MARIA (Medical Assistance and Rehabilitation Intelligent Agent) for Medication Adherence in Patients With Heart Failure: Empirical Results From a Wizard of Oz Systematic Conversational Agent Design Clinical Protocol

Background: Nonadherence to medication is a key factor contributing to high heart failure (HF) rehospitalization rates. A conversational agent (CA) or chatbot is a technology that can enhance medication adherence by helping patients self-manage their medication routines at home. Objective: This study outlines the conception of a design method for developing a CA to support patients in medication adherence, utilizing design thinking as the primary process for gathering requirements, prototyping, and testing. We apply this design method to the ongoing development of Medical Assistance and Rehabilitation Intelligent Agent (MARIA), a rule-based Methods: Following the design thinking process, at the ideation stage, we engaged a multidisciplinary group of stakeholders (patients and pharmacists) to elicit requirements for the early conception of MARIA. In collaboration with pharmacists, we structured MARIA's dialogue into a workflow based on Adlerian therapy, a psychoeducational theory. At the testing stage, we conducted an observational study using the Wizard of Oz (WoZ) research method to simulate the MARIA prototype with 20 patient participants. This approach validated and refined our application of Adlerian therapy in the CA's dialogue. We incorporated human-likeness and trust scoring into user satisfaction assessments after each WoZ session to evaluate MARIA's feasibility and acceptance of medication adherence. Dialogue data collected through WoZ simulations were analyzed using a coding analysis technique. defining the extent to which a CA can communicate on behalf of a health care provider regarding medication adherence. Conclusions: The design thinking process provided interactive steps to involve users early in the development of a CA. Notably, the use of WoZ in an observational clinical protocol highlighted the following: (1) coding analysis offered guidelines for modeling CA dialogue with patient safety in mind; (2) incorporating human-likeness and trust in user satisfaction assessments provided insights into attributes that foster patient trust in a CA; and (3) the application of Adlerian therapy demonstrated its effectiveness in motivating patients with HF to adhere to medication within a CA framework. In conclusion, our method is valuable for modeling and validating CA interactions with patients, assessing system reliability, user expectations, and constraints. It can guide designers