AI Approaches for Feature Extraction in Multimodal Biomedical Data

Artificial intelligence (AI) has myriad applications for biomedical engineering and healthcare. In particular, with the diverse modalities and high resolution of physiological data, the potential for data-driven approaches to elucidate meaningful insights is enormous. However, the enormity and profundity of this data is also a hindrance. AI approaches can fail to learn due to "feature dilution", where a model is overloaded with data points, and spurious or noisy features prevent it from successfully aligning to and converging on semantic or predictive features in the dataset. Traditional methods of feature extraction either rely on manual physiological inference or leverage pre-trained AI models for well-characterized applications. We posit that AI itself can be leveraged to augment feature identification through the exploitation of learned representations and intelligent processes. To address this problem, we present three approaches for the application of artificial intelligence to the task of feature extraction in three different modes of biomedical data: 1. AxoDetect (AD), a complete image-to-model software pipeline for in silico nerve stimulation using computer vision (CV); 2. Probability Gradient Thresholding (PGT), a deep learning-based dynamic time window extraction method for the identification of neuronal compound action potential (CAP) spikes from neuronal recordings; and 3. Latent Semantic Atom Analysis (LSAA), a hybrid approach using Matching Pursuit (MP) and DeepSets (DS) to learn patient-specific representations of atherosclerosis in the coronary arteries from human heart sound data. AxoDetect excels at segmenting nerves without requiring labeled datasets, segmenting between ten to one hundred times faster than similar deep learning methods. The models which are functionalized and constructed by the software have statistically indistinguishable results when compared to the gold standard modeling methods (Wilcoxon followed by MWU with Bonferroni correction). A combined RNN-LSTM approach to PGT was able to achieve over 95% point-wise accuracy on the testing set in identifying a data point's belonging to a neuronal spike on in silico recordings. Finally, using LSAA we can not only predict presence or absence of cardiac stent with high accuracy (88% on the test set, 94\% on the full dataset), we can isolate and cluster semantic groups of atoms representing cardiac occlusion.

Name thesis:degree_name

Doctor of Philosophy

Discipline thesis:degree_discipline

Biomedical Engineering

Grantor

University of Houston

Year dc:date.issued

2024

Author dc:creator

• Lloyd, David Anderson 1996-

Advisor dc:contributor.advisor

• Akay, Metin

Committee members dc:contributor.committeemember

• Smith, Scott
• Romero-Ortega, Mario I.
• Du, Yuncheng
• Wu, Tianfu
• Akay, Yasemin

• Neural Engineering
• CAD
• Biomedical Engineering

Language dc:language.iso

English

Handle dc:identifier.uri

https://hdl.handle.net/10657/18287

OAI identifier oai:identifier

oai:uh-ir.tdl.org:10657/18287