Multimodal machine learning for intradialytic hypotension prediction using noncontact radar-derived vital-sign summaries: a single-center internal validation study
Author(s)
Type
Journal Article
Abstract
Background Intermittent blood-pressure (BP) and hemodialysis (HD) machine measurements limit continuously updated assessment of the intradialytic hypotension (IDH) risk. We developed and internally validated a multimodal machine-learning framework integrating noncontact radar-derived vital-sign summaries with HD machine parameters and demographics for 10-min-ahead IDH prediction. Methods This single-center study included 233 HD patients, 6147 sessions, and 132,608 observation windows. Among these, 375 windows (0.28%) preceded clinically documented IDH within 10 min, corresponding to 325 episodes in 46 patients. Models were evaluated using patient-wise nested five-fold cross-validation. The primary early-fusion model combined demographic, HD machine, and radar-derived heart- and respiratory-rate features. Regression and permutation feature importance (PFI) analyses were performed to evaluate the clinical associations and identify key contributors. Results The multimodal early-fusion model achieved an area under the receiver operating characteristic curve (AUROC) of 86.62% (95% confidence interval, 82.19%–89.49%), area under the precision-recall curve (AUPRC) of 3.33% (2.23%–6.09%), recall of 78.72% (67.81%–85.31%), specificity of 78.73% (72.90%–83.37%), and precision of 1.09% (0.71%–1.44%). It detected 260 of 325 IDH episodes (80.00%) with 4.53 false-positive prediction windows per session. A clinical baseline using recent systolic blood pressure and ultrafiltration-rate information achieved a higher AUPRC (4.21%). HD machine features accounted for most predictive information; incremental contributions from radar and demographics were modest and statistically uncertain. Conclusions Multimodal 10-min-ahead IDH risk estimation was feasible under internal validation, although most short-term predictive information was derived from HD machine parameters. Further work is needed to establish the incremental clinical value of continuous radar-derived monitoring.
Date Issued
2026-10-16
Date Acceptance
2026-10-01
Citation
IEEE Journal of Translational Engineering in Health and Medicine, 2026
ISSN
2168-2372
Publisher
IEEE
Start Page
1
End Page
1
Journal / Book Title
IEEE Journal of Translational Engineering in Health and Medicine
Copyright Statement
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
License URL
Publication Status
Published online
Date Publish Online
2026-10-16