Performance metrics for intervention-triggering prediction models do not reflect an expected reduction in outcomes from using the model
This work addresses a critical issue for clinical researchers and practitioners by highlighting the limitations of current evaluation methods in predicting real-world utility, making it an incremental but important contribution to the field of medical AI.
The paper tackles the problem that standard machine learning metrics for clinical risk prediction models do not reflect actual reductions in patient outcomes when models are used to trigger interventions, especially in dynamic settings. It synthesizes evaluation approaches, showing that without interventional data, estimates are either meaningless, require strong assumptions, or provide only best-case bounds, as demonstrated with simulated and real data.
Clinical researchers often select among and evaluate risk prediction models using standard machine learning metrics based on confusion matrices. However, if these models are used to allocate interventions to patients, standard metrics calculated from retrospective data are only related to model utility (in terms of reductions in outcomes) under certain assumptions. When predictions are delivered repeatedly throughout time (e.g. in a patient encounter), the relationship between standard metrics and utility is further complicated. Several kinds of evaluations have been used in the literature, but it has not been clear what the target of estimation is in each evaluation. We synthesize these approaches, determine what is being estimated in each of them, and discuss under what assumptions those estimates are valid. We demonstrate our insights using simulated data as well as real data used in the design of an early warning system. Our theoretical and empirical results show that evaluations without interventional data either do not estimate meaningful quantities, require strong assumptions, or are limited to estimating best-case scenario bounds.