CROI 2025 Abstract eBook

Abstract eBook

Poster Abstracts

1309 Developing Machine Learning Algorithms to Predict Treatment Interruptions in HIV Care in Uganda

applicability in other settings and its clinical utility compared to existing data to-care initiatives.

Alex Mirugwe 1 , Solomon Ssevvume 1 , Edward Bichetero 1 , Alice Namale 1 , Evelyn Akello 1 , Paul Katongole 1 , Rose Baryamutuma 1 , Paul Mbaka 2 , Enos Sande 3 , Arthur Fitzmaurice 4 , Kenneth Musenge 3 1 Makerere University–University of California San Francisco Research Collaboration, Kampala, Uganda, 2 Ministry of Health Uganda, Kampala, Uganda, 3 US Centers for Disease Control and Prevention Kampala, Kampala, Uganda, 4 US Centers for Disease Control and Prevention Dar es Salaam, Dar es Salaam, United Republic of Tanzania Background: Achieving consistent retention of people living with HIV (PLHIV) in care remains a challenge in Uganda, despite substantial progress towards UNAIDS 95-95-95 targets. This study used advanced machine learning and deep learning techniques applied to deidentified longitudinal treatment and demographic data routinely collected in HIV clinics to predict clients at greatest risk of missing treatment appointments. Methods: This retrospective study used a longitudinal dataset of 66,206 PLHIV who initiated HIV care during 2000-2023, with 1,479,121 clinical visits from 86 health facilities in Uganda, extracted from the central case-based surveillance database. We trained and compared the performance of traditional machine learning models (Decision Tree, Random Forest, AdaBoost, and XGBoost) and the Bidirectional Encoder Representations from Transformers (BERT), model, which is more suitable for analyzing sequential data (e.g., medical records). Feature importance was analyzed using the shapely additive explanations method to identify the most influential predictors. We also investigated the impact of various sampling techniques, including undersampling, oversampling, and synthetic minority oversampling to address class imbalance and improve model performance. Model performance was evaluated using accuracy, precision, recall, F1-score, and AUC-ROC metrics. Results: The dataset comprised 43,132 (65.1%) females with a median age of 36.0 years and 23,074 (34.9%) males with a median age of 41.0 years. Each client had an average of 22.34 clinical visits, with 10.7% of appointments missed and 74.9% missing at least one appointment. The BERT model demonstrated superior performance, achieving an AUC score of 0.96, 94.8% accuracy, 97.1% precision, 100% recall, and an F1-score of 94.2%. In comparison, the XGBoost model with undersampling achieved an AUC score of 0.90, 80.7% accuracy, 97.1% precision, 80.8% recall, and an F1-score of 88.2%. Feature importance analysis showed that historical treatment adherence, visit frequency, treatment duration, and visits on the current regimen are the most influential predictors of appointment interruption. Conclusions: This study highlights the efficacy of transformer-based models like BERT in handling sequential clinical data and improving patient retention predictions. Integrating these predictive models into electronic medical systems could facilitate proactive treatment strategies, enabling the identification of clients at risk of missing their appointments.

The figure, table, or graphic for this abstract has been removed. 1308 An AI-Powered Preventive Intervention for Stigma and Suicidal Ideation in HIV Self-Management Diego S. Villanueva Guzman 1 , Yuanchao Ma 1 , Sofiane Achiche 2 , Kim Engler 1 , David Lessard 1 , Bertrand Lebouche 1 1 Research Institute of McGill University Health Centre, Montreal, Canada, 2 Polytechnique Montréal, Montreal, QC, Canada Background: Self-harm and suicidal ideation are significant health concerns among people with or vulnerable to HIV (PWH), who frequently experience stigma and related mental health challenges such as stress, anxiety, and depression. The AI-powered MARVIN chatbot provides HIV self-management related knowledge and assists with medication adherence (antiretroviral therapy and pre-exposure prophylaxis). It also provides a safe space for PWH to ask sensitive questions which is amenable to mental health support. To enable MARVIN to manage high-risk messaging linked to stigma and mental health issues, we developed a preventive intervention module to address extreme user intentions. Methods: Following the CO-STAR framework, ChatGPT was prompt-tuned to identify 3 types of message intent: self-harm, insult, and non-extreme (i.e., any other intent). To test its performance, we compiled three public hate speech databases from an online catalog (hatespeechdata.com) and combined them with MARVIN-user conversations and a synthetic dataset (N=1000 for each class). We computed precision, recall, and F1 Score for each class, as well as overall accuracy. After integrating ChatGPT into MARVIN, three PWH, two engineers, and a doctor, participated in a two-hour test by performing 14 conversational scenarios and completing a two-item questionnaire on conversation clarity and user satisfaction. Results: With one-shot prompting, ChatGPT attained 97.00% and 94.90% for recall on self-harm and insult intent, respectively. The overall accuracy reached 95.57%, with the remaining metrics shown in Figure 1. The ChatGPT-MARVIN hybrid model then successfully generated appropriate responses containing 1) emergency contact information for self-harm intents; 2) messages guiding users to use stigma-free expressions for insult messages; and 3) a response reviewed by a medical expert for a non-extreme intent. All six testers found MARVIN’s responses to be clear and concise and were satisfied with the overall experience. However, one patient participant suggested including links to additional resources. Conclusions: Testing this anti-stigma preventive module integrated into MARVIN underscored its strong capacity to correctly identify extreme intents and generate concise, satisfying responses. Our findings suggest that AI-driven chatbots, viewed as safe spaces, could play a pivotal role in promoting self management strategies while mitigating stigma and providing users with non-judgmental mental health support.

Poster Abstracts

1310 Exposure and Engagement Drive Impact: Results From a Digital Health Trial With Rwandan Adolescents Rebecca Hemono 1 , Lauren Hunter 1 , Emmyson Gatare 2 , Laetitia Kayitesi 2 , Therese Bagwaneza 2 , Raissa Umutoni 2 , Natacha Mugeni 2 , Stefano Bertozzi 1 , Rebecca Hope 3 , Sandra I. McCoy 1 1 University of California Berkeley, Berkeley, CA, USA, 2 YLabs Rwanda, Kigali, Rwanda, 3 YLabs Global, San Francisco, CA, USA Background: Digital health interventions are a promising approach for increasing access to health information and services, but low or non-sustained engagement can limit effectiveness. We examined how exposure to and

CROI 2025 435