Design and Optimization of Biomedical Signal Classification Based on Semi-supervised Learning


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date21 May 2019


Biomedical signals, which contain significant information about our body and our health, have become a promising topic for the study of disease diagnosis and rehabilitation. Biomedical signal classification can be generally regarded as a question of machine learning. Numerous methods have been proposed for solving this issue. In addition, a number of mature pattern recognition techniques, which have achieved considerable success in computer vision and other areas, are appearing in the field of biomedical signal classification. However, biomedical signal classification has its limitations and defects. Compared to image tagging, biomedical signal annotation has a much higher cost (professional knowledge is required), larger individual differences and much worse stability (due to muscle fatigue, movement interference, etc.). Conventional biomedical signal classification methods generally consist of preprocessing, feature engineering, classifier training and prediction. Feature engineering is a fundamental to the application of biomedical signal classification. Although there are a number of feature extraction methods, feature ranking and feature selection can still present a big challenge. It is difficult to provide complete information using only feature combination.

To address these problems, we concentrate on the design and optimization of classifiers based on semi-supervised learning, which can maximize biomedical annotation utilization and maintain classifier performance. In addition, we also implemented automatic feature learning in the system, which can be a substitute for manual feature engineering. The system can learn the most effective feature or the most representative patterns using a trained neural network. Therefore, we introduced a convolutional neural network (CNN) into the biomedical signal classification to learn the representative features automatically and to train the classifier efficiently.

To deal with the challenges of efficient feature extraction in electromyography (EMG) signals and classifier performance degradation, we have proposed a self-recalibrating neural-prosthesis system which can be automatically updated to maintain a stable performance over time without the need for user retraining. It can extract useful information from the dimensionally reduced spectrogram of EMG signals and use the label-updated stored data to recalibrate the classifier. It has surpassed the best performance ever obtained for a combination of four features on NinaPro database, a publicly accessible database. Our proposed system was evaluated using the NinaPro database comprising hand movement data for 40 intact and 11 amputee subjects. Our system was able to achieve ~10.18% (intact, 50 movement types) and ~2.99% (amputee, 10 movement types) increases in classification accuracy averaged over five testing sessions with respect to the unrecalibrated classifier. These results suggest that the proposed system could be a useful tool for facilitating long-term adoption of prosthetics for amputees in real-life applications.

For electrocardiography (ECG) signals, we have proposed a convolutional neural network with novel ECG input, a dual-beat coupling matrix, which captured both beat morphology and beat-to-beat correlation in the ECG. The classification system was evaluated for the detection of supraventricular ectopic beats (SVEB or S beats) and ventricular ectopic beats (VEB or V beats) using the MIT-BIH arrhythmia database. Our proposed CNN system improved the sensitivity and the positive predictive rate for S beats by more than 12.2% and 11.9%, respectively, over the top performing algorithms. We have also undertaken some further study on this proposed system. We have proposed a patient specific normal beat estimation method, which can give some personal information to the classifier. We also proposed a label update procedure based on semi-supervised learning to improve classification accuracy continuously. Our system can eliminate the requirement for personalized labeled data and could potentially be implemented on portable devices for the long-term monitoring of cardiac arrhythmia. Additionally, the proposed improved system based on semi-supervised learning could potentially be implemented on other classification models and could benefit from big data on ECG signals.

In this thesis, we also offer some thoughts about the high-efficiency expression of biomedical signals. This is because a raw biomedical signal may not be suitable for direct classifier input. In conjunction with semi-supervised learning and feature learning, we developed the classification system towards a fully-automatic and human-level system. The proposed systems could contribute to helping patients or amputees. In the meantime, they could also reduce workloads for the doctors and offer some useful insights.