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

Authors: Shuhao Mei, Yao Long, Xiaoyu Xiao, Shan Cao, Xiaobo Han, Shijia Geng, Jian‐Sheng Sun, Yuxi Zhou, Shenda Hong Date: 2026-03-24 Paper ID: openalex:2507.16145

Summary

SpiroLLM is introduced as the first multimodal Large Language Model designed to interpret respiratory spirogram time series for Chronic Obstructive Pulmonary Disease (COPD) diagnosis and reporting. The model utilizes a novel framework where a SpiroEncoder extracts morphological features from the time series, which are then aligned with standard Pulmonary Function Test (PFT) numerical values via a SpiroProjector into a unified latent space consumable by an LLM backbone. Experiments on the UK Biobank dataset show SpiroLLM achieves a diagnostic AUROC of 0.8977 and exhibits significantly higher robustness when core data is missing compared to text-only baselines. This work establishes a new paradigm for creating interpretable and reliable clinical decision support systems by deeply integrating physiological signals with generative language models.

Key Contributions

• Development of SpiroLLM, the first multimodal large language model capable of understanding and reasoning about respiratory spirogram time series data.
• Introduction of the SpiroEncoder and SpiroProjector to extract morphological features and align them with PFT numerical values in a unified latent space.
• Achieving a diagnostic AUROC of 0.8977 on COPD classification using spirogram analysis, demonstrating strong clinical performance.
• Demonstration of superior robustness against missing core data compared to text-only models, maintaining a 100% valid response rate.

Limitations

The paper focuses specifically on COPD reporting; generalization to other respiratory diseases or medical time-series requires further testing. Clinical deployment requires rigorous prospective validation beyond the reported retrospective analysis.

Open Questions & Future Work

• spiro-llm-demographic-generalization
• spiro-llm-prospective-clinical-assessment

Key Concepts

• spiro-encoder: A dedicated component used to extract morphological features from spirogram time series data to enable LLM understanding.
• spiro-projector: A projection layer that aligns the morphological features extracted from the SpiroEncoder with standard PFT numerical values into a shared latent representation space.

Archivist Review

Archivist review kept only candidates judged central to the paper and reusable across future work. Approved 2 concept(s), 2 open question(s), and 0 dataset(s), with 1 rejected candidate note(s).

Approved Concepts

• SpiroEncoder: It is the core component responsible for extracting meaningful morphological features from the spirogram time series data, which is essential for the multimodal integration.
• SpiroProjector: This module bridges the gap between the extracted time-series features and the LLM’s input space, making the fusion of modalities explicit and unified.

Approved Open Questions

• Verify demographic generalization: Lack of demographic diversity in training data limits the model’s applicability and fairness across global populations, a critical consideration for clinical tools.
• Conduct prospective clinical pilot studies},{background:: Assessing the model’s real-world impact on clinical efficiency and report utility via prospective testing is necessary to move from technical demonstration to validated clinical integration.

Rejected Candidates

• [dataset] UK Biobank (UKB) (uk-biobank-ukb) - low_impact: The dataset is a standard, large-scale resource commonly used in biomedical research and does not possess unique methodological significance warranting a standalone vault entry.

Links

• Abstract
• PDF