Walking gait performance is a major health indicator, and gait analysis can reveal valu- able insights in medicine. The usage of wearable sensors enables pervasive gait analysis applications independent of location and time. Muscle fatigue leads to increased fall risk, and the early detection of fatigue can prevent falls of elderly people or rehabilita- tion patients. In real-life settings, people may walk differently compared to a laboratory environment since they usually walk while performing other tasks. Traditional gait analysis methods do not provide insights on a microscopic level and cannot be used for automatic gait analysis. Therefore, changes in gait characteristics caused by fatigue and dual-task are investigated in this thesis using transparent neural networks. Raw IMU gait data was collected from n = 16 young, healthy persons on nine different body locations with a fatigue and dual-task intervention. A convolutional neural net- work was used to detect fatigue and dual-task in gait in an end-to-end system using acceleration and angular velocity data from the feet and sacrum. Layer-wise Relevance Propagation (LRP) was explored to address prediction explainability which has ever since hindered the acceptance of neural networks among practitioners. The results show that gait changes are individual to the subject for both fatigue and dual-task. Within one subject, the model performs best on sacrum data. For unknown subjects, using both feet and sacrum sensors performs best. LRP patterns in fatigued gait provided valuable physiological insights and indicate that people are affected differently by fatigue. For example, the terminal swing phase was most characteristic to predict fatigue for some subjects, indicating a fatigued tibialis anterior. The thesis shows the potential of com- bining neural networks and LRP on IMU data for accurate, explainable, and pervasive gait classification.