MIT projects explore machine learning applications to improve EHRs
Two new studies from MIT's Computer Science and Artificial Intelligence Laboratory shed light on ways machine learning can improve electronic health records and predictive analytics to help physicians make more informed decisions.
As doctors grapple with a profusion data across multiple systems, with charts documented in varying degrees of consistency, the challenges of putting it all to use for real-time decision-making is acute.
Teams at CSAIL have tackled a pair of projects they say could help make EHRs work better for hospital clinicians. Both models were made possible by MIMIC, an open dataset developed by the MIT Lab for Computational Physiology that has deidentified health data for 40,000 critical care patients.
One project uses machine-learning for an approach called "ICU Intervene," which processes troves of data from the intensive-care-unit and applies deep learning processes to sift through lab results, vitals demographic information and more to help physicians make real-time predictions.
"The system could potentially be an aid for doctors in the ICU, which is a high-stress, high-demand environment," said MIT PhD student Harini Suresh, the paper's lead author. "The goal is to leverage data from medical records to improve health care and predict actionable interventions."
ICU Intervene offers hourly predictions of five different interventions that cover a wide variety of critical care needs, such as breathing assistance, improving cardiovascular function, lowering blood pressure, and fluid therapy, according to the report. The data are represented with values that indicate how far off a patient is from the average.
"Much of the previous work in clinical decision-making has focused on outcomes such as mortality, while this work predicts actionable treatments," said Suresh. "In addition, the system is able to use a single model to predict many outcomes."
Going forward, MIT researchers plan to improve ICU Intervene to offer more individualized care and give more advanced reasoning for its decisionmaking.
A second approach, called "EHR Model Transfer" looks to facilitate the deployment of predictive models across different platforms. Researchers showed that such models can be "trained" on one EHR system and used to make predictions in another.
Most existing machine-learning models need data to be encoded in a consistent way, researchers point out; the fact that hospitals often switch EHR systems can mean problems for predictive analytics. The EHR Model Transfer approach uses natural language processing technology to identify clinical concepts that are encoded differently across systems and then mapping them to a common set of clinical concepts, enabling analytics to works across different versions of EHR platforms.
"Machine-learning models in health care often suffer from low external validity, and poor portability across sites,” said Shah. “The authors devise a nifty strategy for using prior knowledge in medical ontologies to derive a shared representation across two sites that allows models trained at one site to perform well at another site. I am excited to see such creative use of codified medical knowledge in improving portability of predictive models."