Work

Coach-BP: Computer Vision-Based Assessment and Correction of Blood Pressure Measurement Technique

Year: 2026

Type: preprint

Abstract: Summary Background Accurate blood pressure (BP) measurement, essential for hypertension care, requires proper technique. Posture errors can raise systolic BP by more than 20 mm Hg, yet technique is rarely verified objectively. We aimed to develop and evaluate Coach-BP, a computer vision system that detects technique errors from video to support real-time coaching. Methods We enrolled adults attending an academic cardiovascular prevention center between February and April 2026 and filmed BP measurements with front and side cameras. We analyzed 408 encounters from 225 participants. Coach-BP uses machine learning models trained on clinician labels to assess five American Heart Association criteria using pose, visual context, and audio-visual features. We evaluated its accuracy against expert clinician ratings by leave-one-participant-out cross-validation. We then examined robustness to simulated camera artifacts and interpreted the features important for each criterion. Findings At baseline, 138 (62%) of 223 participants had at least one technique error and 54 (24%) two or more. Coach-BP achieved a mean area under the receiver operating characteristic curve (AUROC) of 0·93 (95% CI 0·90–0·94), with 0·97 (0·94–0·99) for legs uncrossed, 0·97 (0·93–1·00) for no talking, 0·92 (0·85–0·98) for back supported, 0·91 (0·86–0·95) for feet flat, and 0·86 (0·81–0·91) for arm supported at heart level. Mean balanced accuracy was 0·88 (0·85–0·90). Performance was robust to simulated camera shift, zoom, and rotation. Its models relied on features clinicians would assess, such as cuffed-hand height and foot position. One model classified back support similarly from either camera. Interpretation Coach-BP accurately detected errors in the five guideline-recommended criteria of BP measurement technique from clinic video, enabling objective assessment of each criterion. Funding University of California San Francisco Cardiology Scholars Award. Research in context Evidence before this study We searched PubMed and Europe PMC from database inception to Oct 10, 2026, without language restriction, for studies in adults that automatically assessed any guideline-recommended patient condition during seated upper-arm blood pressure measurement, using terms including “blood pressure”, “measurement technique”, “posture”, “computer vision”, “pose estimation”, and “sensor”. A systematic review of 328 studies and a randomized crossover trial of arm position have quantified the effect of individual errors, several of which raise systolic blood pressure by 5 mm Hg or more, and studies that observed or filmed patients measuring their blood pressure found errors in almost all patients. Three studies assessed measurement conditions automatically, all with the same sensor-equipped chair, which recorded rest time, back support, crossed legs, and ambient noise during self-measurement. They were small (41–113 participants), and only one compared the chair’s output with direct observation. In one, adherence was high only for the conditions on which the chair gave live guidance. None used video, and the chair did not assess arm or foot position. The evidence on the effect of errors includes a randomized trial, whereas the evidence on automated assessment is limited to these small observational studies of one device. Added value of this study In 225 adults contributing 408 blood pressure measurement encounters, technique errors were present at the first reading in 138 (62%) of 223 participants and did not differ significantly across the self-reported characteristics examined, including home blood pressure monitoring. Among 171 participants re-measured after staff-delivered coaching, the number with any technique error fell from 128 (75%) to 12 (7%). Coach-BP assessed all five guideline-recommended posture and behavior criteria directly from clinic video, with a mean area under the receiver operating characteristic curve (AUROC) of 0·93 under leave-one-participant-out cross-validation. Coach-BP matched or improved on rule-based baselines for all five criteria, with gains for feet flat, the arm criterion, back supported, and no talking. Performance remained stable when the image was shifted, rescaled, or rotated, suggesting robustness to differences in how the camera frames the patient. One back-support model performed similarly from the side and front cameras, so a single front camera placed as in the study could assess all five criteria. Implications of all the available evidence Technique errors were common, including among patients who already monitor blood pressure at home, and most were corrected after brief staff coaching. Coach-BP shows that five criteria of blood pressure measurement technique can be assessed automatically and accurately from ordinary video. Together, these findings support automated feedback at the time of measurement as a scalable way to improve technique, particularly at home, where no one checks it. Prospective studies should now test this in home settings and more diverse populations.

Source: medRxiv

DOI: 10.64898/2026.10.29.26366628