Use of sentiment analysis in SMS and social media to understand HIV prevention needs among young women in Kenya

Massachusetts General Hospital, Boston MA

Investigators

Abstract

Project Summary. The impact of effective HIV prevention tools depends on their use by people when they are at risk for HIV; however, many individuals are uncertain about their HIV prevention needs. Currently used risk scores and criteria (e.g., the VOICE HIV risk score for young women in sub-Saharan Africa) have been developed and validated in various populations, only to perform poorly when implemented more broadly. Novel approaches to assessment of HIV prevention needs are needed. Sentiment analysis involves natural language processing, text analysis, and computational linguistics to systematically identify, extract, quantify, and study affective states and subjective information. This technique has been successfully applied to understand various types of human behavior but has not previously been applied to HIV risk in Africa. This R21 application proposes to develop a sentiment analysis tool to assess HIV prevention needs based on SMS and/or social media messages among young women in Kenya. Importantly, participants in the recently completed Monitoring PrEP Among Young Adult women (MPYA) trial in Kenya found providing WhatsApp messages for sentiment analysis to be acceptable, and preliminary data are promising for identification of HIV risk. Because a smart phone app can process SMS and/or social media data for sentiment analysis in 10 minutes, this type of tool has great potential for use in routine care. Key next step questions include the type and amount of social media data needed for optimal correlation with HIV status and the best means to ethically and practically integrate this tool in routine clinics. We hypothesize sentiment analysis will provide novel insights into HIV prevention needs that may ultimately be shown to predict HIV incidence. We propose the following aims: 1. Explore ethical factors that may influence sentiment analysis of SMS and/or social media messages through individual qualitative interviews with 32 young women (16 who would and 16 who would not provide SMS/social media data) and a focus group discussion among five Kenyan bioethicists. 2. Conduct topic modeling and network analysis of SMS and/or social media messages (e.g., WhatsApp, Instagram, Twitter) to predict HIV prevention needs among 400 young women (age 18-24) seeking HIV testing, PrEP, and other sexual and reproductive health services in Kisumu, Kenya. We will use automated structural topic modelling to determine “topics” (i.e., word clusters) and assess for association with other risk assessments (e.g., VOICE score) as the primary outcome and HIV test results as an exploratory outcome. Data collection and analysis will conform to Aim 1 findings. 3. Assess practical factors that may influence use of a sentiment analysis tool in routine care through a needs assessment with focus group discussions of 20 clinical staff and 10 young women from Aim 2. This study will yield an optimized sentiment analysis tool for assessing HIV prevention needs that is tailored for clinic implementation and ready for large-scale testing on HIV incidence in a future R01-funded trial.