Trustworthy Clinical Decision Support in Endocrinology: A Multi-Class Machine Learning Framework with Counterfactual Explainability for Thyroid Nodules and Metabolic Dysfunction

• Ezekiel Abiodun

  Covenant University

  Author

Thyroid nodules and metabolic dysfunction represent two of the most prevalent and clinically challenging endocrine conditions globally, affecting over 65% of the general population for thyroid nodules and more than 500 million adults for diabetes-related metabolic disorders. Current diagnostic approaches rely heavily on invasive procedures and subjective image interpretation, while existing machine learning models for endocrine diagnostics operate as opaque "black boxes" that undermine clinician trust and fail to provide actionable clinical guidance. This study presents a comprehensive multi-class machine learning framework integrating Gradient Boosting Decision Trees (GBDT) with counterfactual explanation mechanisms to deliver accurate, interpretable, and trustworthy clinical decision support. The proposed framework was validated on a retrospective dataset of 1,649 patients with thyroid nodules, achieving an overall classification accuracy of 94.92% for multi-class thyroid disorder differentiation, with the Gradient Boosting model demonstrating superior performance (AUC = 0.82, precision = 0.814) compared to traditional approaches. Feature importance analysis identified thyroid-stimulating hormone (TSH), thyroxine (TT4), free triiodothyronine (FT3), and thyroid peroxidase antibody (TPOAB) as the most influential predictors. The counterfactual explanation module generated clinically actionable recommendations for metabolic risk modification, achieving 85.8% validity and 87.3% effectiveness in preventing hyperglycemic events. This framework addresses critical gaps in endocrine AI by combining robust multi-class classification with clinically interpretable counterfactual reasoning, offering a foundation for transparent, patient-centered decision support in endocrinology.

• pdf

Copyright (c) 2026 Ezekiel Abiodun (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.