*Purpose: Prehabilitation is recommended for frail and pre-frail LT candidates. Although home-based programs are generalizable, their success depends on the adherence to an exercise prescription. Thus, an objective on-training adherence metric is needed. We aimed to test whether an AI algorithm could identify patterns of physical activity across supervised training of varying intensity monitored with a PAT.

*Methods: We included consecutive LT outpatients during routine physical therapy (PT) consultation. Subjects wore a PAT (Fitbit 3) while sitting and transferring between clinic areas (sedentary to light activity), or while doing a 6-min walk test (6MWT; moderate-intensity activity). Some subjects trained with an EL-FIT (Exercise and Liver FITness) app video (Borg’s self-reported intensity). Cadence (steps/minute) and heart rate reserve (HRR) were used to classify activity intensity. We used a discriminant analysis classifier to differentiate 6MWT vs. other tasks, while we used a binary decision tree to classify the minutes spent exercising with an EL-FIT video vs. all other minutes. In both cases, we used leave-one-patient-out to avoid overfitting (i.e., we trained on all subjects except for the test subject and then classified the activity on the test subject; same procedure was repeated for all).

*Results: We recorded 5544 minutes among 139 subjects (age 57±12, male 55%, MELDNa 14±7). Frailty was present in 12%, pre-frailty in 57%. Average PAT usage was 50 (24-102) min/subject, and 46 (33%) exercised with a video. Recorded time corresponded to sitting/transfer (75%), 6MWT (15%), and exercise videos (10%). The percentage of time spent per training intensity and activity type is shown in Table. Compared to baseline, 6MWT showed the greatest change for the two methods, whereas exercise videos showed a small change with HRR only (p<0.001 across cadence and HRR groups). Self-reported intensity did not correlate with cadence (rho=0.06), HRR (rho=0.09), or video difficulty rank (rho=0.07). AI was able to differentiate 6MWT vs. other activities with 91% sensitivity and 96% specificity and exercising with an EL-FIT video vs. other activities with 89% sensitivity and 24% specificity.

*Conclusions: AI can identify physical activity patterns from PAT, particularly when trainin consists of consecutive minutes of at least moderate intensity (e.g., 6MWT). Such technology can help monitor patients undergoing home-based prehabilitation.

« Back to 2022 American Transplant Congress