Authors (including presenting author) :
Leung MKM(1), Lam RPK(2), Cheung KMH(1), Chin JW(3), Wong KL(3), Qiu CC(4), Woo JCY(2), Rainer TH(2), Tsang TC(1)
Affiliation :
(1)Accident & Emergency Department, Queen Mary Hospital
(2)Department of Emergency Medicine, HKUMed
(3)Department of Clinical AI, PanopticAI
(4)Department of Industrial Engineering and Decision Analytics, HKUST
Introduction :
Automating vital signs acquisition can increase the operational efficiency of accident and emergency departments (A&E). Smartphone-based remote photoplethysmography (rPPG) enables the estimation of vital signs by using artificial intelligence (AI) algorithms to analyse the subtle changes of light reflected from the skin in facial videos. However, its accuracy in A&E settings remains unclear.
Objectives :
To evaluate (1) the accuracy of a proprietary AI-based rPPG in contactless estimation of heart rate (HR), respiratory rate (RR), SpO2, blood pressure (BP); and temperature; and (2) patient comfort and satisfaction with different measurement methods.
Methodology :
A prospective observational cross-sectional study was conducted at Queen Mary Hospital A&E. Adult patients of triage category 4 (Semi-urgent) and 5 (Non-urgent) were recruited. Vital signs were measured manually by a research nurse using standard hospital equipment as reference. Simultaneously, 25-second facial videos were recorded using an iPhone 14 for contactless estimation of vital signs (temperature from an infrared camera). The accuracy of contactless estimations was evaluated using Pearson correlation coefficient (r) and root mean square error (RMSE). Patient satisfaction and comfort, assessed using a visual analogue scale, were compared between measurement methods using the Wilcoxon signed-rank test.
Result & Outcome :
From October to November 2024, 360 videos were obtained from 126 patients (79 women and 47 men; mean age 54 years). Contactless HR estimation had a high level of agreement with manual measurement (r 0.992, p< 0.01, RMSE 1.82 bpm). The respective values for other vital signs were: RR (r 0.589, p< 0.01, RMSE 3.48 breaths/min), SpO2 (r 0.173, p< 0.01, RMSE 1.65%), systolic BP (r 0.710, p< 0.01, RMSE 15.77 mmHg) and diastolic BP (r 0.677, p< 0.01, RMSE 7.85 mmHg), temperature (r 0.555, p< 0.01, RMSE 0.28°C). Comfort and satisfaction ratings were significantly higher with contactless measurement compared to manual measurement.
Conclusions:
AI-based rPPG technology is highly accurate in HR estimation, but contactless measurement of other vital signs in A&E requires additional work to improve the generalizability and robustness. This proof-of-concept project demonstrates the great potential of AI-based rPPG in automating vital sign acquisition in clinical settings, which can save manpower and reduce infection risks.
