Feature-Based Design of Priority Queues: Digital Triage in Healthcare

We study data-driven classification where a classifier assigns jobs (e.g., patients or medical images) based on observed features to priority queues for human review. Traditional classifiers are designed to minimize misclassification loss functions but may underperform when integrated with workflows that can be modeled by queueing systems for two key reasons: First, conventional loss functions do not capture the externalities inherent in queueing systems that amplify the impact of classification errors: misclassifying an urgent patient as non-urgent impacts the wait of other patients classified as urgent and non-urgent. The second problem is that the queueing system design (which includes the number of priority queues and the assignment of job types to priority queues) is typically optimized ex-ante assuming perfect classification. We propose an integral approach where the classifier and/or the queueing system are optimized to minimize the workflow's average waiting cost. We demonstrate the value of our approach using a real data-set covering 560,486 patient visits to three emergency rooms over three years. We theoretically characterize the optimal number of priority queues and the optimal prioritization policy as a function of the classifier's accuracy for tractable problem instances. Compared to the use of off-the-shelf classifiers, the integral approach significantly reduces average delay costs in highly utilized systems with significant heterogeneity in delay costs. It compensates for classifier inaccuracies by assigning job types differently and to fewer priorities than the traditional approach. Competing Interest Declaration: The authors declare that they have no competing financial, professional, or personal interests that might have influenced the design and the results of this study.