Advanced deep learning techniques for biomedical signal and image analysis for healthcare

This work presents a comprehensive investigation into advanced machine and deep learning approaches for biomedical signal analysis, spanning multiple modalities and application domains. The research addresses challenges associated with noise, variability, and computational constraints in the processing of physiological data and imaging. In particular, the work encompasses: • EEG/fNIRS under Simulated Space Conditions: Novel methods for analyzing EEG and fNIRS signals are proposed to assess functional connectivity in simulated microgravity environments, providing insights into the impact of altered gravitational forces on central nervous system performance. • ECG-Based Arrhythmia Detection: Deep learning models employing stacked time–frequency scalogram images are developed for accurate classification of cardiac arrhythmias, demonstrating significant potential for early diagnosis and clinical intervention. • Facial Image-Based BMI Prediction: A lightweight ensemble framework (PatchBMI-Net) is introduced for non-intrusive estimation of body mass index from facial images, enabling real-time health monitoring via mobile platforms. • sEMG-Based Hand Gesture Recognition: The work benchmarks both traditional and deep learning classifiers for the classification of hand gestures from sEMG signals, with a focus on enhancing the control of assistive and rehabilitative devices. • Sensor Fusion for Human Activity Recognition: Innovative sensor fusion techniques are proposed to integrate multi-modal wearable sensor data, leading to improved human activity recognition performance in everyday and clinical settings. • Chest X-ray Classification: In addition, a deep learning framework for chest X- ray analysis is developed to assist in the automated detection of pulmonary abnormalities, enhancing diagnostic capabilities in radiology. Collectively, these contributions advance the field of biomedical signal analysis by addressing the challenges of data heterogeneity, computational efficiency, and real-time inference. The proposed methodologies demonstrate robust performance across diverse biomedical applications and hold promise for deployment in resource-constrained environments, thereby offering novel solutions for personalized healthcare, telemedicine, and beyond.

Author

• Aarotale, Parshuram N.

Identifier

hdl:10057/30051

OAI identifier oai:identifier

oai:soar.wichita.edu:10057/30051