Gait acceleration-based diabetes detection using hybrid deep learning

Diabetes is a medical condition affecting multiple organs and systems due to high blood glucose levels in the body which cause diabetic neuropathy and diabetic foot ulcers. Conventionally, diabetes is detected using invasive methods such as pricking the finger and measuring blood glucose. However, invasive methods are not convenient and can cause pain to patients. An alternative method to detect diabetes is to use gait analysis as abnormalities can be analysed in gait patterns to predict diabetes severity. To our best knowledge, no studies have investigated the use of gait acceleration for diabetes detection using hybrid deep learning models. Current research utilises hybrid models with non-gait data such as electrocardiography data and the Pima Indians Diabetes Database. This paper aims to classify diabetes by utilising acceleration data from wearable sensors placed on the hip, knees, and ankles, employing a hybrid deep learning model CNN-LSTM. The proposed CNN-LSTM model consists of two convolutional layers and two LSTM layers. By combining the two models, CNN-LSTM can extract important features and learn patterns for classification. The performance of CNN-LSTM is compared with CNN and LSTM models using accuracy, precision, recall, F1 score, AUC and ROC. Compared to existing methods, CNN-LSTM has achieved a higher accuracy of 91.25%, surpassing that of current methods. Hence, this paper demonstrates that non-invasive techniques for diabetes detection hold the potential to replace traditional invasive methods. In the future, muscle activation and muscle forces can be investigated together with gait acceleration to improve the model performance on diabetes detection.