Development and Internal Validation of an Interpretable Machine Learning Framework for Composite Cognitive-Neurological Risk Screening in Mine Workers

Development and Internal Validation of an Interpretable Machine Learning Framework for Composite Cognitive-Neurological Risk Screening in Mine Workers

Reza nazari1 Sayfaddin Moosazadeh2 Naeimeh Akbari-Gharalari3 Farshad Nezhadshahmohammad4

1) M.Sc.,Faculty of Engineering, Urmia University, Urmia, Iran
2) Associate Professor,Department of Mining Engineering, Faculty of Engineering, Urmia University, Urmia, Iran
3) Assistant Professor,Neurophysiology Research Center, Cellular and Molecular Medicine Research Institute, Urmia University of Medical Sciences, Urmia, Iran
4) Associate Professor,Department of Mining Engineering, Faculty of Engineering, Urmia University, Urmia, Iran

IOI Code : AAHF-ABKKAH
Publication : National Conference on Architecture and Art Technologies, Smart City, Soft and Creative Industries and Technological Culture(icact.ir)
Abstract :
    Background and Aim: Mine workers are exposed to a range of occupational hazards, including mineral dusts, metals and metalloids, and psychosocial job stressors. The cognitive and neurological consequences of such exposures may initially manifest as mild, nonspecific symptoms, making broad surveillance through specialized clinical assessments often impractical due to cost, time, and limited access. This study aimed to develop and internally validate a non-invasive, interpretable framework for multidimensional neurocognitive risk screening and classification among mine workers. Methods: This descriptive-analytical study was conducted on 242 workers from three mines in Iran. Demographic and occupational data were collected alongside three instruments: the Q16, GHQ-28, and MNSI. To define an operational screening label, each participant s status on all three instruments was determined using established cut-off points (Q16 ≥ 6, GHQ-28 ≥ 51, and MNSI ≥ 7) and subsequently aggregated through majority voting; a participant classified as positive on at least two instruments was assigned to the higher-risk class. After encoding and standardization, the training set was balanced using the Synthetic Minority Over-sampling Technique (SMOTE). A 90–32–90 autoencoder was then employed for latent representation learning, and its output was fed into five base classifiers: random forest, XGBoost, support vector machine, K-nearest neighbors, and decision tree. Finally, a stacked ensemble model combined these classifiers, and SHapley Additive exPlanations (SHAP) were used for model interpretation. Results: A total of 155 participants (64.0%) were assigned to the lower-risk class and 87 (36.0%) to the higher-risk class. Using a random seed of 42, the data were partitioned into 194 training and 48 test samples. The training set initially contained 155 class-zero and 39 class-one samples; after applying SMOTE, it was balanced to 311 samples (155 class-zero and 156 class-one). The stacked ensemble model achieved an accuracy of 95.83%, precision of 90.91%, sensitivity of 90.91%, specificity of 97.30%, an F1 score of 90.91%, and an area under the curve of 0.9902 on the test set, with only two misclassifications recorded. Compared with the best-performing base model—XGBoost, which attained 93.75% accuracy—the ensemble model demonstrated superior performance. The latent representation learning stage yielded a 3.1 percentage point improvement in accuracy relative to the baseline approach. In the SHAP analysis, Enc_10, Enc_31, and Enc_13 exhibited the highest importance; among the original features, MNSI and Q16 showed correlations of 0.769 and 0.758 with the target label, respectively. Conclusion: The findings suggest that integrating multidimensional screening data with machine learning and SHAP-based interpretability can provide a foundation for a decision-support tool in mining occupational health practice. Nevertheless, the target label in this study represents an operational questionnaire-based reference rather than a gold standard or independent clinical diagnosis. Accordingly, the present results should be interpreted as an internal validation of a screening outcome classification and warrant external validation and comparison against independent clinical assessment before any clinical application or large-scale implementation.

Keywords : mine workers neurotoxicity neurocognitive screening q16 ghq 28 mnsi machine learning stacking ensemble shap occupational health