High recurrence of symptoms and common diseases, coupled with a dearth of skilled professionals, has led to the consideration of LLM-enabled medical chatbots as viable means to expand access to timely, quality care in high volume settings.
My research in LLM-enabled healthcare delivery has explored:
What can be done to shift preferences away from chatbots that overprescribe medication. As efforts are made to align chatbot responses with internationally recognized medical guidelines, a gap may arise between clinically-validated advice and patients' treatment expectations that are formed based on diverse local medical practices. Our two-phase experiment compared transcripts of a simulated chatbot offering clinically validated advice with another that overprescribed medication. In Phase 1, we found that 54% of participants showed a preference for the overprescribing bot. In Phase 2, we incorporated context-aware nudges, demonstrating a significant shift in preferences towards the clinically-validated bot. Our mixed-methods findings and analysis contribute to the ongoing dialogue on designing medical technologies that navigate the tension between clinical accuracy and patient acceptance in marginalized communities. This work is under submission.
How well an LLM-enabled, voice-based EHR system can serve doctors in rural areas. Voice-based EHR systems are increasingly being used in high resource settings to ambiently convert doctor-patient conversations into EHRs. However, the manner in which doctors converse with patients from low socioeconomic statuses may limit the ability of LLMs to perform this ambient conversion. Further, doctors’ ability to take time away from their tasks to edit generated EHRs is limited due to the extremely high volume of patients and the lack of computing facilities within rural clinics. This work explores the creation and piloting of a mobile, voice based EHR system to doctors that treat patients of different socio-economic strata, to understand the new potential that is enabled by voice-based EHRs as well as the gaps that are left after the introduction of such a system. This work is in progress.
How well a range of LLMs can handle health questions in Indic languages or Indic English. Increasingly, medical chatbots are being deployed to serve populations that speak a diverse array of languages. In this work, we interrogate their ability to process and disseminate complex health information to Indic language speakers, paying particular attention to where and how they fail. This work is in progress.