Description of Session
Jacaranda Health’s two-way text messaging platform (PROMPTS) educates and empowers mothers in Kenya to seek care at the right time. It is rapidly scaling, with 40,000 users projected by the end of 2019. A feature of the platform that has high demand is the ability for mothers to have their questions answered by helpdesk agents. The helpdesk has responded to >30,000 unique questions to date. Although many of the questions are general in nature and non-urgent, almost thirty percent of incoming messages indicate the need for further and often urgent medical attention – such as bleeding or sepsis. We asked: How can these red-flag questions or comments be identified quickly, while maintaining the overall efficiency and quality of the helpdesk? In this session, we will share our approach to using Natural Language Processing (NLP) on the rich conversational data contained in the platform to rapidly identify mothers who need additional care. Our goal is a response time of less than 1 hour for these urgent questions. We designed and tested an NLP bot, with an integrated Swahili translation component, that reads each incoming question and assigns a priority level and provides a suggested response to each message. The priority levels and suggested responses are used by the helpdesk agents to respond more quickly to high priority messages (e.g. bleeding). Preliminary assessments indicate that integration of this bot has resulted in a 50% reduction in response time to urgent questions. We will discuss how this approach leverages a powerful source material for Machine Learning: a well-characterized corpus of data of individuals and their health outcomes that can be used to better understand complex conversations and intents.
