Ewout W. Steyerberg

Ben Van Calster, Gary S. Collins, Andrew J. Vickers et al.

A myriad of measures to illustrate performance of predictive artificial intelligence (AI) models have been proposed in the literature. Selecting appropriate performance measures is essential for predictive AI models that are developed to be used in medical practice, because poorly performing models may harm patients and lead to increased costs. We aim to assess the merits of classic and contemporary performance measures when validating predictive AI models for use in medical practice. We focus on models with a binary outcome. We discuss 32 performance measures covering five performance domains (discrimination, calibration, overall, classification, and clinical utility) along with accompanying graphical assessments. The first four domains cover statistical performance, the fifth domain covers decision-analytic performance. We explain why two key characteristics are important when selecting which performance measures to assess: (1) whether the measure's expected value is optimized when it is calculated using the correct probabilities (i.e., a "proper" measure), and (2) whether they reflect either purely statistical performance or decision-analytic performance by properly considering misclassification costs. Seventeen measures exhibit both characteristics, fourteen measures exhibited one characteristic, and one measure possessed neither characteristic (the F1 measure). All classification measures (such as classification accuracy and F1) are improper for clinically relevant decision thresholds other than 0.5 or the prevalence. We recommend the following measures and plots as essential to report: AUROC, calibration plot, a clinical utility measure such as net benefit with decision curve analysis, and a plot with probability distributions per outcome category.

Shubhayu Bhattacharyay, Ioan Milosevic, Lindsay Wilson et al.

When a patient is admitted to the intensive care unit (ICU) after a traumatic brain injury (TBI), an early prognosis is essential for baseline risk adjustment and shared decision making. TBI outcomes are commonly categorised by the Glasgow Outcome Scale-Extended (GOSE) into 8, ordered levels of functional recovery at 6 months after injury. Existing ICU prognostic models predict binary outcomes at a certain threshold of GOSE (e.g., prediction of survival [GOSE>1] or functional independence [GOSE>4]). We aimed to develop ordinal prediction models that concurrently predict probabilities of each GOSE score. From a prospective cohort (n=1,550, 65 centres) in the ICU stratum of the Collaborative European NeuroTrauma Effectiveness Research in TBI (CENTER-TBI) patient dataset, we extracted all clinical information within 24 hours of ICU admission (1,151 predictors) and 6-month GOSE scores. We analysed the effect of 2 design elements on ordinal model performance: (1) the baseline predictor set, ranging from a concise set of 10 validated predictors to a token-embedded representation of all possible predictors, and (2) the modelling strategy, from ordinal logistic regression to multinomial deep learning. With repeated k-fold cross-validation, we found that expanding the baseline predictor set significantly improved ordinal prediction performance while increasing analytical complexity did not. Half of these gains could be achieved with the addition of 8 high-impact predictors (2 demographic variables, 4 protein biomarkers, and 2 severity assessments) to the concise set. At best, ordinal models achieved 0.76 (95% CI: 0.74-0.77) ordinal discrimination ability (ordinal c-index) and 57% (95% CI: 54%-60%) explanation of ordinal variation in 6-month GOSE (Somers' D). Our results motivate the search for informative predictors for higher GOSE and the development of ordinal dynamic prediction models.

Alexandros Rekkas, David van Klaveren, Patrick B. Ryan et al.

The Predictive Approaches to Treatment Effect Heterogeneity statement focused on baseline risk as a robust predictor of treatment effect and provided guidance on risk-based assessment of treatment effect heterogeneity in the RCT setting. The aim of this study was to extend this approach to the observational setting using a standardized scalable framework. The proposed framework consists of five steps: 1) definition of the research aim, i.e., the population, the treatment, the comparator and the outcome(s) of interest; 2) identification of relevant databases; 3) development of a prediction model for the outcome(s) of interest; 4) estimation of relative and absolute treatment effect within strata of predicted risk, after adjusting for observed confounding; 5) presentation of the results. We demonstrate our framework by evaluating heterogeneity of the effect of angiotensin-converting enzyme (ACE) inhibitors versus beta blockers on three efficacy and six safety outcomes across three observational databases. The proposed framework can supplement any comparative effectiveness study. We provide a publicly available R software package for applying this framework to any database mapped to the Observational Medical Outcomes Partnership Common Data Model. In our demonstration, patients at low risk of acute myocardial infarction received negligible absolute benefits for all three efficacy outcomes, though they were more pronounced in the highest risk quarter, especially for hospitalization with heart failure. However, failing diagnostics showed evidence of residual imbalances even after adjustment for observed confounding. Our framework allows for the evaluation of differential treatment effects across risk strata, which offers the opportunity to consider the benefit-harm trade-off between alternative treatments.