Objective Addressing the personalized differences in motion gait between patients and healthy individuals, as well as the issue of abnormal gait recognition, this paper proposes an effective solution based on deep learning. Methods The study adopts a model that fuses Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory Networks (BiLSTM) networks. It collects unilateral ankle joint movement data at different walking speeds within a comfortable range from healthy elderly individuals and stroke patients. The model uses signals from inertial sensors and electromyography sensors as inputs, while also analyzing gait features and comparing gait differences between the two groups. The effectiveness of the model is validated by comparing the classification performance of GRU, LSTM, and CNN-BiLSTM models in abnormal gait recognition. Results The CNN-BiLSTM model performs excellently in abnormal gait recognition, achieving an overall accuracy of 95.40%. Conclusion The results suggest that the algorithm proposed in this study can be applied to early disease detection and other related fields, providing support for early diagnosis and precise monitoring of diseases.