Class-Level Complementarity of Clinical Text and Structured Variables in Spanish Emergency Triage

Emergency department triage with the Emergency Severity Index (ESI) depends on clinician judgment and has been associated with error rates above 30%, motivating automated decision support. The free-text reason for consultation is strongly predictive. Yet, the added value of structured clinical variables, when combined with text, remains largely unexplored for Spanish-language data and for the full five-level ESI scale. We study this complementarity on 164,288 real, highly imbalanced emergency records from a Chilean regional hospital. A stacking architecture combines a domain-adapted Spanish clinical RoBERTa encoder with a CatBoost model over 16 structured variables, evaluated on an independent held-out set. Clinical text is the dominant signal (macro-F1 0.612), far above the tabular model (0.467). Adding structured variables leaves macro-F1 statistically unchanged (∆ = +0.009) but redistributes predictions (McNemar p < 10^−4), trading precision for sensitivity: C1 recall rises by 6.5 pp while C1 macro-F1 falls by 1.5 pp, and C5 recall drops from 0.228 to 0.206 even as C5 precision and macro-F1 improve. Using the ϕ coefficient with per-class recovery and loss rates, we show that the modalities are non-redundant but asymmetric: the tabular branch recovers over 45% of the text model’s errors in C2 and C3 and only 5.3% in C5, while missing 34.9% of the cases text already resolves. Complementarity is therefore strongly class-dependent, a pattern that aggregate metrics conceal.