SmartHIVCare: A Bilingual Retrieval-Augmented Multimodal Chatbot for ART Education and Adherence Support in Low-Resource Settings

Abstract

Background: Human immunodeficiency virus (HIV) remains a major public health challenge in sub-Saharan Africa, where linguistic diversity and limited digital health resources constrain patient education and antiretroviral therapy (ART) adherence. Amharic, one of Africa’s most widely spoken languages, remains underrepresented in clinically oriented conversational AI systems. Methods: We developed SmartHIVCare, a bilingual (Amharic–English) multimodal conversational system integrating retrieval-augmented generation, multilingual semantic retrieval, Amharic-specific text normalization, and speech-based interaction for ART education and adherence support. Performance was evaluated using automated response quality metrics, clinician-based human evaluation, latency analysis, error analysis, and an ablation analysis on 50 bilingual queries (25 English, 25 Amharic). Results: SmartHIVCare achieved BLEU scores of 0.448 ± 0.386 (95% CI: 0.341–0.555) and a BERTScore of 0.820 ± 0.114 (95% CI: 0.789–0.852). Human evaluation yielded high ratings for accuracy (4.8/5, 4.6/5), clarity (4.6/5, 4.6/5), safety (4.6/5, 4.6/5), and clinical usefulness (4.8/5, 4.4/5) for English and Amharic, respectively. Response latency scaled with modality: 3.2 s (text-to-text), 4.1 s (text-to-speech), 5.8 s (speech-to-text), and 6.4 s (speech-to-speech). Conclusions: SmartHIVCare demonstrates the feasibility of retrieval-grounded bilingual conversational AI for HIV education in underrepresented languages. This proof-of-concept evaluation focuses on technical feasibility and response quality and requires validation through larger real-world clinical studies.