Development of Machine Learning Methods for Understanding, Detecting, and Preventing iatrogenesis
Summary
Background: Iatrogenesis is the side effect and risk associated with medical care. It includes the causation of a disease, a harmful complication, or other ill effects by any medical activity, including diagnosis, intervention, error, or negligence. The Institute of Medicine (2000) estimated that about 98,000 patient deaths due to substandard care in hospitals. A more recent projection was as high as 400,000, making it one of the leading causes of death in the US.
Patient Safety Focus: Ensuring patient safety requires a new model for healthcare institutions. We need to develop a Learning Healthcare System that learns from previous events, supports timely access to accurate clinical information and clinical decision support, and establishes the institutional knowledge management infrastructure that proactively help prevent harm.
Predicting iatrogenesis: A key component in preventing iatrogenic harm is understanding the types and causes of harm. Today, iatrogenesis is often identified retrospectively after the harm has already occurred. Voluntary reporting and manual chart review are helpful underestimates the true incidence, challenging to scale, and often requires staffing and expertise requirements. With the advent of electronic health records (EHR), information technology has the potential to identify adverse events. Our project will attempt to leverage machine learning (ML) approaches to understand, detect, predict and prevent iatrogenic events prospectively using clinical decision support systems.
Role of ML and the EHR: With the increasing availability of EHR-derived computable healthcare data, more advanced information technology applications such as ML and artificial intelligence (AI) hold great promise in augmenting the current limitations of preventing adverse events. Institutions need a better way of identifying harmful events before they happen, and proactively institute changes to reduce the probability of harm from occurring. Using EHR data, ML methods such as random forest classifiers and deep learning, clustering, use of natural language processing, among others, show potential in identifying iatrogenic harm.
Goal: Develop the Machine Learning technology to Understand, Detect, and Prevent iatrogenesis.
Methods: We will use the extensive amount of patient safety and EHR data captured by Arkansas Children’s information systems to 1) identify the properties and characteristics of patient safety events and 2) to determine the attributes and variables that contribute to specific types of adverse events. Successful models identified in the project will be integrated into care delivery via clinical decision support mechanisms. We will leverage the computing resources available at Arkansas Children’s and UAMS to support the effort, including the exploration of machine learning tools available at Amazon Healthlake (https://aws.amazon.com/healthlake/).
Task #1: Link Patient Harm dataset with EHR data: We will build a patient-level, de-identified ML training dataset that combines incident cases from the patient safety reporting system and big data (including metadata, log files, notes, clinical) from the EHR.
Task #2: ML Feature Engineering: We will perform descriptive data analyses and employ ML methods to identify the types of iatrogenic harm (dependent variables) and the related features in the EHR (independent variables). We will then utilize these features to develop a predictive model for iatrogenic harm events.
Keywords:
- machine learning
- artificial intelligence
- iatrogenesis
- adverse events
- patient safety
- electronic health record (EHR)
- clinical informatics
- clinical decision support systems
Researchers:
- Feliciano Yu (Author)
- Melody Greer
- Daniel Liu
- Jonathan Bona