Collaborative Research: CISE MSI: RDP: III: Towards Robust and Human-Aligned Deep Learning for Medical-Sensor Time Series

Time-series data arises frequently in medical applications such as sensor monitoring over time. While deep neural networks have been extensively employed to analyze such medical time-series data, current networks often rely on spurious features that misalign with medical expertise. The spurious correlations between time-series data features and labels are biased in observed data, which undermines the robustness of the deep network to generalize to new patients in complex and dynamic environments. The project aims to address the need for reliable deep learning in ubiquitous medical-sensor applications with the goal of technological improvement in healthcare quality. The educational plan will foster broader participation among undergraduate and graduate students - particularly within the Hispanic community in South Texas - through dedicated research, education, and outreach initiatives. The primary goal of this project is to facilitate robustness of time-series deep-learning models via systematically defining and mitigating spurious correlations between input confounders and output decisions. Specifically, this project aims to achieve the research goal by developing robust techniques via following: 1) identifying input confounders prevalent at various levels of time-series data - including point, segment, and structure levels - to understand their spurious correlations with target labels; 2) designing knowledge-editing mitigation strategies to locate neuron groups responsible for spurious correlations so as to efficiently correct them; and 3) investigating the approach in two medical-sensor applications: monitoring for Parkinson's disease and detection of falls in the elderly. The research outcome will yield open-source tools and potentially benefit a wide range of sensor-based medical-monitoring and diagnosis tasks. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.