RESUMEN
When integrating AI tools in healthcare settings, complex interactions between technologies and primary users are not always fully understood or visible. This deficient and ambiguous understanding hampers attempts by healthcare organizations to adopt AI/ML, and it also creates new challenges for researchers to identify opportunities for simplifying adoption and developing best practices for the use of AI-based solutions. Our study fills this gap by documenting the process of designing, building, and maintaining an AI solution called SepsisWatch at Duke University Health System. We conducted 20 interviews with the team of engineers and scientists that led the multi-year effort to build the tool, integrate it into practice, and maintain the solution. This "Algorithm Journey Map" enumerates all social and technical activities throughout the AI solution's procurement, development, integration, and full lifecycle management. In addition to mapping the "who?" and "what?" of the adoption of the AI tool, we also show several 'lessons learned' throughout the algorithm journey maps including modeling assumptions, stakeholder inclusion, and organizational structure. In doing so, we identify generalizable insights about how to recognize and navigate barriers to AI/ML adoption in healthcare settings. We expect that this effort will further the development of best practices for operationalizing and sustaining ethical principles-in algorithmic systems.
RESUMEN
Machine learning has been suggested as a means of identifying individuals at greatest risk for hospital readmission, including psychiatric readmission. We sought to compare the performance of predictive models that use interpretable representations derived via topic modeling to the performance of human experts and nonexperts. We examined all 5076 admissions to a general psychiatry inpatient unit between 2009 and 2016 using electronic health records. We developed multiple models to predict 180-day readmission for these admissions based on features derived from narrative discharge summaries, augmented by baseline sociodemographic and clinical features. We developed models using a training set comprising 70% of the cohort and evaluated on the remaining 30%. Baseline models using demographic features for prediction achieved an area under the curve (AUC) of 0.675 [95% CI 0.674-0.676] on an independent testing set, while language-based models also incorporating bag-of-words features, discharge summaries topics identified by Latent Dirichlet allocation (LDA), and prior psychiatric admissions achieved AUC of 0.726 [95% CI 0.725-0.727]. To characterize the difficulty of the task, we also compared the performance of these classifiers to both expert and nonexpert human raters, with and without feedback, on a subset of 75 test cases. These models outperformed humans on average, including predictions by experienced psychiatrists. Typical note tokens or topics associated with readmission risk were related to pregnancy/postpartum state, family relationships, and psychosis.