Probability of somatic disease detection in metallurgic production workers using gradient boosting and fuzzy logic methods

The study object was represented by metallurgical production workers undergoing periodic medical examinations under exposure to a complex of occupational and work-related factors. The aim of the study was to assess the probability of detecting somatic pathology in metallurgical production workers exposed to a complex of occupational and non-occupational factors using machine learning, in order to scientifically substantiate approaches to forming preventive observation groups.

A cross-sectional epidemiological study included 1,899 workers. Age, length of service in harmful working conditions, smoking status, and 32 binary occupational exposure variables were used as input predictors. Clinical, laboratory, and instrumental indicators used for diagnostic verification were excluded from the model to prevent artificial overestimation of predictive accuracy. CatBoost was applied for prediction, SHAP analysis for interpreting predictor contributions, and ANFIS for formalizing the identified dependencies. Model performance was assessed using AUC-ROC, sensitivity, and specificity.

Musculoskeletal diseases and circulatory system diseases prevailed in the structure of detected pathology. The best dis-criminative ability was obtained for diseases of the ear and mastoid process (AUC-ROC = 0.820), musculoskeletal diseases (0.746), and circulatory system diseases (0.722). Models for respiratory, digestive, eye, and nervous system diseases did not demonstrate sufficient predictive performance for practical interpretation. SHAP analysis showed that age made the greatest contribution to prediction; among occupational and behavioral factors, work hardness, occupational noise, and smoking were the most significant. ANFIS modeling made it possible to construct length-of-service risk profiles and identify factor combinations associated with an increased probability of ear, musculoskeletal, and circulatory system diseases.

The proposed approach can be used for preliminary risk stratification during periodic medical examinations, formation of follow-up groups, and substantiation of indications for more detailed examination according to the most relevant pathology profile.