A Machine learning approach for early detection of thyroid disorders in Sri Lankan women

Abstract

Thyroid disease is a significant global health issue, particularly affecting women’s metabolism, hormonal balance, and well-being. Detecting and treating the disease early is key to controlling its progression and avoiding complications. Traditional diagnostic methods like symptom-based assessments and blood tests are often hard to interpret due to complex and voluminous clinical data. Recently, machine learning techniques have shown promising potential in enhancing early diagnostic capabilities. This study presents a machine learning-based thyroid disease diagnosis system for Sri Lankan women, using 402 clinical samples from hospitals in the Uva and Western provinces. The system classifies individuals into three diagnostic categories: hyperthyroidism, hypothyroidism, and healthy. Feature selection using SelectKBest, refined with clinical expert input, identified eight key features from the initial 22 attributes. Four machine learning models, RF, ANN, DT, and SVM were fine-tuned and evaluated. Among them, the RF classifier achieved the highest performance, with an accuracy of 91%, and precision, recall, and F1-score of 90%. A comprehensive Sri Lankan women’s thyroid disease dataset was also compiled, offering a valuable foundation for future research and public health analysis. The proposed system shows the potential of machine learning for early and accurate thyroid disease diagnosis, especially in resource-limited settings like Sri Lanka.