Professor Jean Feng, UC San Fransisco
Friday April 28, 2023
10:35 am - 11:10 am
FMonitoring machine learning (ML)-based risk prediction algorithms in the presence of confounding medical interventions
Abstract
Machine learning (ML) algorithms have the potential to derive insights from clinical data and improve patient outcomes. However, these highly complex systems are sensitive to changes in the environment and liable to performance decay. Even after their successful integration into clinical practice, ML algorithms should be continuously monitored and updated to ensure their long-term safety and effectiveness. However, performance monitoring is complicated by confounding medical interventions (CMI): When an algorithm predicts a patient to be at high risk of an adverse event, clinicians are encouraged to administer prophylactic treatment and alter the very target that the ML algorithm aims to predict. A simple approach is to ignore CMI and monitor only the untreated patients, whose outcomes remain unaltered. In general, ignoring CMI may inflate Type I error because (i) untreated patients disproportionally represent those with low predicted risk and (ii) evolution in both the model and clinician trust in the model can induce complex dependencies that violate standard assumptions. To this end, we establish ignorability conditions under which a new and simple score-based cumulative sum (CUSUM) monitoring procedure achieves valid inference. Through theoretical and empirical analyses, we show how the combination of model monitoring and updating, rather than implementing each one alone, provides a path to designing safe and effective ML systems.