So Many Tests, So Little Time: Operationally Constrained Machine Learning for Diagnostic Assessment under Time Constraints

In many application areas, practitioners must make diagnostic decisions under operational constraints. For example, when diagnosing concussion—one of the most common types of traumatic brain injuries—clinicians may assess patients under time constraints imposed by athletic, emergency department, or military settings. Yet, many existing machine learning (ML) techniques fail to account for such operational constraints in the construction of a diagnostic battery. Accordingly, this research proposes a novel approach to generating ML models that enforce operational constraints such as time and interpretability. Motivated by the need to incorporate stochasticity in the time associated with each concussion subtest (i.e., feature), we model this prediction problem as a chance-constrained mixed integer program and propose several reformulations to enhance tractability. We then apply our models to data from the NCAA–DoD CARE Consortium, a large, multi-site study of sports-related concussion among collegiate athletes and military cadets. Compared to diagnostic batteries commonly used in current practice, the concussion diagnosis batteries designed by our methods are both time-efficient (achieving >97% reduction in time needed to administer the battery) and highly accurate (maintaining area under the curve > 0.93). As such, our method can design practically relevant ML models that can be applied broadly across many industries and application areas.