Xiang Jang, a PhD candidate in Systems Engineering at the University of Virginia, presents his research on applying Bayesian risk analysis to large multimodal healthcare data, primarily focusing on COVID-19 related projects. The talk introduces the concept and importance of large multimodal data in the era of AI, highlighting its role in integrating diverse information sources for complex domains like healthcare. Jang then details three specific research contributions: modeling COVID-19 transmission dynamics using wastewater data and a novel SEIR (Susceptible, Exposed, Infected, Recovered) model, analyzing the mortality risk of long COVID patients through Bayesian generalized linear regression, and investigating healthcare inequality in COVID-19 hospitalization rates using a Bayesian hierarchical model across various U.S. sites. His findings demonstrate the utility of Bayesian methods in understanding disease characteristics, predicting outcomes, and identifying disparities within large, complex datasets to promote equitable healthcare decisions.
Summarizing the Transcript: Bayesian Risk Analysis on Large Multimodal Healthcare Data
0:00 Introduction of Speaker and Topic: Michael Hoy introduces Xiang Jang, a PhD candidate at UVA Systems Engineering, whose research focuses on Bayesian machine learning, Markov models, and AI for health. The talk will cover his work with the UVA IT Thrive organization on the National COVID Cohort Collaboration (N3C) project.
1:26 Overview of the Presentation: The talk is structured in two main parts: an introduction to large multimodal data and its significance, followed by an exploration of Bayesian risk analysis and its application to various research projects in healthcare.
1:48 Understanding Large Multimodal Data: Multimodal data refers to vast and complex datasets that integrate multiple types of information, such as text, images, audio, video, and numerical data from various sources. This integration is crucial in the era of AI and large language models, especially in complex domains like healthcare, where single-modality data often fails to capture the full picture.
2:20 Multimodal Data's Role in COVID-19: The COVID-19 pandemic significantly accelerated the generation and integration of large-scale multimodal data. This includes diverse information sources like lab tests, imaging, pathology, symptoms, and other clinical or physiological data, which are vital for understanding disease transmission and characteristics.
3:54 Multimodal vs. Unimodal AI Models: Multimodal AI models integrate diverse inputs (e.g., text, image, video) to address complex problems, offering a comprehensive approach. In contrast, unimodal AI models are limited to a single type of data input.
4:41 Healthcare Applications of Multimodal Data: When applied to artificial intelligence and machine learning models, multimodal data can assist with various healthcare tasks, including disease detection and screening, diagnosis, treatment response prediction, and drug development.
5:37 Global Multimodal Healthcare Databases: Several large-scale multimodal databases exist globally, many of which expanded significantly during the COVID-19 pandemic. Examples include the European Health Data and Evidence Network, UK Biobank, Germany's National Pandemic Cohort Network, and Japan's National Database, which contain extensive healthcare and insurance claim information.
6:53 Key US Databases: N3C and IThrive: The speaker primarily focuses on the National COVID Cohort Collaborative (N3C), an NIH-backed database compiling over 22 million patient records (including 8.9 million COVID-positive) and 33 billion rows of electronic health record data from over 90 US sites. This data also includes socio-economic information like zip codes and social vulnerability indices. IThrive, an NIH-sponsored Virginia-based research institute, aims to replicate N3C methodologies for data within the state, utilizing a similar cloud computing environment.
11:57 Introduction to Bayesian Risk Analysis: Bayesian risk analysis is a decision-making framework based on Bayes' Theorem. It models uncertainty by treating parameters as random variables, allowing continuous updates of beliefs by combining prior knowledge with evidence observed from data. Unlike frequentist approaches, Bayesian methods provide a distribution for predictions, reflecting inherent uncertainty.
15:19 Project 1: COVID-19 Transmission Dynamics: Jang's research included a novel Bayesian SEIR compartmental model incorporating wastewater viral load and reported confirmed cases from New Haven, CT (March-May 2020). This approach helps estimate transmission rates, particularly in regions with insufficient testing, by modeling hidden states like exposed and infected populations, which are not directly observable.
20:51 Project 2: Characteristics of Long COVID Mortality: This project investigated the mortality risk among long COVID patients, a chronic condition defined by the CDC (ICD10 code U9.99). Using a matched dataset of 47,000 long COVID patients and COVID-positive controls from N3C, a Bayesian generalized linear regression model indicated that long COVID patients have a statistically distinguishable, albeit small, reduction in survival probability over time.
24:52 Project 3: COVID-19 Healthcare Inequality: Using a Bayesian hierarchical model, the research analyzed data from approximately 800,000 patients with COVID-associated emergency department (ED) visits across 54 N3C sites. The findings revealed that patients from underrepresented racial and ethnic groups were significantly less likely to be hospitalized after their ED visits, despite potentially severe conditions, highlighting existing systemic healthcare disparities.
28:25 Conclusion: Importance of Bayesian Risk Analysis: Bayesian risk analysis applied to large multimodal data is crucial for robustly capturing uncertainty, integrating diverse information, providing quantitative evidence, and facilitating personalized, equitable, and data-driven decisions within healthcare and other complex fields.
29:11 Q&A: Feature Importance in Latent Processes: In response to a question regarding feature importance in models with long latent processes and delayed outcomes (like death), Jang explained that in linear models, importance can be inferred from coefficients, while tree-based models use specific metrics. He noted that comprehensive Electronic Health Record (EHR) databases help address latency by collecting numerous observations during patient hospital visits.