Perfusion Assessment of Healthy and Injured Hands Using Video-Based Deep Learning Models.

Abstract

Assessing in-field hand trauma is challenging, and inaccurate perfusion assessment can substantially impact the patient and health system. Technology that enhances perfusion assessment could improve in-field triage. The authors present noncontact, video-based deep learning methods to classify perfused and ischemic fingers in control and acute trauma settings. The authors obtained iPhone video from 2 cohorts of subjects. The first group were control participants, some of whom were evaluated during cycles of tourniquet-induced ischemia. The second group were acutely injured patients in the authors' emergency department. For both groups, imaging photoplethysmography waveforms were extracted using a deep learning model, after which the waveform's spectrogram was classified as either perfused or ischemic using a ResNet-18 classifier. This was then compared with clinical ground-truth labels. The authors captured videos of 48 controls, including 14 evaluated during tourniquet-induced ischemia and 15 acutely injured patients. Over 5-fold cross-validation of control subjects, the authors' algorithms correctly classified ischemic finger regions with a sensitivity of 72%, a positive predictive value of 74%, and an accuracy of 90%. The authors then tested on videos of acutely injured patients, without controlling hand pose, skin cleanliness, or other variables, and achieved a sensitivity of 33%, a positive predictive value of 24%, and an accuracy of 77%. Under controlled settings, deep learning methods for perfusion classification performed well. In hospital settings-with uncontrolled lighting, hand pose, and injuries-classification performance degraded. This technology is promising but additional approaches that account for acute trauma-related variables are needed for clinical applicability as a triage tool.