Transport-Calibrated Ordinal Triage for Chest Radiograph Follow-Up Recommendation Under Hospital Shift

• Sajid Nadeem

  Department of Computer Science, Abasyn University, Ring Road, Charsadda Link, Peshawar 25000, Pakistan

  Author

• Kamran Yousaf

  Department of Information Technology, Hamdard University, Main Madinat al-Hikmah, Hakim Mohammed Said Road, Karachi 74600, Pakistan

  Author

Chest radiograph triage requires models to distinguish ordinary negative examinations from findings that warrant near-term review, urgent escalation, or interval follow-up. Current multimodal systems can produce fluent descriptions, yet their ordering of clinical urgency is often unstable when hospital prevalence, acquisition style, and report conventions change. This paper presents a retrospective-style empirical study of transport-calibrated ordinal triage for chest radiograph follow-up recommendation. The task was formulated as four ordered categories: no follow-up, routine follow-up, expedited follow-up, and urgent review. We evaluated 38,612 frontal chest radiograph encounters distributed across three hospital-derived domains with non-identical prevalence, label noise, and portable-image rates. A multimodal report encoder was combined with an ordinal cumulative-link head, a site-adversarial nuisance remover, and an optimal-transport calibration layer that aligned class-conditional score geometry across domains. Compared with a conventional cross-entropy classifier, the proposed model improved macro ordinal agreement from 0.612 to 0.684, reduced severe under-triage from 6.9\% to 3.8\%, and lowered domain calibration error from 0.119 to 0.063. The improvement persisted in a held-out hospital simulation, where urgent-review sensitivity increased from 81.2% to 87.6% without a significant rise in routine over-escalation. Ablation results showed that ordinal structure and transport calibration contributed separately. The findings suggest that follow-up recommendation can benefit from modeling clinical order and domain movement jointly rather than treating triage as a flat multi-class prediction problem.

• PDF