Predicting Disease Severity using Novel Digital Mobility Outcomes: A Deep Learning Approach in COPD Patients

Background: Deep learning models have the capacity to identify intricate patterns, dependencies, and interactions that may not be adequately captured by single statistics.

Aim: To develop deep learning models to assess the capacity of Digital Mobility Outcomes (DMOs) in detecting COPD disease severity.

Methods: This is a cross-sectional study evaluating the ability of each DMO (cadence, stride duration, stride length, walking speed, single support duration) to predict COPD severity according to the GOLD stages. A total of 31,456 walking bouts collected over 7-days from 17 patients were used. Multi-layer perceptron neural network models with a 10-fold cross-validation, 3 layers with 100 neurons each, were developed to evaluate the performance of DMOs to predict disease severity. DMOs served as features, with individual models built for each DMO. Statistical descriptors (mean, median, min, max, SD, coefficient of variation, percentiles (25th, 75th, 90th, 95th), age, and height) were used as inputs for walking bouts longer than 10 and 30 sec, aggregated on a weekly level. Performance metrics included Cohen's Kappa, F1 Score, Recall, Precision, and Accuracy to evaluate the models.

Results: Models exhibited high levels of performance (Table).

Conclusions: The advanced capability of deep learning models to handle complex relationships among features (statistical descriptors for each DMO), contributed to high success in predicting disease severity.