Exploratory evaluation of the systemic inflammatory response index and internally evaluated machine learning models for prognosis in acute pesticide poisoning.

PURPOSE: Acute pesticide poisoning (APP) is a life-threatening emergency with substantial mortality. This study aimed to evaluate whether the systemic inflammatory response index (SIRI) is associated with short-term prognosis in APP and to explore whether SIRI-containing machine learning models can support preliminary risk stratification. Because only 17 deaths occurred in this small single-center retrospective cohort and no external validation cohort was available, the analysis was designed and interpreted as exploratory rather than as development of a clinically deployable prediction model. METHODS: A total of 127 patients with APP admitted to our hospital between October 2018 and December 2024 were retrospectively enrolled, including 110 survivors and 17 non-survivors. SIRI was calculated as absolute neutrophil count × absolute monocyte count/absolute lymphocyte count. All hypothesis tests were two-sided. Missing laboratory data were handled by complete-case screening for key variables and median imputation for variables with less than 10% missingness in model development. Logistic regression, random forest (RF), and support vector machine (SVM) analyses were performed only as exploratory internal modelling analyses using a predefined 70:30 training-internal hold-out split and repeated 10-fold cross-validation within the training set. Because the hold-out set was small and contained only a few deaths, ROC-derived cut-offs, threshold-dependent sensitivity/specificity, calibration metrics, and decision-curve results were not used as clinical validation evidence. RESULTS: SIRI was lower in the non-survival group than in the survival group (2.4 ± 3.1 vs. 6.1 ± 5.8; two-sided P = 0.042), consistent with the trend shown in Figure 3A. In multivariable logistic regression adjusted for RBC, D-dimer, APTT, AST, creatinine, glucose, albumin, and pesticide type, lower SIRI remained associated with mortality (adjusted OR = 0.86, 95% CI: 0.74-0.99, P = 0.041). However, given the low number of fatal events, the regression estimates, variable selection results, and machine learning performance metrics should be interpreted as statistically unstable candidate signals rather than externally generalizable evidence. CONCLUSION: Low SIRI was associated with mortality in this single-center APP cohort after adjustment for selected confounders, and exploratory models incorporating SIRI suggested that routine hematological and inflammatory variables may contain preliminary prognostic information. Nevertheless, because the death group included only 17 patients, mechanistic immune-inflammatory assays were not available, and no independent external validation cohort was included, SIRI should be regarded as a candidate blood-count-derived prognostic signal rather than a mechanistically proven biomarker. The RBC-SIRI pattern should not be used for clinical decision-making before prospective multicenter validation, calibration updating, and mechanistic verification.