Munib Gives an Oral Presentation at ICML2025 @ AIW

In July 2025, the Actionable Interpretability workshop at ICML 2025 brought together researchers and practitioners to discuss how interpretability can drive real-world advances in artificial intelligence. The workshop encouraged contributions that demonstrate practical improvements in model robustness, alignment, and applications, and fostered dialogue around the challenges of translating interpretability research into actionable outcomes.

At the workshop, Munib presented a novel multi-modal graph learning framework for competing risks survival analysis using electronic health records (EHRs). His unified, end-to-end model learns modality-specific spatio-temporal graph representations across diverse data types, including time-series, demographics, diagnostic histories, and radiographic text, and fuses them into a patient graph. The approach introduces a composite training objective to enhance risk discrimination and calibration over time, while providing fine-grained interpretability across temporal and modality dimensions. Tested on five real-world datasets, Munib’s model outperformed leading baselines and achieved up to 8% improvements in cause-specific concordance, offering significant advances in both performance and clinical interpretability.

For more details, please refer to the video.