BIDS Showcases at the 2025 AI at Yale Symposium

On May 9, 2025, Yale’s Kline Tower was transformed into a hub of interdisciplinary innovation during the Envisioning AI at Yale: An Interdisciplinary Symposium. This university-wide event brought together thought leaders and early-career researchers to explore how artificial intelligence is shaping science, medicine, the environment, education, and the arts.

The Department of Biomedical Informatics & Data Science (BIDS) made a significant impact at the symposium, highlighting its leadership at the intersection of AI and healthcare. BIDS faculty, postdocs, students, and research staff contributed a rich range of posters, talks, and exhibits, showcasing breakthroughs in biomedical data science.

Featured BIDS Highlights

Presentation: Target-Guided Transformers: Beyond Next-Token Prediction

At the core of today’s large language models (LLMs) is the deceptively simple task of next-token prediction—guessing the next word in a sequence. While effective, this approach can restrict a model’s long-range coherence and controllability.

Stutz introduced an innovative method called target-guided transformers, which enhance next-token prediction by also learning sequence-level outcomes. This dual-task training enables the model to weigh how each token affects the broader intent or result of the sequence. She demonstrated applications in reinforcement learning (e.g., improving strategies from imperfect gameplay) and language generation (e.g., sentiment control).Looking ahead, her research suggests potential applications in clinical modeling, where patient outcomes could guide treatment recommendation systems.

Presentation: Unsupervised Segmentation of Histopathology Using Robust AI Imaging Biomarkers

In digital pathology, machine learning models often struggle with medical images that lack consistent orientation. Traditional algorithms are invariant to translation—but not rotation or reflection—limiting their reliability.

Chen proposed a symmetry-aware convolutional kernel that generates equivariant histopathological biomarkers via unsupervised segmentation. Validated using prostate tissue slides from the Gleason 2019 Challenge dataset, her method significantly improved model robustness and diagnostic consistency.This approach paves the way for more generalizable and scalable tools in pathology—potentially enhancing diagnoses across a broad range of diseases.

Poster: ClaimsGPT-Instruct: A Large Concept Model for Patient Trajectory Forecasting

Zhou explored how LLMs trained on structured electronic health record (EHR) data—such as ICD codes and medication histories—can be adapted for clinical forecasting. His model, ClaimsGPT-Instruct, demonstrated improved performance in predicting disease progression and in-hospital mortality.

The work underscores the potential of foundation models to power real-time decision support tools in medicine and public health, moving one step closer to truly intelligent, patient-centric care systems.

Poster: TEXT-WATER: Contextualizing Data for Environmental Exposure to Emerging Water Contaminants

Emerging water contaminants (ECs) are an urgent public health issue. These unregulated substances, increasingly detected in drinking water, have been linked to serious health conditions, including cancer and reproductive, neurological, and developmental disorders.

Professor Cheung demonstrated how natural language processing (NLP) methods, including LLMs and biomedical ontologies, can systematically extract exposure and risk data on ECs from the scientific literature. His work exemplifies how AI can enhance environmental health monitoring and support evidence-based policy.

To view the full list of BIDS participants and their contributions to the symposium click the links below:

• Exhibits and Resource Tables: ai.yale.edu/symposium-exhibits
• Poster Sessions: ai.yale.edu/symposium-posters#PosterSession