Deep Learning Unravels Differences Between Kinematic and Kinetic Gait Cycle Time Series from Two Control Samples of Healthy Children Assessed in Two Different Gait Laboratories

We investigate the application of deep learning in comparing gait cycle time series from two groups of healthy children, each assessed in different gait laboratories. Both laboratories used similar gait analysis protocols with minimal differences in data collection. Utilizing a ResNet-based deep learning model, we successfully identified the source laboratory of each dataset, achieving high classification accuracy across multiple gait parameters. To address inter-laboratory differences, we explored various preprocessing methods and time series properties that may be detected by the algorithm. We found that standardization of time series values was a successful approach to decrease the ability of the model to distinguish between the two centers. Our findings also reveal that differences in the power spectra and autocorrelation structures of the datasets play a significant role in model performance. Our study emphasizes the importance of standardized protocols and robust data preprocessing to enhance the transferability of machine learning models across clinical settings, particularly for deep learning approaches.