Enhanced interpretable thyroid disease diagnosis by leveraging synthetic oversampling and machine learning models

Area of Science:

• Endocrinology and Metabolism
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Thyroid disorders, including hyperthyroidism and hypothyroidism, affect millions globally, impacting metabolism and quality of life.
• Early and accurate diagnosis is critical for effective management and treatment of thyroid conditions.
• Current diagnostic methods can be improved, particularly in handling imbalanced datasets.

Purpose of the Study:

• To propose an effective artificial intelligence (AI)-based approach for the early diagnosis of thyroid illness.
• To address challenges in diagnosing thyroid disorders, including class imbalance.
• To enhance the transparency and interpretability of AI diagnostic models.

Main Methods:

• Utilized an open-access thyroid disease dataset comprising 3,772 patient observations.
• Employed the nominal continuous synthetic minority oversampling technique (SMOTE-NC) for data balancing.
• Applied a fine-tuned Light Gradient Booster Machine (LGBM) technique, forming the proposed SNL (SMOTE-NC-LGBM) approach.
• Incorporated Explainable AI (XAI) using Shapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• The proposed SNL approach achieved a high accuracy performance score of 0.96, outperforming state-of-the-art methods.
• Advanced machine learning and deep learning methods were used for comparative performance evaluation.
• Hyperparameter optimizations were conducted to further enhance diagnostic performance.

Conclusions:

• The SNL approach demonstrates a significant advancement in the early and efficient diagnosis of thyroid disorders.
• The use of AI, particularly with explainable mechanisms, revolutionizes thyroid illness detection.
• This research provides a robust tool to aid specialists in overcoming thyroid disorders early.