Medical Learning Through Machine Learning (ML2)

Objectives and Rationale: Emergency care medicine both on the battlefield and at trauma centers faces two related concerns: What are new patterns of injury being seen by first response providers, and are the providers adequately trained to handle these evolving situations? In this project, we will focus on gaps in medical care that can be identified, analyzed, predicted, and turned into tools for improvements in trauma skills training of medical personnel operating in the emergency room (ER) setting and/or in the field. We expect that our research results will inform competency assessment for both military and civilian healthcare systems. The objective of this project -- Medical Learning through Machine Learning ([ML]2) -- is to apply machine learning (ML) and artificial intelligence (AI) techniques to real hospital and trauma care data, in an attempt to understand factors influencing patient outcomes, and translate our findings to medical case studies and better training materials.

We hypothesize that our research will enable us to infer the level of capability needed by a healthcare provider for that provider to come to an appropriate decision and perform expected interventions. We further hypothesize that aggregating these analyses across many providers intervening with many patients will lead to the identification of learning gaps where additional or enhanced training can lead directly to improved patient outcome. To test these hypotheses, we propose the following objectives:

Objective 1: Data collection, integration and curation. For this proposal, three organizations will serve as our primary verified data sources. The Carolina Data Warehouse for Health (CDW-H) is a data repository for the 10 hospitals in the University of North Carolina Health Care System containing data on patient demographics, encounters, diagnoses, procedures, medications, labs, full-text notes and financials. The North Carolina Trauma Registry (NCTR) houses data that is collected on all trauma patients from numerous North Carolina hospitals, including all trauma centers, to facilitate trauma system development. The Department of Defense Joint Trauma Registry (DoDTR) is a data repository for DoD trauma-related injuries, many reported from battlefield experiences.

Objective 2: Develop models for curated data to enable competency assessment. We will develop predictive models using ML and AI tools. We will explore clinical data to forecast injury progression depending on condition and medications used. We will also identify gaps in training responsible for such needs, which will direct future treatment guidelines and curriculum development.

Objective 3: Integrate all available data and modeling tools within the developed platform. We will design and implement [ML]2 as a comprehensive online data management system with the ability to import data from numerous sources (such as hospital records), clean up, edit, curate, and semantically enrich data, auto-detect errors, and integrate AI and ML tools.

Applicability and Impact: Two major concerns will be addressed in this research. First, too many providers fail to treat patients according to established clinical practice guidelines (CPGs). CPGs are a distillation of the medical literature and experience. There are differences between civilian and military CPGs because of the different type of casualties and available medications, tools, and treatments, but there are also many similarities (and military and civilian practices inform each other). Adherence to CPGs is poor under many emergency care situations. Once identified, adverse treatment informs training gaps, especially those that need to be addressed immediately. Machine learning of CDW-H, NCTR, or DoDTR data can not only suggest that patient outcomes are less desirable than expected, but also give a clue as to why -- the data show what trauma care event occurred, when it happened, and who performed it. We will scour that data to determine training of medical providers that is missing that could result in a higher survival rate. Second, we expect to pick up on important patterns of injury much sooner than is possible via human observation. Today, there is little in the way of feedback from statistics in these databases to inform practice in the field. Once we identify anomalies, gaps, outliers, or trends, we can relate them back to training.