Development and validation of a Ramathibodi Poison Center scoring system to predict follow-up necessity in hospitalized patients: a retrospective study.

Introduction: The Ramathibodi Poison Center manages a substantial workload of 30,000-40,000 annual inquiries, placing significant strain on limited human resources. We aim to develop and validate a practical risk assessment score (Ramathibodi Poison Score) to predict in-hospital mortality among hospitalized poisoned patients, thereby facilitating evidence-based patient prioritization and resource allocation.

Methods: A retrospective study used a 9-year dataset partitioned into derivation cohort (70%) and internal validation cohort (30%), while a subsequent 1-year dataset (2025) served for temporal validation. Predictors were identified using multivariable logistic regression, and coefficients were transformed into a simplified integer-based score.

Results: A total of 105,342 cases were included in the primary analysis, with 5,578 cases for temporal validation. Independent predictors of mortality included pesticide exposure (the strongest predictor), unknown toxic substance, undetermined or intentional exposure, advanced age (>60 years), and male sex. The score demonstrated robust discrimination with AUC values of 0.828 (derivation), 0.831 (internal), and 0.753 (temporal). Calibration yielded an observed-to-expected (O:E) ratio of 1.000 in the derivation cohort, whereas initial temporal validation demonstrated an underestimation of mortality (O:E 1.528) due to shifting epidemiology. Following logistic recalibration, the O:E ratio in the temporal cohort was 1.000. The score stratified patients from low risk (score 0-5; 0.62% mortality, negative predictive value (NPV) 99.4%) to high risk (score ≥15; 22.22% mortality).

Discussion: The score addresses existing severity score limitations by stratifying risk before clinical deterioration. The robust NPV at the low-risk threshold confirms the safety of on-demand surveillance. Furthermore, initial temporal miscalibration highlights the dynamic nature of toxicology, emphasizing the need for periodic local recalibration against emerging threats.

Conclusion: The Ramathibodi Poison Score utilizes simple variables to predict mortality with robust discrimination. It safely identifies low-risk patients for de-escalated surveillance. With necessary periodic recalibration, it provides a dynamic, evidence-based framework for efficient resource allocation as an adjunct to clinical judgment.