Abstract: FR-PO0032
Comparing Large Language Model and Human Coding for Qualitative Analysis of Mindfulness in CKD: Insights from "The Mindful Kidney" Project
Session Information
- Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
October 23, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Gudino, Paola, Yale University, New Haven, Connecticut, United States
- Griffith, Frances J., University of Kentucky, Lexington, Kentucky, United States
- Hanson, Kobe Richard, Yale University, New Haven, Connecticut, United States
- Ash, Garrett I., Yale University, New Haven, Connecticut, United States
- Sarav, Menaka, Yale University, New Haven, Connecticut, United States
Background
Mindfulness-based interventions are effective for reducing the psychosocial burden of chronic kidney disease (CKD), but traditional qualitative analysis of patient experience is resource-intensive. Large Language Models (LLMs) offer a potential solution for scaling qualitative research. This study evaluates the feasibility and accuracy of LLM-assisted coding compared to gold-standard human analysis within "The Mindful Kidney" project.
Methods
We conducted semi-structured interviews with 5 CKD patients regarding CKD-related digital health experiences for physical and mental health. We compared two workflows for qualitative analysis: 1) Clinician-led deductive coding using human and LLM interrater reliability checks and 2) LLM-led inductive coding using Claude Sonnet 4.6 (Anthropic) via Claude Code.
Results
There was concordance in the two workflows
Clinician-led deductive coding identified the following patterns:
Experiences with CKD: Most participants described a spectrum of emotions and working to practice health-promoting lifestyle behaviors.
Prior experiences with digital health: All participants reported previous positive experiences
User needs and preferences for digital health: All participants reported interest in learning and using digital health tools
LLM inductive analysis surfaced 7 cross-cutting patterns:
Emotional response to CKD varied with patients' life history
Participants were managing their CKD in social isolation.
Patients experienced CKD either through either through lab numbers (e.g., eGFR trends) or through physical symptoms.
Participants wanted health data but preferred passive, nonintrusive monitoring.
Digital health and mental health interventions should be empowering and self-directed, not burdensome.
Structural barriers — language, cost, disability benefit systems — outweighed technical obstacles to using digital health.
The inductive themes largely corroborate the deductive coding while surfacing contextual dimensions that the codebook does not fully capture.
Conclusion
LLMs provide a scalable, high-efficiency adjunct for qualitative nephrology research. By automating deductive tasks and facilitating theme discovery, AI-assisted workflows can accelerate the integration of patient-centered insights into digital health interventions for the CKD population.
Funding
- Private Foundation Support