Routine blood pressure (BP) monitoring, crucial for health assessment, faces challenges such as limited access to medical-grade equipment and expertise. Portable cuff BP devices, on the other hand, are cumbersome to carry all day and often cost-prohibitive in less developed countries. Besides, these sphygmomanometer-based devices can cause discomfort and disrupt blood flow during measurement. This study explores the use of smartphones for continuous BP monitoring, focusing on overcoming the trust barriers associated with the opacity of machine learning models in predicting BP from low-quality PPG signals. Our approach included developing models based on cardiovascular literature, using simple statistical methods to estimate BP from smartphone PPG signals with comprehensive data pre-processing, applying SHAP for enhanced interpretability and feature identification, and comparing our methods against standard references using Bland-Altman analysis. Validated with data from 125 participants, the study demonstrated significant correlations in waveform features between smartphone and reference BP monitoring devices. The cross-validation of linear regression [MAE=9.86 and 8.01 mmHg for systolic blood pressure (SBP) and diastolic blood pressure (DBP), respectively] and random forest model (MAE=8.91 and 6.68 mmHg for SBP and DBP) using waveform-only variables demonstrated the feasibility of using a smartphone to estimate BP. Although SHAP analysis identified key feature sets, Bland-Altman results did not fully meet established thresholds (84.64% and 94.69% of MAE<15 mmHg for SBP and DBP, respectively). The study suggests the potential of smartphone cameras to enhance the accuracy and interpretability of machine learning models for daily BP estimation, but also indicates that smartphone PPG-based BP prediction is not yet a replacement for traditional medical devices.

翻译：常规血压监测对健康评估至关重要，但面临医疗级设备与专业知识获取受限等挑战。便携式袖带血压计全天佩戴不便，且在欠发达国家往往成本过高。此外，这些基于血压计的装置在测量时会引起不适并干扰血流。本研究探索利用智能手机进行连续血压监测，重点解决因机器学习模型在基于低质量PPG信号预测血压时存在不透明性而引发的信任障碍。我们的方法包括：基于心血管文献开发模型，采用简单统计方法通过全面数据预处理从智能手机PPG信号估算血压；应用SHAP方法增强可解释性与特征识别；并通过Bland-Altman分析将本方法与标准参考值进行对比。经125名参与者数据验证，研究显示智能手机与参考血压监测设备的波形特征存在显著相关性。使用仅含波形变量的线性回归模型（收缩压MAE=9.86 mmHg，舒张压MAE=8.01 mmHg）与随机森林模型（收缩压MAE=8.91 mmHg，舒张压MAE=6.68 mmHg）的交叉验证，证明了使用智能手机估算血压的可行性。尽管SHAP分析识别出关键特征集，Bland-Altman结果尚未完全达到既定阈值（收缩压和舒张压MAE<15 mmHg的比例分别为84.64%和94.69%）。研究表明智能手机摄像头具有提升日常血压估算机器学习模型准确性与可解释性的潜力，但亦指出基于智能手机PPG的血压预测尚不能替代传统医疗设备。