From treadmill to outdoor overground walking: Enhancing ground contact timing detection for older adults using transfer learning

Identification of ground contact timings (GCT) is critical for monitoring mobility in older adults. Laboratory methods are precise but limited to controlled environments, restricting their applicability in real-world settings. Treadmills allow extended measurements but fail to reflect the variability of overground walking. We evaluated the performance of deep learning models trained on treadmill data from young adults and their generalizability to treadmill and outdoor walking in older adults. We also explored transfer learning to enhance predictions by fine-tuning models with older adults' treadmill and outdoor walking data. Foot-mounted inertial measurement unit (IMU) walking data was collected from 20 young adults on treadmills and 26 older adults on treadmills and outdoor level, incline, and decline terrains. Ground truth GCTs were derived using pressure insoles (young adults) and manually-annotated motion capture (older adults). A fully connected neural network, a convolutional neural network (CNN), and a bidirectional long short-term memory network were trained on IMU data. Transfer learning was applied incrementally by fine-tuning the best-performing model with older adults' data. Model performance was evaluated on unseen outdoor data from 6 participants using F1-score and mean absolute error (MAE). The CNN achieved the highest F1-scores (0.9864 treadmill, 0.9637 outdoor level, 0.9538 incline, and 0.9029 decline walking) and the lowest MAE. Fine-tuning improved treadmill F1-scores up to n=10, while outdoor level scores plateaued at n=5. Decline walking showed poorer performance, highlighting the need for advanced modeling strategies. These findings underscore the potential of transfer learning for real-world mobility monitoring.