AI Symposium Agenda

From silos to community:  accelerating healthcare by training AI for flexible, privacy-preserving collaboration

The ability to develop language models have enabled rapid progress across diverse domains and tasks. Yet, their adoption remains difficult in healthcare due to insufficient in-domain data and privacy constraints that restrict data sharing. Healthcare organizations face a fundamental challenge: Language models require training on large, diverse datasets, but medical data must remain private and proprietary, creating organizational silos that can limit model performance. Dr. Ali Farhadi will present FlexOLMo, a novel architecture they've developed at Ai2 that will enable healthcare institutions to collaboratively train language models without data migration and while preserving data opt-out rights. Built on fully-open architecture, code, and foundational weights trained on public data, FlexOLMo allows each data owner to contribute private modular components trained locally with clear data attribution, while their model architecture seamlessly enables opting in or out of specific sub-components at inference time, allowing organizations to flexibly configure model access based on their collaboration agreements. FlexOLMo offers a path from isolated data silos to flexible collaborative development, addressing key barriers to healthcare AI advancement while maintaining privacy and institutional control.

08:40 a.m. - 09:05 a.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Enhancing precision cancer medicine with biologically guided artificial intelligence

Precision cancer medicine, which has the overarching goal of using molecular, pathologic, and clinical data to match the patients with the optimal therapies, has begun to transform cancer care in many domains. However, there remain significant challenges implementing this strategy for patients, particularly related to (i) synthesizing all prior knowledge about molecular states that are relevant to selective treatment response, (ii) relating these properties to patient-specific molecular, pathologic, and clinical patterns, and (iii) delivering these insights in a proper manner at the point-of-care. Increasingly, novel artificial intelligence (AI) strategies grounded in biological and clinical principles are making significant impact in addressing each of these challenges. In this presentation, Dr. Eliezer Van Allen will share emerging AI technologies for enhancing these approaches, with concrete examples on how they are informing the present and future of precision cancer medicine.

09:05 a.m. - 09:30 a.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Measuring Data Privacy and Profiling Model Weaknesses 

Rigorous evaluation is central to reliably deploying AI in medical settings. This talk will discuss two recent projects on evaluating AI datasets and models. First, on evaluating data: Koh will introduce a reidentification attack for sanitized text data and show that state-of-the-art methods for data sanitization convey a false sense of privacy. Second, on evaluating models: Koh will describe EvalTree, an automated method for profiling model weaknesses to precisely identify where it fails and provide actionable guidance for improvement.

09:30 a.m. - 09:55 a.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Adapting and using language models for medical information retrieval

Language models have become a powerful tool for research in clinical oncology. However, there are many variables in how to best adapt and utilize these tools for information retrieval and utilization remains. This talk will cover experiments on the use and application of both proprietary and open language models in clinical oncology information retrieval, as well as an example of how they can be useful.

10:10 a.m. - 10:35 a.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Artificial intelligence in cancer care and clinical research

This talk will review the rapid evolution in AI technology over the last decade and summarize how it can be applied to routinely generated clinical data for patients with cancer to drive discovery and expand access to clinical trials.

10:35 a.m. - 11:00 a.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Large language models and AI to advance cancer phenotype extraction and cardiotoxicity prediction

This talk will introduce cancer-domain large language models developed to extract cancer phenotypes and generate diagnosis from electronic health records. This advancement holds significant implications for predicting cardiotoxicity related to cancer treatment. 

11:00 a.m. - 11:25 a.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Moderated by: Elizabeth Krakow, MD, CM, MS

Featuring all speakers from morning sessions + audience

11:25 a.m. - 12:05 p.m. 

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

12:05 a.m. - 12:45 p.m.

Veridical data science towards trustworthy AI

Data Science is central to AI and has driven most of recent advances in biomedicine and beyond. Human judgment calls are ubiquitous at every step of a data science life cycle: problem formulation, data cleaning, EDA, modeling, and reporting. Such judgment calls are often responsible for the "dangers" of AI by creating a universe of hidden uncertainties well beyond sample-to-sample uncertainty. To mitigate these dangers, veridical (truthful) data science is introduced based on three principles: Predictability, Computability and Stability (PCS). The PCS framework and documentation unify, streamline, and expand on the ideas and best practices of statistics and machine learning.

PCS will be showcased through collaborative research in finding genetic drivers of a heart disease and improving uncertainty quantification in supervised learning over conformal prediction. 

12:45 p.m. - 1:30 p.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Refreshments Available

1:45 p.m. - 1:55 p.m.

Weintraub Building Great Hall
Fred Hutch Cancer Center

Towards modeling everything for personalized cancer care

Personalized cancer care means delivering the right intervention to each patient at the right time, balancing potential benefits against harms.  Using cancer screening as a case study, Dr. Adam Yala presents work to advance this Pareto frontier across three dimensions: (1) predicting patient outcomes from rich clinical data, (2) designing risk-tailored intervention strategies, and (3) evaluating and translating these strategies into practice.

01:40 p.m. - 02:05 p.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Multimodal analysis as a frontier of computational oncology

In this talk, Dr. Sohrab Shah will discuss how multimodal data integration is advancing computational oncology research at different scales. Dr. Shah will show progress in integrating data from patient diagnostic information for improved risk prediction models and at the level of single cell data for improved understanding of tumor-immune interactions and spatial biology in cancer. In particular,  this talk will focus on i)  'late fusion' models for H&E whole slide image + text integration for predicting risk of recurrence in breast cancer in the real world data setting; and ii) a new graph-based neural network method for encoding spatial measurements for spatial transcriptomic and multiplexed immunofluorescence data.

2:05 p.m. - 2:30 p.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center

Charting a path to AI-augmented clinical oncology

In this talk, Dr. Robet Grant will explore how AI will improve cancer care, using examples from his research program focused on prediction models of treatment-related toxicities.

2:30 p.m. - 2:55 p.m.

Pelton Auditorium, Weintraub Building
Fred Hutch Cancer Center