Testing a Novel Design Framework for Patient-Facing Machine Learning-Based Predictions of Heart Failure Decompensation.

Overview

abstract

BACKGROUND: Policies facilitating the return of personal health data mean there is an urgent need to investigate strategies to safely present machine learning-based predictions to patients. OBJECTIVES: The authors aimed to design and test the usability of a patient-facing smartphone application prototype displaying a predictive algorithm of cardiac decompensation to patients with cardiac implantable electronic devices (CIEDs). METHODS: We created a design framework for presenting algorithm output and implemented it in high-fidelity prototypes of a smartphone app. Prototypes showed 3 conditions with varying degrees of change in cardiac decompensation risk: significant, moderate, or little to no change. We conducted a mixed-methods usability evaluation of the prototype at a large, urban health system. English-speaking adults with implanted, actively transmitting CIEDs participated in evaluation sessions. The primary endpoint was patient objective comprehension of the algorithm's output. Secondary endpoints were risk perception, behavioral intention, and information-seeking. RESULTS: Twenty participants (mean age 54 years, 40% female, 65% White, and 35% Black or African American) completed the study. Comprehension of the algorithm output was high across conditions (80% to 85%), but comprehension of the algorithm's threshold varied (60% to 85%), as did comprehension of the CIED sensors contributing to the algorithm (63% to 93%). In response to the condition showing a significant change, the majority of participants would be "moderately" (40%) or "very" worried (30%) about worsening cardiac status, and 80% would call their doctor. CONCLUSIONS: The prototype demonstrated high levels of patient comprehension, appropriate risk perception, and behavioral intention, suggesting its viability for empowering patients with actionable health insights.