Designing a Comprehensive Machine Learning-driven Prescriptive Clinical Decision Support System for Shock and Associated Critical Care Conditions

University Of Pittsburgh At Pittsburgh, Pittsburgh PA

Investigators

Abstract

Title Designing a Comprehensive Machine Learning-driven Prescriptive Clinical Decision Support System for Shock and Associated Critical Care Conditions PI: Yoon, Joo Heung (University of Pittsburgh) Abstract The thought process of intensive care unit (ICU) clinicians to make everyday decision is quite complex with a lot of concurrent acute and chronic clinical conditions to consider, which requires challenging risk-stratification and prioritization steps among competing interests and confounders. Also, those decisions are high-stake and need to be timely. With the recent advancement of machine learning in medicine, various diagnostic and therapeutic models were attempted in hope of streamline those decision-making processes. However, most of machine learning (ML) models cannot be readily implemented to bedside, with less than 5% of models have gone through appropriate prospective study or clinical trials. The challenges are, i) inefficient ML model development pipeline with incomplete data, repetitive workflow with pre-processing, labeling, and featurization, all of which are extremely resource intensive, ii) the lack of overarching decision supporting engine that handles competing interests among simultaneous different decisions, and iii) inadequate post-implementation analysis with unclear safety profile of using ML models at bedside. Through this Maximizing Investigators' Research Award proposal, we will focus on resolving those critical bottlenecks in ML-driven decision support strategy, with the example of shock and associated critical condition, which are very frequently encountered among critically ill patients. First, to address the efficient and approachable ML model design, we will develop a multimodal foundation model for ICU patients. Self-supervised pre-trained models using a large amount of various clinical data, using waveform and electronic health record (EHR). Second, by using the foundation model, we will generate a series of downstream detection, prediction, and management models in a very time-efficient way. Then we plan to build a dynamic priority engine to determine next set of actions prior and during the shock period, through which the treatment priorities will be determined by the importance and urgency of each clinical problem using reinforcement learning, based on the inference generated by downstream models. Just like real-life clinical decisions are changing with time, the priority set by our engine will change as well. Third, we will silently deploy the priority engine to the data stream of deidentified ICU patients, in which setting neither clinicians nor patients will aware the model's presence as it only runs in background and the detection, prediction, or management plans will not be shared, but available for outcome analysis. After the completion of this proposal period, we ultimately plan for a scalable multi-center clinical trial on our downstream models and prioritization engine for shock and other condition management. We will make the entire models, pipeline, and results publicly available. Current proposal will set a roadmap for implementable ML-driven clinical decision support system for researchers and clinicians to utilize in critical care environment.