Non-invasive triglyceride detection: Using a combination of complementary multivariate photoplethysmogram features.

Objective: Long-term monitoring and effective management of triglycerides (TG) are crucial to reducing morbidity in patients with cardiovascular diseases. However, frequent invasive TG detection increases patient inconvenience and discomfort. Many efforts have been made to estimate TG based on photoplethysmogram (PPG), but current studies dissatisfy the practical application due to unclear quantitative relationship between PPG features and TG. This study investigated a framework for non-invasive TG detection based on finger PPG signals to quantify the contribution of PPG features to TG estimation. Methods: 58 features for TG detection are extracted from the PPG of 133 cardiovascular patients. The relationship between PPG features and TG is analyzed for quantitative purposes, and the pathological significance of features is further combined to remove irrelevant ones. To improve the accuracy of TG estimation, a complementary feature selection method is proposed that uses the complementary coefficient and feature stability as evaluation indicators to select the optimal feature combination from the mixed features. Finally, the Transformer algorithm strengthens the link between complementary features for TG estimation. Results: Our results demonstrated that many PPG features are moderately correlated with TG, and feature K shows the highest correlation score (r = 0.48). When using a subset of 6 selected PPG features, the performance of TG estimation based on complementary features is significantly better than that using normal mixed features, which combined with the Transformer model reduced MAE and SDE to 0.37 mmol/L and 5.26 mmol/L, respectively, and increased PCC to 0.86. Conclusion: The proposed PPG-based TG detection method clarifies the quantitative relationship between PPG features and TG and provides a new idea for non-invasive TG detection.