1030-P: Feature Selection of Gait Measures Discriminating Prediabetes from Healthy Controls Using Machine Learning

Introduction & Objective: Key challenges in using body-worn inertial sensors to monitor gait characteristics are an excessive number of measures. This study aimed to determine the best combination of gait measures to discriminate prediabetes from healthy control (HC) subjects.

Methods: In this study, 108 individuals with prediabetes (age: 71.20 ± 5.11 years) and 63 healthy controls (age: 70.40 ± 6.25 years) undertook a 400-meter fast walk test. Wearing Opals inertial sensors from APDM Wearable Technologies, participants had six sensors attached to their feet, wrists, sternum, and lumbar regions. Employing the best subset selection method with forward stepwise regression, 70% of the training dataset was used for feature selection, validated with the remaining 30% for testing the model's performance, assessed by the area under the curve (AUC), accuracy, and F1 score on the validation dataset.

Results: From a pool of 37 gait measures, the best subset identified three measures from the training dataset: Cadence, total number of turns, and transverse range of motion of the trunk. A logistic regression model was built with these three gait measures using the training dataset and tested on the validation dataset. The logistic regression model on validation dataset showed AUC=0.77 (sensitivity=0.65 and specificity=0.85), accuracy=0.78 and F1 score=0.68. The performance of the proposed model with these 3 gait measures is similar to performance of model train and tested using all 37 gait measures.

Conclusion: The optimal combination of gait measures distinguishing prediabetes from HC gait involves three measures from Feet and Lumbar locations. These insights deepen our understanding of prediabetes-related gait deficits, aiding informed clinical decisions and intervention assessments.