poor demographic data quality and lack of data access. This audit required 115 person-hours across 8–10 months. Our recommendations for performing reliability and fairness audits include verifying data validity, analyzing model performance on intersectional subgroups, and collecting clinician-patient linkages as necessary for label generation by clinicians. Those responsible for AI models should require such audits before model deployment and mediate between model auditors and impacted stakeholders.

Open questions and research gaps for monitoring and updating AI-enabled tools in clinical settings

As the implementation of artificial intelligence (AI)-enabled tools is realized across diverse clinical environments, there is a growing understanding of the need for ongoing monitoring and updating of prediction models. Dataset shift-temporal changes in clinical practice, patient populations, and information systems-is now well-documented as a source of deteriorating model accuracy and a challenge to the sustainability of AI-enabled tools in clinical care. While best practices are well-established for training and validating new models, there has been limited work developing best practices for prospective validation and model maintenance. In this paper, we highlight the need for updating clinical prediction models and discuss open questions regarding this critical aspect of the AI modeling lifecycle in three focus areas: model maintenance policies, performance monitoring perspectives, and model updating strategies. With the increasing adoption of AI-enabled tools, the need for such best practices must be addressed and incorporated into new and existing implementations. This commentary aims to encourage conversation and motivate additional research across clinical and data science stakeholders.

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

IMPORTANCE

Despite the potential of machine learning to improve multiple aspects of patient care, barriers to clinical adoption remain. Randomized clinical trials (RCTs) are often a prerequisite to large-scale clinical adoption of an intervention, and important questions remain regarding how machine learning interventions are being incorporated into clinical trials in health care.

OBJECTIVE

To systematically examine the design, reporting standards, risk of bias, and inclusivity of RCTs for medical machine learning interventions.

EVIDENCE REVIEW

In this systematic review, the Cochrane Library, Google Scholar, Ovid Embase, Ovid MEDLINE, PubMed, Scopus, and Web of Science Core Collection online databases were searched and citation chasing was done to find relevant articles published from the inception of each database to October 15, 2021. Search terms for machine learning, clinical decision-making, and RCTs were used. Exclusion criteria included implementation of a non-RCT design, absence of original data, and evaluation of nonclinical interventions. Data were extracted from published articles. Trial characteristics, including primary intervention, demographics, adherence to the CONSORT-AI reporting guideline, and Cochrane risk of bias were analyzed.

FINDINGS

Literature search yielded 19 737 articles, of which 41 RCTs involved a median of 294 participants (range, 17-2488 participants). A total of 16 RCTS (39%) were published in 2021, 21 (51%) were conducted at single sites, and 15 (37%) involved endoscopy. No trials adhered to all CONSORT-AI standards. Common reasons for nonadherence were not assessing poor-quality or unavailable input data (38 trials [93%]), not analyzing performance errors (38 [93%]), and not including a statement regarding code or algorithm availability (37 [90%]). Overall risk of bias was high in 7 trials (17%). Of 11 trials (27%) that reported race and ethnicity data, the median proportion of participants from underrepresented minority groups was 21% (range, 0%-51%).

CONCLUSIONS AND RELEVANCE

This systematic review found that despite the large number of medical machine learning-based algorithms in development, few RCTs for these technologies have been conducted. Among published RCTs, there was high variability in adherence to reporting standards and risk of bias and a lack of participants from underrepresented minority groups. These findings merit attention and should be considered in future RCT design and reporting.

Pseudoprospective Paraclinical Interaction of Radiology Residents With a Deep Learning System for Prostate Cancer Detection: Experience, Performance, and Identification of the Need for Intermittent Recalibration

OBJECTIVES

The aim of this study was to estimate the prospective utility of a previously retrospectively validated convolutional neural network (CNN) for prostate cancer (PC) detection on prostate magnetic resonance imaging (MRI).

MATERIALS AND METHODS

The biparametric (T2-weighted and diffusion-weighted) portion of clinical multiparametric prostate MRI from consecutive men included between November 2019 and September 2020 was fully automatically and individually analyzed by a CNN briefly after image acquisition (pseudoprospective design). Radiology residents performed 2 research Prostate Imaging Reporting and Data System (PI-RADS) assessments of the multiparametric dataset independent from clinical reporting (paraclinical design) before and after review of the CNN results and completed a survey. Presence of clinically significant PC was determined by the presence of an International Society of Urological Pathology grade 2 or higher PC on combined targeted and extended systematic transperineal MRI/transrectal ultrasound fusion biopsy. Sensitivities and specificities on a patient and prostate sextant basis were compared using the McNemar test and compared with the receiver operating characteristic (ROC) curve of CNN. Survey results were summarized as absolute counts and percentages.

RESULTS

A total of 201 men were included. The CNN achieved an ROC area under the curve of 0.77 on a patient basis. Using PI-RADS ≥3-emulating probability threshold (c3), CNN had a patient-based sensitivity of 81.8% and specificity of 54.8%, not statistically different from the current clinical routine PI-RADS ≥4 assessment at 90.9% and 54.8%, respectively ( P = 0.30/ P = 1.0). In general, residents achieved similar sensitivity and specificity before and after CNN review. On a prostate sextant basis, clinical assessment possessed the highest ROC area under the curve of 0.82, higher than CNN (AUC = 0.76, P = 0.21) and significantly higher than resident performance before and after CNN review (AUC = 0.76 / 0.76, P ≤ 0.03). The resident survey indicated CNN to be helpful and clinically useful.

CONCLUSIONS

Pseudoprospective paraclinical integration of fully automated CNN-based detection of suspicious lesions on prostate multiparametric MRI was demonstrated and showed good acceptance among residents, whereas no significant improvement in resident performance was found. General CNN performance was preserved despite an observed shift in CNN calibration, identifying the requirement for continuous quality control and recalibration.

Recent Advances and Future Perspectives in the Use of Machine Learning and Mathematical Models in Nephrology

We reviewed some of the latest advancements in the use of mathematical models in nephrology. We looked over 2 distinct categories of mathematical models that are widely used in biological research and pointed out some of their strengths and weaknesses when applied to health care, especially in the context of nephrology. A mechanistic dynamical system allows the representation of causal relations among the system variables but with a more complex and longer development/implementation phase. Artificial intelligence/machine learning provides predictive tools that allow identifying correlative patterns in large data sets, but they are usually harder-to-interpret black boxes. Chronic kidney disease (CKD), a major worldwide health problem, generates copious quantities of data that can be leveraged by choice of the appropriate model; also, there is a large number of dialysis parameters that need to be determined at every treatment session that can benefit from predictive mechanistic models. Following important steps in the use of mathematical methods in medical science might be in the intersection of seemingly antagonistic frameworks, by leveraging the strength of each to provide better care.

Machine Learning for Acute Kidney Injury Prediction in the Intensive Care Unit

Machine learning is the field of artificial intelligence in which computers are trained to make predictions or to identify patterns in data through complex mathematical algorithms. It has great potential in critical care to predict outcomes, such as acute kidney injury, and can be used for prognosis and to suggest management strategies. Machine learning can also be used as a research tool to advance our clinical and biochemical understanding of acute kidney injury. In this review, we introduce basic concepts in machine learning and review recent research in each of these domains.