An uncertainty-driven gated feature selection network (UGFS-Net) for TG level prediction: linking high-altitude exposure to lipid metabolism disorder via elevated TG.

BACKGROUND

High-altitude hypoxia is closely linked to dysregulated lipid metabolism, particularly elevated triglyceride (TG) levels, which increase cardiovascular and metabolic risks. This study proposes an interpretable deep learning model to predict TG levels in high-altitude migrants based on clinically accessible indicators.

METHODS

Data were collected from low-altitude residents (n = 96) and high-altitude migrants (n = 388). An Uncertainty-driven Gated Feature Selection Network (UGFS-Net) was developed for TG prediction, incorporating an uncertainty-driven sample re-weighting and hard example mining strategy. The model was trained via modern optimization techniques and stratified partitioning by the TG distribution. Performance was evaluated using accuracy and calibration metrics, and interpretability was assessed via SHapley Additive exPlanations (SHAP). Five benchmark machine learning models with PCA or LASSO dimensionality reduction were used for comparison.

RESULTS

The UGFS-Net demonstrated a notable performance gain through uncertainty estimation, yielding an increase in R² from 0.7294 to 0.8776 for TG levels prediction under high-altitude. Predicted uncertainty showed significant correlations with errors and effectively distinguished low- from high-reliability samples, with strong calibration (bin-wise r = 0.9164). SHAP analysis highlighted that lipid metabolism, glucose metabolism, and the erythrocyte system collectively form a network that drives TG alterations in high-altitude environments and the Pearson correlation coefficient between gated the attention weights and SHAP importance scores was 0.9093. UGFS-Net consistently outperformed conventional machine learning models.

CONCLUSIONS

This study developed UGFS-Net, an interpretable deep learning model that accurately predicts triglyceride levels in high-altitude migrants (R² = 0.8776) and provides well-calibrated uncertainty estimates, with identified key biomarkers offering clinical insights.