AI Systems for Qualitative Data Interrogation (EpiChat) and Semi-Automated Coding (EpiCode)

Abstract

Background: The Virage project collected approximately 800 one-hour interviews on rabies awareness, including transmission, disease progression, and vaccination. Manually searching, comparing, and coding this corpus would require substantial time and human resources. Objective: EpiChat and EpiCode were developed using large language models (LLMs) to support natural-language interrogation and thematic analysis of interview transcripts. Local LLMs were used to preserve data privacy, avoid sharing data with foreign cloud services, and eliminate token and subscription costs. Methods: EpiChat uses an LLM-based router to direct questions to three specialized pipelines. An SQL pipeline translates structured metadata questions into database queries. For questions regarding counts or proportions, a map-reduce pipeline examines each interview independently and aggregates the results. Explanatory questions are handled using Retrieval-Augmented Generation (RAG). Because conventional RAG retrieves only the highest-ranked passages across the corpus, it may overlook rare perspectives that are important to qualitative research. To address this, EpiChat introduces a deep search mode that formulates the question as a quotation and searches for it into each interview. Verified quotations from matching interviews are then combined into a single response with source citations. EpiCode supports qualitative coding by matching transcript segments to a two-level codebook of main codes and subcodes. Each suggested code includes a model-generated confidence score, supporting quotation, and brief rationale. When no existing code adequately represents the content, the system proposes a new code. All suggestions require researcher review and approval. Results: Preliminary runtime measurements showed that EpiChat processed an interview-level query in approximately four seconds per interview, while EpiCode generated code suggestions for an entire interview in around three minutes. Conclusion: EpiChat and EpiCode provide a privacy-preserving, locally deployed platform for AI-assisted qualitative analysis. Future work will compare AI-generated codebooks with human-developed codebooks and evaluate whether EpiChat produces answers consistent with findings derived from manual coding. Poster presented at the Pasteur Network Annual Meeting (PNAM 2026), Laval, Canada, 3–5 November 2026.