SpiroLLM: Finetuning pretrained LLMs to understand spirogram time series with clinical validation in COPD reporting

Area of Science:

• Artificial Intelligence in Medicine
• Respiratory Medicine
• Medical Diagnostics

Background:

• Chronic Obstructive Pulmonary Disease (COPD) is a significant global health burden.
• Pulmonary Function Tests (PFTs) generate vital spirogram data for lung health monitoring.
• Current AI diagnostic tools lack interpretability, hindering clinical adoption.

Purpose of the Study:

• To develop the first multimodal large language model (LLM) capable of understanding spirograms for COPD diagnosis.
• To enhance the interpretability and clinical trust of AI-driven diagnostic tools.

Main Methods:

• Leveraged UK Biobank data from 234,028 individuals.
• Developed SpiroLLM, integrating a SpiroEncoder and SpiroProjector to process spirogram morphology and PFT values.
• Utilized a unified latent space for feature alignment.

Main Results:

• SpiroLLM achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.8977.
• Demonstrated superior robustness with a 100% valid response rate on incomplete data, compared to 13.4% for text-only models.
• Successfully generated comprehensive diagnostic reports from spirogram data.

Conclusions:

• SpiroLLM represents a novel paradigm for interpretable clinical decision support.
• Deep fusion of physiological signals with LLMs holds substantial potential for advancing respiratory disease diagnostics.
• The multimodal approach significantly improves AI reliability and clinical applicability in PFT analysis.