A machine learning strategy for fast prediction of cardiac function based on peripheral pulse wave.

• A machine learning strategy is proposed for predicting cardiac function based on peripheral pulse waves. • Two high-quality datasets are created for health- and cardiovascular disease-subject groups. • The model enables the prediction of three cardiac function parameters fast and accurately. • The model is validated through consistency analysis and comparison with clinical measurements. Pulse wave has been considered as a message carrier in the cardiovascular system (CVS), capable of inferring CVS conditions while diagnosing cardiovascular diseases (CVDs). Clarification and prediction of cardiovascular function by means of powerful feature-abstraction capability of machine learning method based on pulse wave is of great clinical significance in health monitoring and CVDs diagnosis, which remains poorly studied. Here we propose a machine learning (ML)-based strategy aiming to achieve a fast and accurate prediction of three cardiovascular function parameters based on a 412-subject database of pulse waves. We proposed and optimized an ML-based model with multi-layered, fully connected network while building up two high-quality pulse wave datasets comprising a healthy-subject group and a CVD-subject group to predict arterial compliance (AC), total peripheral resistance (TPR), and stroke volume (SV), which are essential messengers in monitoring CVS conditions. Our ML model is validated through consistency analysis of the ML-predicted three cardiovascular function parameters with clinical measurements and is proven through error analysis to have capability of achieving a high-accurate prediction on TPR and SV for both healthy-subject group (accuracy: 85.3%, 86.9%) and CVD-subject group (accuracy: 88.3%, 89.2%). The independent sample t -test proved that our subject groups could represent the typical physiological characteristics of the corresponding population. While we have more subjects in our datasets rather than previous studies after strict data screening, the proposed ML-based strategy needs to be further improved to achieve a disease-specific prediction of heart failure and other CVDs through training with larger datasets and clinical measurements. Our study points to the feasibility and potential of the pulse wave-based prediction of physiological and pathological CVS conditions in clinical application. [ABSTRACT FROM AUTHOR]