Explainable Machine Learning With Hybrid Feature Selection for Thyroid Disease Classification: A Case Study in Bangladesh

ABSTRACT Thyroid disease is a prevalent endocrine disorder that requires early and accurate diagnosis for effective treatment. Machine learning (ML) techniques have demonstrated remarkable potential in this domain, offering accurate and efficient diagnostic tools. This study explores a systematic approach involving Exploratory Data Analysis (EDA), feature engineering, and extensive model evaluation to detect thyroid disease. A total of 348 patient records collected at the National Institute of Nuclear Medicine and Allied Sciences (NINMAS), Dhaka, Bangladesh, between October and December 2024 were analyzed. The dataset comprised 217 euthyroid, 51 clinical hyperthyroid, 33 subclinical hyperthyroid, 28 clinical hypothyroid, and 19 subclinical hypothyroid cases. The dataset was partitioned into 70% training and 30% testing sets, with 10‐fold stratified cross‐validation performed on the training set for robust model development and evaluation. This study evaluates conventional and ensemble algorithms for multiclass thyroid disease diagnosis, encompassing euthyroid, hyperthyroid, hypothyroid, subclinical hypothyroid, and subclinical hyperthyroid categories. Validation employed patient‐disjoint partitioning, stratified cross‐validation, and an independent hold‐out set for robust generalization. In the full‐feature baseline, CatBoost and Bagging Ensemble classifiers attained a test accuracy of 96.15%, with macro‐averaged F1 of 94.90%, weighted F1 of 96.09%, Cohen's kappa of 93.29%, and Matthews correlation coefficient of 93.37%. A Variance Threshold (VT) with Mutual Information Gain (MI) feature selection framework preserved these metrics for the Bagging Ensemble while reducing dimensionality, enhancing interpretability without compromising accuracy. Explainable AI (XAI) techniques, SHAP and LIME, were used to interpret feature importance and model behavior, ensuring transparency in decision‐making. Interpretability analysis revealed TSH and free thyroxine (FT4) as the most influential biomarkers driving model outputs. Although constrained by limited sample size and single‐center data acquisition, the framework achieved robust internal validation metrics. These results, while promising, remain provisional and require external validation on larger, multi‐institutional cohorts to establish clinical generalizability.