You are here

Erdogmus is co-PI for $1M NIH Grant

May 8, 2019

ECE Professor Deniz Erdogmus is a co-PI for a $1M NIH grant, in collaboration with the University of Southern California, for a "Multimodal Signal Analysis and Data Fusion for Post-Traumatic Epilepsy".

Abstract Source: NIH

The research objective of this proposal, Multimodal Signal Analysis and Data Fusion for Post-traumatic Epilepsy Prediction, with Pl Dominique Duncan from the University of Southern California, is to predict the onset of epileptic seizures following traumatic brain injury (TBI), using innovative analytic tools from machine learning and applied mathematics to identify features of epileptiform activity, from a multimodal dataset collected from both an animal model and human patients. The proposed research will accelerate the discovery of salient and robust features of epileptogenesis following TBI from a rich dataset, collected from the Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx), as it is being acquired by investigating state-of-the-art models, methods, and algorithms from contemporary machine learning theory. This secondary use of data to support automated discovery of reliable knowledge from aggregated records of animal model and human patient data will lead to innovative models to predict post-traumatic epilepsy (PTE). This machine learning based investigation of a rich dataset complements ongoing data acquisition and classical biostatistics-based analyses ongoing in the study and can lead to rigorous outcomes for the development of antiepileptogenic therapies, which can prevent this disease. Identifying salient features in time series and images to help design a predictor of PTE using data from two species and multiple individuals with heterogeneous TBI conditions presents significant theoretical challenges that need to be tackled. In this project, it is proposed to adopt transfer learning and domain adaptation perspectives to accomplish these goals in multimodal biomedical datasets across two populations. Specifically, techniques emerging from deep learning literature will be exploited to augment data, share parameters across model components to reduce the number of parameters that need to be optimized, and use state-of-the-art architectures to develop models for feature extraction. These will be compared against established pipelines of hand-crafted feature extraction in rigorous cross-validation analyses. Developed techniques for transfer learning will be able to extract features that generalize across animal and human data. Moreover, these theoretical techniques with associated models and optimization methods will be applicable to other multi-species transfer learning challenges that may arise in the context of health and medicine. Multimodal feature extraction and discriminative model learning for disease onset prediction using novel classifiers also offer insights into biomarker discovery using advanced machine learning techniques through joint multimodal data analysis.