Do You Trust Me? Development, Implementation and Acceptance of a Machine Learning Model - Opportunities, Challenges and Future Direction
Video
Team Information
Team Members
Murad Megjhani, Associate Research Scientist, Department of Neurology, Columbia University Irving Medical Center
Faculty Advisor: Soojin Park, Associate Professor of Neurology, Vagelos College of Physicians and Surgeons
Abstract
There is growing Interest in developing, implementing, and deploying machine learning (ML) models in real-time across industries and in the hospital. However, there are challenges associated with real-time implementation and model acceptance. These include: a) challenges with mathematical models, e.g., lack of enough training data for the model to be generalizable, b) logistics, e.g., real-time implementations often have to deal with delays in data acquisition, that can lead to incorrect model output, technical issues related to maintenance and integration of model into the electronic health record (EHR), and finally, c) barriers to adoption that include privacy, security, risk assessment, lack of trust in the black box machine learning model and unintended consequences on clinician workflow.
Developing an Algorithm Implementation Pipeline for Applications in Intensive Care (ALPACA), we demonstrated a modular model deployment pathway using the use case of a ContinuOuS Monitoring tool for delayed cerebral IsChemia (COSMIC). A trained and validated temporal machine learning algorithm that detects delayed cerebral ischemia (DCI) after subarachnoid hemorrhage was successfully implemented, and is undergoing rigorous testing for accuracy and acceptance. Currently there are no effective monitoring tools for DCI. The pipeline ingests demographic data and real-time continuous physiological data.
We previously illustrated technical run time challenges and mitigations in deploying ML models. Real-time deployment of the Clinical Decision Support Service (CDSS) accepts data from data stores, runs the machine learning model, and provides model output for a user interface. We are evaluating accuracy in a silent clinical validation, with a pilot study targeting 48 patients (enrolled 11). We are planning a simulation study of acceptance of the risk score by trained clinicians. We have started a Contextual Design approach for user centered design of the clinical decision support, and have enrolled and interviewed 11 of 15 planned clinicians. Thematic extractions from interviews will inform the storyboards for rapid agile iterative design of the User Interface. We plan to develop COSMIC to be Fast Healthcare Interoperability Resource (FHIR) and FHIR-device interoperable, resulting in an innovative plug and play tool that will enable future effectiveness testing across centers and widespread dissemination.
Team Lead Contact
Murad Megjhani: mm5025@cumc.columbia.edu
