Objective: Applying machine learning (ML) to identify if there were clusters of daily-living gait and physical activity measures that differed among responders to the two interventions of the VTIME project.

Background: In the V-TIME project, treadmill training (TT) with virtual reality (TT+VR) reduced fall rates in patients with Parkinson’s disease (PD) more than TT alone. Given this effect and since TT+VR targets both motor and cognitive aspects of mobility, we speculated that fall rate reduction would be related to specific daily-living gait and physical activity measures among those who responded to the intervention in the two groups.

Method: 103 PD patients who reported 2 or more falls in the 6 months pre-intervention were randomly assigned to one of two active interventions: TT alone or TT+VR. After 6-weeks of training, subjects wore a 3D accelerometer on the lower back for 1 week. Measures of daily-living gait quality and activity were extracted. ML models were trained on this data, with the task of classifying responders/non-responders. Subjects were considered as responders if they didn’t fall in the 6 months post-intervention, or if the fall rate was 75% lower.

Results: At baseline, TT and TT+VR subjects had similar age, gender, and disease duration (TT: n=53, 71.2±6.2 yrs, 39.6% women, duration: 9.0±7.4 yrs; TT+VR: n=61, 70.8±6.4 yrs, 39.3% women, duration: 9.1±5.4 yrs). ML achieved good classification of responders/non-responders (n=13, 43) for the TT+VR subjects (sensitivity: 0.85±0.16, specificity: 0.86±0.07, auc: 0.94).  In the TT group, the ML was not successful in classifying responders/non-responders (n=14, 33) (sensitivity: 0.62±0.18, specificity: 0.60±0.11, auc: 0.7). In the TT+VR ML models, the salient features were in domains of gait variability, amplitude, and physical activity (e.g. % of active time during the day); models for TT were mostly related to rhythm features (e.g. step time).

Conclusion: Applying ML to daily-living measures of gait and activity revealed clusters of measures that responded to TT+VR, with a disparate set related to TT. This training specific effect may reflect the richness of TT+VR. While TT focused on rhythm, the TT+VR also targeted walking in cognitively challenging conditions. This might explain the change in a cluster of behavioral measures among the responders in the TT+VR group.

« Back to 2022 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/a-motor-cognitive-intervention-impacts-daily-living-gait-and-activity-in-parkinsons-disease-fallers-differently-than-a-motor-intervention-analysis-using-a-machine-learning-approach/