Energy-Efficient and Robust Algorithms for Biomedical Applications

Citation

Haghi, Benyamin Allahgholizadeh (2024) Energy-Efficient and Robust Algorithms for Biomedical Applications. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/gkx4-s019. https://resolver.caltech.edu/CaltechTHESIS:04072024-202335451

Abstract

Medical devices play a critical role in improving the quality of life for patients and assisting physicians by monitoring, detecting, and helping manage chronic conditions such as epilepsy and spinal cord injuries. To perform these functions effectively, these devices must extract the most relevant information from complex medical data. However, the functionality of these medical devices has been limited by the existing challenges in medical applications. Some of these challenges include the complexity in the analysis of raw medical data, adaptability, non-stationarity, noise, large data volumes, real-time processing, limited resources, and high accuracy demands. Moreover, considering factors such as individual differences, environmental influences, and genetic variations, medical data will cause numerous variations and uncertainties in analyzing and interpreting the medical conditions in different biomedical applications. Medical data analysis is already complex and is further complicated by issues like non-stationarity and noise, especially when using traditional and manual methods. When it comes to the designing, implementation, and utilization of wearable and implantable medical devices, efficiency, accuracy, and adaptability become crucial. Particularly, applications that require fast control of equipment, such as brain-machine interfaces (BMIs), make the need for fast decision-making evident. Medical data have been conventionally managed by reliance on extensive manual labor. However, such manual data management techniques are not scalable, have inefficient procedures, and are more likely to produce errors. Therefore, more advanced, automated methods are required immediately considering the existing challenges of the current medical data analysis techniques.

Such a shift in data processing and management will lead to more trustable procedures that can significantly improve the accuracy and efficiency of medical data analysis. Other than being just an improvement, such transformation signifies a noteworthy point in the development of medical devices. In this view, it is essential to introduce advanced technology and novel methods for medical data processing as well as automation. Therefore, it becomes critical that these high-performance and advanced techniques can efficiently be implemented with minimum effects on hardware for clinical applications. Currently, artificial intelligence (AI) and its subfield machine learning (ML) has led to major transformations in designing and utilization of various medical devices. Among all these biomedical applications, three major area are addressed in this thesis: Brain Machine Interfaces (BMIs), seizure detection, and classification of arrhythmias in cardiac rhythms. We selected these three applications due to their significance and ability to improve patient treatment further. Additionally, we showed how we used machine learning algorithms for each of these applications to address their current challenges.

In our work related to Brain-Machine Interfaces (BMIs), we have been focused on improving the quality of life for individuals with spinal cord injury (SCI) through two studies. In our initial study, we have designed and implemented a deep multi-state Dynamic Recurrent Neural Network (DRNN) decoder for BMI applications. This algorithm decodes neural data recorded from the posterior parietal cortex (PPC) and the motor cortex (M1) of human participants to appropriate control signals to predict computer cursor kinematics on the computer screen. By reducing the amount of history used in predicting the movement kinematics from the recorded neural data, we have demonstrated that improved performance and robustness are preserved while memory and power consumption are reduced. We then compared the performance of DRNN with other decoding techniques to demonstrate that when operating on wavelet-based neural features, our proposed DRNN-based decoder outperforms other decoding techniques. Therefore, DRNN have the potential to be used for more efficient and effective BMIs. After developing DRNN as a decoding technique for BMI applications, we have implemented an efficient feature extraction technique, referred to as Feature Extraction Network (FENet), which has been designed by using convolutional neural networks for optimizing feature extraction and decoding to ensure consistency across electrodes when decoding the recorded neural data to the movement kinematics in BMI systems. After being tested with data recorded from the posterior parietal and motor cortices of three human participants, FENet outperformed existing feature extraction techniques such as threshold crossings and wavelet transforms, and it significantly enhanced both closed- and open-loop cursor controls. We have also evaluated the generalizability of FENet when applied to different datasets, brain regions, and participants. Therefore, the results of our research in BMI technology have the potential to promise the improvement of the quality of life for spinal cord injury (SCI) patients.

Second, we co-designed EKGNet, a convolutional network that combines analog computing and deep learning for detecting heartbeat arrhythmia. EKGNet demonstrated high accuracy while minimizing power consumption, effectively overcoming challenges related to analog circuitry and real-time processing. The experimental findings, using PhysionNet’s MIT-BIH and PTB Diagnostics datasets, showed an average balanced accuracy of 95% for intra-patient arrhythmia classification and 94.25% for myocardial infarction (MI) classification.

Finally, we designed a real-time seizure detector by using XGboost as a technique relies on gradient boosted trees, which can help with the fast and accurate diagnosis of seizure for epileptic patients. With an averaged detection latency of 1.1 seconds, this design attained average F1 scores of 99.23% and 87.86% under various data splitting methods. The energy-area-latency product was 27× lower than the current state-of-the-art solutions, which allowed for adjustments that were specific to each patient and significantly reduced energy consumption.

The results presented in this dissertation demonstrate the potential of AI in addressing the existing challenges in three biomedical applications: brain-machine interfaces (BMI), seizure detection, and heartbeat arrhythmia detection. By addressing these existing challenges including complex biological data management, real-time processing constraints, and limited resources in biomedical applications, AI has the potential to improve the quality of life for patients suffering from neurological disorders and medical conditions. Moreover, the improved precision, operational efficiency, and flexibility caused by the integration of AI into the design of the future biomedical systems will potentially assist healthcare providers to offer enhanced support and treatment to patients. While we have focused on the three above-mentioned biomedical applications, the principles learned from our analysis may be relevant and can be extended to other biomedical applications.

Thesis Files