Augmenting a ResNet + BiLSTM Deep Learning Model with Clinical Mobility Data Helps Outperform a Heuristic Frequency-Based Model for Walking Bout Segmentation

Comprehensive Summary

This study evaluates a deep learning approach that is optimized for identifying walking bouts using a single wearable inertial measurement unit (IMU), or sensor, on patients with knee osteoarthritis (OA). The deep learning model integrates a residual network (ResNet) to extract spatial features from IMU data, and bidirectional long short-term memory (BiLSTM) for temporal-dependent detection of gait events. To improve model robustness and clinical relevance, the authors progressively trained the model using PAMAP2, an open-source dataset, an additional healthy participant data, and a clinical dataset of 32 adults with OA awaiting knee and hip surgery wearing the sensor on their shank. When trained on the clinical dataset, the model was more accurate (0.96 - 0.98) compared to traditional frequency-based models which showed lower, more variable accuracy (0.87 - 0.96). Performance most notably differed at lower gait speeds of 0.3 m/s, where the frequency-based model’s recall (number of positive cases correctly identified) dropped to 0.38 compared to the deep learning model’s 0.96. Additionally, the deep learning model identified almost two times as many strides (25,279) versus the frequency-based model (14,826), signifying improved sensitivity of real-world gait patterns.

Outcomes and Implications

Knee OA is one of the most common musculoskeletal conditions, and wearable IMU sensors can offer practical applications in monitoring gait during everyday life. However, reliable identification of when patients are walking remains a clinical challenge, especially in slower or more irregular gait patterns. This study demonstrates that deep learning can not only improve accurate detection of gait, but is able to handle slower gait speeds such as those of knee OA patients or older populations. The authors note that a major challenge still is the computational cost of ResNet and BiLSTM, which may limit battery life, though future light-weighting strategies to minimize the size of the model, such as pruning or quantization, may address this issue. While the clinical data set only incorporated static activities and walking without fully free-living data, the model’s strong performance displays promise for long-term, real-world gait monitoring. Improved detection of everyday gait may allow for more reliable assessment of disease progression and response to treatment, supporting more personalized and timely care for people with knee OA.