A Continual Pre-training Approach to Tele-Triaging Pregnant Women in Kenya

Access to high-quality maternal health care services is limited in Kenya, which resulted ∼36,000 and neonatal deaths 2018. To tackle this challenge, Jacaranda Health (a non-profit organization working on Kenya) developed PROMPTS, an SMS based tele-triage system for pregnant puerperal women, has more than 350,000 active users Kenya. PROMPTS empowers women living far away from doctors hospitals send messages get quick answers (through human helpdesk agents) questions about their medical symptoms pregnancy status. Unfortunately, ∼1.1 million are received by every month, makes it challenging agents ensure that these can be interpreted correctly evaluated level of emergency timely responses and/or treatments need. This paper reports a collaborative effort with develop state-of-the-art natural language processing (NLP) framework, TRIM-AI (TRIage Mothers using AI), automatically predict the (or severity condition) mother content messages. leverages recent advances multi-lingual pre-training continual code-mixed (between English Swahili), achieves weighted F1 score 0.774 real-world datasets. been successfully deployed field since June 2022, being used prioritize provision most critical conditions. Our preliminary A/B tests show ∼17% accurate at predicting high-risk conditions sent Kenyan mothers, reduces helpdesk’s workload ∼12%.