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

Abstract

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. Standard applications of digital triage are sequential. First, a classifier predicts job types (e.g., diseases on a chest X-ray). Next, the jobs are fed into priority queues based on their predicted types.

This state of practice suffers from three key deficiencies: 1) classification is focused on predicting clinical types rather than their clinical priorities; 2) the design of priorities assumes perfect knowledge of the clinical types and is then taken as given. Instead, this design should account for the fact that clinical types are imperfectly observed, and prioritization will be based on estimates of the clinical types; 3) job classification is performed without considering that ``cost'' of error depends on a downstream priority-queue. The mapping of features to types is typically calibrated to minimize a standard prediction error (loss) that does not capture the externalities inherent in queuing systems and that amplify the impact of classification errors---misclassifying a non-urgent job as urgent impacts the wait of other urgent jobs.

To mitigate the deficiencies of this current practice, we propose ``direct triage'' where the classifier predicts priority queues (rather than diseases) from features and is trained to minimize the workflow's average waiting cost. For tractable problem instances, we analytically characterize the optimal number of priority queues and the optimal prioritization policy as a function of the classifier's accuracy and the server's utilization.

We demonstrate the value of our direct approach using a dataset containing over 100,000 chest 2D X-ray images labeled with clinical findings. Compared to the use of off-the-shelf classifiers,
the direct approach significantly reduces average waiting costs in highly utilized systems with non-negligible heterogeneity in delay costs.

Our work advocates for a change of practice for clinicians: label each job (patient case or image) with its correct urgency level (in addition to its correct disease or type) to train classifiers for prioritization.

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.

Note: Funding Statement: The work of the second author was supported by NSF grant DIS-1935809. Declaration of Interests: 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.

Keywords: data-driven operations, priority queues, digital healthcare, automated triage, feature-based classification, image processing, disease detection on images