By Mark Sendak

Machine Learning (ML) and Augmented Intelligence (AI) crossed the chasm at Duke Health in 2019. The technologies and clinical integrations are now mainstream with great expectations to improve care delivery and outcomes. First was the launch of Sepsis Watch™ on November 5, 2018, after the DIHI team spent two and a half years developing and validating a deep learning model and building infrastructure to support real-time model integrations. This milestone marked the first time a deep learning technology was integrated into routine clinical care in the United States. The six-month pilot brought Duke University Hospital to the top decile in performance for the Centers for Medicare and Medicaid Services sepsis measure. Amidst the pilot success, five ML/AI projects were selected for funding through the DIHI RFA, including projects led by multiple clinical stakeholders involved in the Sepsis Watch™ program. In June 2019, the team at DIHI, in partnership with the emergency departments at Duke Regional Hospital and Duke Raleigh Hospital, implemented Sepsis Watch™. In parallel, DIHI integrated two new ML/AI models for predicting steroid-induced hyperglycemia and inpatient mortality into clinical care.

The ecosystem at Duke Health and Duke University is coalescing around the opportunity to lead the nation in developing and integrating ML/AI into clinical care. Ai.health, announced in June 2019, will harness talent, energy, and resources to scale high impact collaborations to improve health care. Central to these efforts will be DIHI infrastructure, including a data pipeline, metadata curation engine, and model deployment platform that together enable rapid development, evaluation, and integration of ML/AI into clinical care. The work of health care data janitors and data engineers will now be supported with essential tools and utilities. Beyond technology, DIHI has been leading the development of best practices to ensure the effective and rigorous development of health care ML/AI. Examples include the refinement of data quality assurance processes to ensure health care data are fit for use for ML/AI, transparent “Model Facts” labels to accompany ML/AI technologies, and standard processes for temporal and external validation ML/AI integrated into clinical care. DIHI contributed to the forthcoming guidelines set forth by the Machine Learning in Health Care (MLHC) community and will be hosting the MLHC conference in August 2020.

Central to DIHI’s mission is training the workforce to guide healthcare into the 21st century. For the first time, DIHI led training programs concurrently for undergraduate students, graduate students, and medical trainees. Undergraduate students participated in a course taught by DIHI and the Social Science Research Institute to learn methods for evaluating health care innovations. Masters students participated in a course taught by DIHI and Duke Biomedical Engineering to learn about health care data science and gained hands-on experience building ML/AI models on electronic health record data. Medical students participated in the DIHI Clinical Research & Innovation Scholarship to work on interdisciplinary teams building next-generation technologies. For the first time, all five medical student scholars are co-authors on oral presentations that will be presented at MLHC 2019.

2018 to 2019 was a year filled with building and integrating products, building capabilities and capacity, and breaking down barriers to transform health and healthcare. We’re thrilled to see what we can do in 2020 and look forward to seeing you in Durham for MLHC 2020.