Redesign of telephone triage for non-traumatic chest pain using deterministic rules derived through machine learning: development and temporal validation

Objective.

To develop and validate a telephone triage system based on deterministic rules and dichotomous questions for non-traumatic chest pain, aimed at reducing undertriage and the sex-related diagnostic gap.

Method.

Retrospective observational study conducted at the Centro de Emergencias Sanitarias 061 de Andalucía (Andalusia, Spain). A total of 66,384 calls triaged using the current structured guideline (GTTA23) were included, with the prehospital clinical diagnosis used as the reference standard. The multiple-choice questions were recoded into 36 dichotomous variables. Using gradient boosting, we derived 15 emergency rules (P1, immediate advanced life support) and 8 non-emergency rules (NO-P1, deferrable cases); the remaining cases were referred for coordinator review (ESC). The development cohort comprised 53,323 episodes recorded between 2020 and 2023, and the validation cohort comprised 13,061 episodes recorded in 2024. We followed the TRIPOD+AI and SAGER guidelines.

Results.

In the validation cohort, system sensitivity (P1 ø ESC vs NO-P1) was 97.8% (95% CI, 96.96–98.42), compared with 74.0% for the current system (D = +23.82 percentage points; P < .0001). Undertriage dropped from 6.7% down to 4.0%, and the number of false negatives decreased from 415 to 35. The gap between the sexes decreased from 16.09 to 6.24 percentage points (DD = −9.85 percentage points; P = .0005). The system reduced immediate advanced life support dispatch from 52.8% to 28.5%; 35.2% of cases were resolved deterministically, while 64.8% were referred to the coordinator.

Conclusions.

The system reduces undertriage to < 5% and attenuates the sex-related diagnostic gap while preserving transparency and interpretability. Its impact on workload, coordination times, and resource use requires prospective validation.