Intelligent Classification of Intraoperative Hemodynamic Phenotypes Based on CADFlow Doppler Monitoring

Accurate and continuous assessment of intraoperative hemodynamic status is fundamental to perioperative safety. Marked intraoperative hemodynamic instability is closely associated with postoperative complications and mortality, making precise hemodynamic management a central goal of perioperative medicine. Since the 1970s, hemodynamic profiling has relied primarily on parameters obtained from the pulmonary artery catheter (PAC); the four-quadrant framework built on pulmonary artery occlusion pressure (PAOP) and cardiac index (CI) - such as the Forrester-Diamond classification - distinguishes "warm-dry", "warm-wet", "cold-dry" and "cold-wet" phenotypes and remains a classic element of critical-care teaching. However, PAC placement is an invasive procedure carrying risks such as pulmonary-artery rupture and arrhythmias, which limits its routine use in surgical patients.

Machine learning (ML) has increasingly been explored for hemodynamic phenotype classification, and unsupervised methods have shown particular promise. Geri et al. identified five cardiovascular phenotypes in septic shock by combining clinical and echocardiographic data; Guinot et al. identified three congestion-related phenotypes in ICU patients using clinical, biomarker and Doppler data; and Kouz et al. and Jian et al. identified endotypes of intraoperative hypotension in surgical cohorts using hierarchical clustering and autoencoder-based deep clustering, respectively. Nevertheless, these studies share common limitations: most depend on invasive arterial or PAC-derived parameters; they capture arterial-side flow information only and lack venous indices needed to assess preload and right-heart function; and the complexity of ML "black-box" models raises the need to benchmark them against simpler, more cost-effective approaches.

CADFlow is a novel wireless, wearable Doppler-ultrasound patch that continuously and noninvasively monitors carotid artery and jugular vein hemodynamic parameters simultaneously. Its fully noninvasive design (a probe applied to the neck) avoids the risks of arterial puncture or PAC insertion, eliminates operator-dependent variability, and provides both carotid parameters (VTI, ccFT, peak systolic velocity Vs, dP/dt) and jugular parameters (Jvti, mean jugular flow jTAV, jugular pulsatility index jPI, jugular filling index JVI). This dual arterial-and-venous information is not available from arterial-only Doppler or invasive monitoring, and the resulting multimodal feature set is expected to improve the resolution, physiological interpretability and performance of unsupervised phenotype classification.

This study aims to apply CADFlow multimodal data and unsupervised learning to automatically discover intraoperative hemodynamic phenotypes and, on this basis, to develop a real-time phenotype classification model, thereby providing new tools and evidence for intelligent perioperative hemodynamic management