Patient Experience Recommender System for Persuasive Communication Tailoring

To maximize patient perspective and effectively support lifestyle choices, we will develop the "Patient Experience Recommender System for Persuasive Communication Tailoring." PERSPeCT is an adaptive computer system that will assess a patient's individual perspective, understand the patient's preferences for health messages, and provide personalized, persuasive health communication relevant to the individual patient.

Investigators propose to overcome key weaknesses in existing top-down expert-driven health communication interventions by applying advanced machine learning algorithms to adaptively recommend messages based on the "collective intelligence" of thousands of patients. This work will leverage a paradigm-shifting "Web 2.0" approach to adaptive personalization with the potential for broad impact on the field of computer tailored health communication (CTHC).

Using knowledge from scientific experts, current CTHC interventions collect baseline patient "profiles" and then use expert-written, rule-based systems to target messages to subsets of patients. These market segmentation interventions show some promise in helping certain patients reach lifestyle goals. Although theoretically sound, rule-based systems may not account for socio-cultural concepts that have intrinsic importance to the targeted population, thus limiting their relevance. Further, the rules do not adapt to patient feedback.

Outside healthcare, companies like Google, Amazon, Netflix and Pandora have made extensive use of adaptive recommendation systems to provide content with enhanced personal relevance. These systems use machine learning algorithms to derive personalized recommendations from a variety of data sources including preference feedback collected from individual users.

Within the scope of this Patient-Centered Outcomes Research Institute (PCORI) pilot, investigators will address the challenges of adapting machine learning recommender systems to CTHC in the specific context of patient decision support for smoking cessation. Investigators have chosen this domain because smoking is a major preventable cause of death, and because we have an existing database of 1,000 persuasive messages developed in a current federal grant (R01 CA129091). Specific study aims are to:

Aim 1: Collect Explicit Feedback data in order to train PERSPeCT Recruit 700 smokers using multiple, complimentary strategies, and using a web interface, ask smokers to provide (a) Perspectives on smoking and quitting and socio-cultural context information and (b) Ratings of the influential aspect of smoking cessation messages.

Aim 2: Design, Implement and Validate a customized recommendation framework This will involve (a) developing and implementing a machine learning recommender system that integrates patient profiles, message metadata, web site views and influence ratings, and (b)training the model and validating its predictive performance.

Aim 3: Conduct a pilot randomized trial (n = 120 smokers) of PERSPeCT. Investigators hypothesize that the PERSPeCT system will (H1) Select messages of increasing influence as smokers provide more message ratings and (H2) Select messages with better influence than a rule-based CTHC system when smokers provide a sufficient number of ratings CTHC systems support patient decisions about behaviors, lifestyles, and choices. PERSPeCT addresses areas of interest for PCORI, namely: 1) Identifying, testing, and/or evaluating methods that can be used to assess the patient perspective when researching behaviors, lifestyles, and choices within the patient's control; and 2) Developing, refining, testing, and/or evaluating patient-centered approaches, including decision support tools. The study team is uniquely positioned to accomplish these ambitious aims within the scope of this PCORI pilot because investigators will utilize an existing database of persuasive messages from a previous study, two years of data on the effectiveness of these messages and a trans-disciplinary team with expertise in health communication, web systems engineering, and machine learning recommender systems.