A novel interpretable feature set optimization method in blood pressure estimation using photoplethysmography signals

• Published in: Biomedical signal processing and control 2023-09, Vol.86, p.105184, Article 105184
• Main Authors: Liu, Jian, Hu, ShuaiCong, Xiao, Zhijun, Hu, Qihan, Wang, Daomiao, Yang, CuiWei
• Format: Article
• Language: English
• Subjects: Blood pressure estimation; Feature optimization; Photoplethysmography (PPG); Shapley Additive Explanations
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2023.105184

[Display omitted] •The paper extracts features from PPG and its derivatives, while also incorporating statistical information from subjects. All 172 features from 10 dimensions are provided in Appendix A.•The SHAP algorithm is utilized to enhance the interpretability of the feature optimization process.•The optimized LightGBM model is combined with SHAP to calculate the SHAP values in each feature.•The paper proposes an interpretable method for feature importance ranking. The most appropriate number of features is obtained based on the performance of the model and feature importance. Appendix B provides the ranking of all features.•The results show that all evaluation metrics of the model improved after feature optimisation, achieving accurate SBP and DBP estimation. Blood pressure (BP) estimation based on photoplethysmography (PPG) signals enables continuous and comfortable BP measurement, which is important for the clinical management of hypertension. The purpose of this study is to propose a novel and interpretable feature optimization method to improve the performance of the PPG-based model for BP estimation. The PPG signals of 152 subjects were selected from a public database. Feature detection was performed on the signals after FIR band-pass filtering. A total of 172 features extracted from ten feature dimensions were used to construct a feature set, including features from the raw PPG signals, PPG derivative signals, and statistical information. Light Gradient Boosting Machine (LightGBM) was used as this work's prediction model. To further improve the performance of the LightGBM model, Shapley Additive Explanations (SHAP) were utilized for feature optimization to achieve the purpose. The number of features corresponding to the lowest model error was defined as the optimal feature set, and this was utilized for training the model to obtain BP estimation. Compared to the validated BP, the mean and standard deviation (SD) of the estimation errors for systolic blood pressure (SBP) and diastolic blood pressure (DBP) were −0.73 ± 6.50 mmHg and 0.37 ± 3.83 mmHg, respectively. According to the British Hypertension Society (BHS) criteria, SBP and DBP are within the range of B and A grades, respectively. We propose a novel feature optimization method to reduce the feature dimension for BP estimation. The algorithm can effectively prevent overfitting and improve the model's performance.