Needs-driven, utility-oriented, standards-based operationalization of artificial intelligence for clinical decision support: a framework with application to suicide prevention

Abstract

While artificial intelligence (AI) technologies increasingly permeate our daily lives, the adoption and impact of AI have fallen short of expectations in healthcare. The challenges of operationalizing AI in healthcare are complex and include interaction design (e.g. poorly designed user interfaces), model formulation (e.g. algorithmic bias, limited practical utility, trustworthiness or interpretability), and workflow context (e.g. a lack of integration into existing workflows; limited actionability). Critically, AI projects must demonstrate overall utility, balancing their costs with the benefits they confer. To achieve this utility, informatics efforts are needed before, during, and after predictive model development, to mediate effective, sustainable, and interoperable AI deployment to support clinical workflows. In this work, I investigated how human-centered design methods, needs-driven model development, utility-oriented evaluation methods, and standards-based software design can be leveraged collectively to address the unique challenges faced by healthcare AI, and achieve clinically impactful AI implementations. The two key contributions resulting from it are (1) a generalizable framework for the needs-driven operationalization of AI to support healthcare workflows and clinical decision making, and (2) the application of this framework to conceive, implement and evaluate AI support for suicide prevention. To apply this framework, I used human-centered design methods to assess technological support needs for Caring Contacts, an evidence-based suicide prevention intervention, revealing opportunities for AI-based cognitive support. Using neural transfer learning from publicly available social media data, I developed accurate natural language processing models for risk-based prioritization of patient messages. Through utility-oriented evaluation metrics, I demonstrated that this model has the potential to positively impact clinical practice. Incorporating this model, I devised a standards-based, reusable, interoperable, workflow-integrated information system for cognitive support of Caring Contacts. I developed blueprints for a FHIR data representation model and information system architecture, and implemented and shared an open-source software application. Together, this work contributes towards bridging the historical implementation gap by furthering methods for the design, development, and delivery of AI-supported interventions, and by guiding future attempts to realize the potential of AI in clinical settings.