DiagnoLLM: Integrating Bayesian Deconvolution and Large Language Models for Interpretable Disease Diagnosis

ACL ARR 2025 May Submission5532 Authors

Abstract: While deep learning models have shown strong performance in clinical disease classification, their black-box nature limits adoption in high-stakes healthcare settings. We present DiagnoLLM, a novel framework that combines Bayesian deconvolution with Large Language Model (LLM)-driven interpretability to bridge this gap. First, DiagnoLLM applies GP-unmix, a Gaussian Process-based hierarchical model, to infer cell-type-specific gene expression profiles from bulk RNA-seq and single-cell data. A deep learning model trained on these features achieves high predictive performance in Alzheimer's Disease (AD) classification (88.0\% accuracy). To enhance transparency, we introduce an LLM-based interpretability plug-in that generates faithful, audience-specific diagnostic reports grounded in model outputs, eQTL signals, and domain knowledge. The resulting reports align with clinical reasoning while maintaining fidelity to underlying predictions. DiagnoLLM demonstrates that LLMs, when used as structured narrative generators rather than classifiers, can play a critical role in building trust in biomedical AI. Code and data are available at: https://anonymous.4open.science/r/DiagnoLLM-0B4C.

Paper Type: Short

Research Area: NLP Applications

Research Area Keywords: Interpretable Disease Diagnosis, Bayesian Deconvolution, Large Language Models

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 5532