Adaptive Content Tuning of Social Network Digital Health Interventions Using Control Systems Engineering for Precision Public Health: Cluster Randomized Controlled Trial (Preprint)

Abstract

BACKGROUND Social media has emerged as an effective tool to mitigate preventable and costly health issues with social network interventions (SNIs) but a precision public health approach is still lacking to improve health equity and account for population disparities. OBJECTIVE This study aimed (i) to develop an SNI framework for precision public health using control systems engineering to improve the delivery of digital educational interventions for health behavior change and (ii) to validate the SNI framework to increase organ donation awareness in California taking into account underlying population disparities. METHODS This study developed and tested an SNI framework which uses publicly available data at the ZCTA level to uncover demographic environments using clustering analysis which is then used to guide digital health intervention using the Meta business platform. The SNI delivered five tailored organ donation related educational contents through Facebook to four distinct demographic environments uncovered in California with and without an Adaptive Content Tuning (ACT) mechanism, a novel application of the Proportional Integral Derivative (PID) method, in a cluster randomized trial (CRT) over a 3-month period. The daily number of impressions (ie, exposure to educational content) and clicks (ie, engagement) were measured as a surrogate marker of awareness. A stratified analysis per demographic environment was conducted. RESULTS Four main clusters with distinctive sociodemographic characteristics were identified for the State of California. The ACT mechanism significantly increased the overall click rate per 1000 impressions (beta=.2187;P<.001), with the highest effect on Cluster 1 (beta=.3683; P<.001) and the lowest effect on Cluster 4 (beta=.0936.;P=0.053). Cluster 1 is mainly composed of a population that is more likely to be rural, white, and have a higher rate of Medicare beneficiaries while Cluster 4 was more likely to be urban, Hispanic, and African-American, with high employment rate without high income and a higher proportion of Medicaid beneficiaries. CONCLUSIONS The proposed SNI framework, with its adaptive content tuning mechanism, learns and delivers, in real-time, for each distinct subpopulation, the most tailored educational content and establishes a new standard for precision public health to design novel health interventions with the use of social media, automation, and machine learning in a form that is efficient and equitable. CLINICALTRIAL This study was approved by the Institutional Review Board (IRB) Office of University of California, Davis, US (1596733-2). The study was registered on ClinicalTrials.gov (NTC04850287).

Adaptive Content Tuning of Social Network Digital Health Interventions Using Control Systems Engineering for Precision Public Health: Cluster Randomized Controlled Trial

“…Organ donor registration counts obtained from California's Donor Registry (Donate Life California) were used as an external validation for the clustering analysis. The acquired demographic data were used as an input to a machine learning methodology [33] to stratify the target population into groups of ZCTAs with similar demographics, considering a diverse set of possible clustering retrievals. The methodology consists of a clustering analysis that automatically selects the most suitable number and composition of groups using a range of clustering algorithms and metrics (Figure 1).…”

“…In its first step, a set of clustering retrievals for each k is obtained for each algorithm. Five clustering metrics were then simultaneously considered: Silhouette [37], CH-index [38], DB-index [39], WB-index [40], and Infoguide [33] to determine the candidate set of clustering retrievals. Finally, the chosen clustering retrieval was obtained considering its goodness of fit when used in a prediction model in which the outcome was the organ donor registration count.…”

“…Finally, the chosen clustering retrieval was obtained considering its goodness of fit when used in a prediction model in which the outcome was the organ donor registration count. Refer to the clustering methodology for more detail [33].…”

Association of Neighborhood and Environmental Factors With Clinical Phenotypes and Outcomes in Heart Failure With Preserved Ejection Fraction

“…Our community clustering algorithm 4,5 used 24 sociodemographic and socioeconomic variables previously described, 5 across 5 main domains of SDoH, at ZIP code tabulation areas (ZCTA) level (from the American Community Survey, https://data.census.gov/cedsci/ [accessed 7 March, 2022]). Prominent variables were education level, access to health care, household income, and employment.…”