Development and validation of an interpretable machine learning model for predicting incident gestational hypothyroidism using clinical laboratory markers
Related Experiment Videos
Liran Shen1, Ru Liu2, Yunbiao Zhang1
1Department of Medical Laboratory Center, Shanxian Central Hospital, Heze, China.
Objective:
Traditional risk factors have limited performance for early identification of gestational hypothyroidism (GHT), and evidence remains limited on prediction using clinical laboratory markers.
Methods:
This single-center retrospective observational study included 407 pregnant women without GHT at baseline, including 164 with incident GHT and 243 without GHT during follow-up. Candidate predictors were extracted from electronic medical records and laboratory information systems. Least Absolute Shrinkage and Selection Operator (LASSO) regression, the Boruta algorithm, and Random Forest variable-importance ranking were used for feature selection. Twelve ML models were developed and compared. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), classification metrics, calibration curves, decision curve analysis (DCA), clinical impact curves, and residual analysis. Shapley Additive Explanations (SHAP) was used for interpretation.
Results:
Nine predictors were retained: zinc (Zn), iron (Fe), copper (Cu), calcium (Ca), vitamin D (vitD), vitamin E (vitE), albumin (ALB), alanine aminotransferase (ALT), and alkaline phosphatase (ALP). In the validation set, Lasso had the highest AUC [0.918; 95% confidence interval (CI), 0.871-0.964]. Light Gradient Boosting Machine (LightGBM) had a similar AUC (0.916; 95% CI, 0.866-0.965) and achieved the highest accuracy, positive predictive value (PPV), specificity, F1 score, and Youden's J statistic. LightGBM also showed acceptable calibration performance in the internal validation set, clinical net benefit, and residual distribution; therefore, it was selected as the final model. SHAP identified Zn, ALP, ALB, Cu, vitE, ALT, Fe, Ca, and vitD as the leading contributors.
Conclusion:
A LightGBM model based on clinical laboratory markers showed balanced performance for predicting incident GHT. External prospective validation is required before clinical implementation.