Novel Techniques to Assess Predictive Systems and Reduce Their Alarm Burden

The performance of a binary classifier ("predictor") depends heavily upon the context ("workflow") in which it operates. Classic measures of predictor performance do not reflect the realized utility of predictors unless certain implied workflow assumptions are met. Failure to meet these implied assumptions results in suboptimal classifier implementations and a mismatch between predicted or assessed performance and the actual performance obtained in real-world deployments. The mismatch commonly arises when multiple predictions can be made for the same event, the event is relatively rare, and redundant true positive predictions for the same event add little value, e.g., a system that makes a prediction each minute, repeatedly issuing interruptive alarms for a predicted event that may never occur. We explain why classic metrics do not correctly represent the performance of predictors in such contexts, and introduce an improved performance assessment technique ("u-metrics") using utility functions to score each prediction. U-metrics explicitly account for variability in prediction utility arising from temporal relationships. Compared to traditional performance measures, u-metrics more accurately reflect the real-world benefits and costs of a predictor operating in a workflow context. The difference can be significant. We also describe the use of "snoozing," a method whereby predictions are suppressed for a period of time, commonly improving predictor performance by reducing false positives while retaining the capture of events. Snoozing is especially useful when predictors generate interruptive alerts, as so often happens in clinical practice. Utility-based performance metrics correctly predict and track the performance benefits of snoozing, whereas traditional performance metrics do not.