Comparison of machine learning and logistic regression in predicting mortality from acute poisoning in young adults: a multicenter study identifying herbicide exposure as the predominant risk determinant.

OBJECTIVE

Compare the effectiveness of machine learning algorithms and traditional logistic regression in predicting the mortality risk of young patients with acute poisoning, and establish a risk stratification nomogram.

METHODS

This multicenter retrospective study derived a derivation cohort of 406 young adults with acute poisoning from Wenzhou and an external validation cohort of 150 patients from Lishui. LASSO regression was used to screen predictive factors from 43 candidate variables. Compare the predictive performance of 14 machine learning algorithms (including RandomForest, XGBoost, CatBoost, LightGBM, SVM, etc.) with logistic regression on a training set (7:3 random split). The model evaluation indicators include AUC, sensitivity, specificity, and calibration, and conduct internal and external verification.

RESULTS

LASSO identified six independent predictive factors: white blood cell count, creatinine, herbicide poisoning, invasive mechanical ventilation, liver dysfunction, and shock. The discriminative power of the final logistic regression is comparable to that of the optimal machine learning model (internal validation AUC 0.885, external validation AUC 0.971), and it is well calibrated (Brier score 0.058-0.082, Hosmer Lemeshow test p > 0.05). A nomogram for predicting the 28-day mortality risk of young patients was constructed based on Logistic regression.

CONCLUSION

Logistic regression performs similarly to complex machine learning algorithms in predicting the risk of death from acute poisoning in young adults, with better interpretability and clinical practicality. The nomogram constructed based on this is a simple and effective early risk stratification tool, which can serve as one of the reference tools for clinical decision-making assistance.