Explainable Machine Learning Techniques To Predict Amiodarone-Induced Thyroid Dysfunction Risk: Multicenter,...

Area of Science:

• Computational medicine and bioinformatics.
• Development and validation of predictive models for adverse drug reactions.

Background:

• Amiodarone-induced thyroid dysfunction is a life-threatening and unpredictable adverse effect.
• Traditional regression models have shown suboptimal performance in predicting such adverse effects.
• Machine learning algorithms utilizing time-series data offer potential for improved prediction accuracy.

Purpose of the Study:

• To develop and validate machine learning models for forecasting individual risk of amiodarone-induced thyroid dysfunction.
• To optimize a machine learning-based risk stratification scheme using resampling and adjusted decision thresholds.

Main Methods:

• Utilized multicenter electronic health records from patients receiving amiodarone (2013-2017).
• Developed 16 machine learning models, including extreme gradient boosting, incorporating stationary and dynamic features over time.
• Employed resampling techniques and readjusted clinical decision thresholds to optimize model performance, evaluated using metrics like AUROC and AUPRC.

Main Results:

• The extreme gradient boosting oversampling model achieved the highest predictive performance (AUROC: 0.934).
• Optimized decision threshold (0.627) improved F1-score to 0.699, correctly identifying 275 true-positive high-risk patients.
• Key predictors included shorter treatment duration, higher TSH and HDL-C, and lower free T4, alkaline phosphatase, and LDL levels.

Conclusions:

• Machine learning models, enhanced by resampling methods, can accurately predict amiodarone-induced thyroid dysfunction.
• These models provide a valuable tool for individualized risk prediction and clinical decision support in managing amiodarone therapy.