CVN+MICS Seminar: Srinivasan Parthasarathy

Abstract
Deep electrocardiogram (ECG) classifiers achieve impressive performance in detecting cardiac conditions; however, they often obscure the specific waveform regions that influence their predictions. To balance classification accuracy with clinical transparency, we present LYNX-ECG, a hierarchical visibility graph neural network that links key points on the time series based on geometric visibility. By preserving important morphological and timing anchors while significantly reducing storage requirements through a simple yet novel sparsification approach, mimicking clinical expertise, LYNX-ECG achieves an impressive macro F1 score of 0.889 on CODE-Test (one of the largest and most diverse ECG datasets of its kind), surpassing five highly competitive baseline foundation models while storage is reduced by an order of magnitude.
Because each retained node in the sparsified graph retains its precise timestamp, the model can provide region-level attribution, timing interventions, and counterfactual audits through structural deletion of clinically defined regions. In most classified exams, the clinically relevant region has the largest deletion effect, demonstrating alignment with clinical practice. Our work draws on core ideas from data mining, systems, machine learning algorithms, and visualization. It shows that sparse visibility graphs can effectively integrate high-performing ECG classification with explanation audits directly tied to clinically recognized cardiac structures.
Joint work with A. Xavier, W. Meira, M. Zaki, G. Evangelisto, A. Ribeiro, and A. Khan.
Bio
Srinivasan Parthasarathy is a University Distinguished Scholar and Professor at The Ohio State University whose research has significantly shaped modern data analysis and AI through advances in both theory and systems. He is known for influential contributions including the ECLAT family of architecture-conscious algorithms and the FRACTAL and MILE systems for large-scale graph analytics and graph representation learning. His work has received 17 best paper awards or nominations from major conferences in the field. His work has been applied across clinical healthcare, public-sector, and commercial settings, influencing tools and technologies used by millions. Recently, his research on ethical and trustworthy AI has informed discussions at the Federal Trade Commission and the Federal Communications Commission and has been shared with congressional stakeholders. He has served as PC Chair for both SIAM Data Mining and ACM SIGKDD and as steering chair for SIAM Data Mining for two terms. He is an IEEE Fellow, co-director of the Center for Responsible AI and Governance, and a Distinguished Fellow of the Robert Bosch Center for Data Science and AI and CeRAI.