Triage-Time Machine Learning for Computed Tomography Acquisition Prioritization in the Emergency Department

To evaluate whether triage-time machine learning (ML) can improve computed tomography (CT) acquisition prioritization beyond first-in-first-out (FIFO) ordering and the Emergency Severity Index (ESI).

Using two emergency department datasets (Site A: 118,385 visits; Site B: 425,087 visits), we identified 43,051 and 57,728 CT patients. Gradient-boosted classifiers trained on 32 triage-time features predicted 6-hour deterioration. Because true concurrent queues could not be reconstructed, we simulated 10-, 15-, and 20-patient CT queues. Shifts were sampled until 1000 deteriorator-positive shifts were accrued per queue-size scenario. The primary endpoint was the mean deteriorator queue percentile; the secondary endpoint was the first-quartile recall. A fully crossed experiment assessed transferability.

Under FIFO, the mean deteriorator queue percentile remained near mid-queue at both sites (44.9%-47.4%). Site-specific ML reduced this to 14.2%-15.2% at Site A and 15.7%-17.1% at Site B, versus 21.2%-22.6% and 32.9%-34.8% for ESI. First-quartile recall with ML was 75.0%-78.4% at Site A, and 72.1%-76.1% at Site B. Cross-site transfer remained weakest for the Site A model applied to Site B (29.7%-30.9%). Internal models were well calibrated. At Site A, 72.8% of deteriorating CT patients had at least one affirmed actionable CT finding.

In simulation, triage-time ML robustly improves CT scanner access across plausible congestion levels. ESI captures part of the available signal, but local retraining is preferred because cross-site transfer degrades performance.