SpiroLLM: Multimodal COPD Reporting

• SpiroLLM is a multimodal large language model that integrates respiratory waveforms, quantitative pulmonary function metrics, and patient demographics to generate pulmonologist-style COPD reports.
• It employs a hybrid CNN-BiLSTM encoder and a SpiroProjector to map spirogram features into a pretrained LLM’s embedding space, enhancing diagnostic accuracy and interpretability.
• Evaluated on a UK Biobank cohort, SpiroLLM outperforms conventional classifiers by improving sensitivity and robustness, especially under conditions of incomplete data.

Searching arXiv for the specified paper and closely related work to ground the article. SpiroLLM is a multimodal LLM for interpreting spirogram time series and generating COPD diagnostic reports, introduced in "SpiroLLM: Finetuning Pretrained LLMs to Understand Spirogram Time Series with Clinical Validation in COPD Reporting" (Mei et al., 22 Jul 2025). It is presented as the first multimodal LLM designed to understand spirogram time series for COPD reporting, with the explicit aim of bridging signal-based pulmonary function analysis and narrative clinical report generation. The system combines respiratory waveform encoding, numerical pulmonary function test (PFT) values, demographic variables, and a pretrained LLM backbone to produce clinician-style reports while retaining diagnostic utility and interpretability. The model is trained and validated on a UK Biobank cohort of 234,028 individuals and is positioned as a clinical decision-support framework rather than a pure classifier (Mei et al., 22 Jul 2025).

1. Clinical setting and problem formulation

Chronic Obstructive Pulmonary Disease is described as a major chronic respiratory disease with persistent airflow limitation and as a leading global cause of disability and mortality (Mei et al., 22 Jul 2025). In this setting, respiratory spirogram time series obtained during pulmonary function tests are central to early detection of respiratory disease and to longitudinal monitoring of lung function. The paper frames expert spirometry interpretation as both clinically important and operationally difficult: interpretation is time-consuming and uneven across care settings, while most existing AI systems for COPD diagnosis from respiratory curves remain classification-oriented and do not provide an explanatory diagnostic rationale (Mei et al., 22 Jul 2025).

The work identifies two complementary limitations in prior approaches. First, deep-learning systems operating on spirometry curves can classify COPD but are typically "black boxes." Second, standard LLMs can generate fluent clinical prose but cannot directly interpret raw respiratory waveforms. SpiroLLM is designed to fuse these capabilities so that a model can read spirometry, reason over structured PFT values, and generate a comprehensive diagnostic report in a form intended to resemble pulmonologist reporting (Mei et al., 22 Jul 2025).

This suggests that the primary novelty is not merely multimodal ingestion, but the coupling of physiological signal interpretation with report generation in a clinically constrained domain. A plausible implication is that the model targets the longstanding separation between signal-processing pipelines and narrative reporting systems in pulmonary diagnostics.

2. Dataset construction and supervision strategy

The model is trained and validated on a UK Biobank cohort of 234,028 individuals, screened through strict quality control so that spirometry validity and measurement consistency were preserved (Mei et al., 22 Jul 2025). The cohort includes COPD cases and non-COPD controls selected under the same criteria, and the final dataset is split 8:1:1 into training, validation, and test sets using stratified sampling. The paper presents this scale as enabling supervised fine-tuning of a report-generating model in a domain where fully manual expert annotation would otherwise be prohibitively expensive (Mei et al., 22 Jul 2025).

A major component of the training pipeline is semi-automated target report construction. Three information sources are combined: morphology descriptions generated from curve images by Qwen2.5-VL-72B, quantitative pulmonary metrics computed with SpiroUtils, and clinical guideline snippets retrieved from a GOLD 2025 knowledge base through retrieval-augmented generation. These are integrated by DeepSeek-V3 into a structured target report $r_i^*$ (Mei et al., 22 Jul 2025). SpiroUtils computes standard metrics such as FVC, FEV1, FEF25–75%, and FEF75%, and derives predicted values and Z-scores using GLI 2012 reference equations (Mei et al., 22 Jul 2025).

This supervision design is central to the paper’s methodology. Rather than relying on large-scale handcrafted reports, it uses an overview pipeline anchored in morphology descriptions, quantitative lung-function metrics, and guideline language. The authors present this as a way to reduce annotation burden while keeping targets aligned with authoritative clinical reasoning (Mei et al., 22 Jul 2025).

3. Architecture and multimodal alignment

SpiroLLM has two principal multimodal components: the SpiroEncoder and the SpiroProjector (Mei et al., 22 Jul 2025). The SpiroEncoder reuses the prior DeepSpiro encoder, described as a hybrid CNN-BiLSTM time-series network. The CNN captures local waveform patterns in the spirometry sequence, and the BiLSTM models temporal context, producing an embedding

$$E_i \in \mathbb{R}^{L_i \times D_{feat}} = E_s(s_i \mid \theta_E),$$

where $s_i$ is the raw pulmonary function time series and $\theta_E$ denotes encoder parameters (Mei et al., 22 Jul 2025). The embedding is intended to preserve both fine-grained curve morphology and longer-range trajectory structure.

The SpiroProjector is a lightweight MLP alignment layer that maps SpiroEncoder features into the embedding space of the pretrained LLM. It is formulated as

$$\mathbf{P}_{i} = \operatorname{SpiroProjector}(\mathbf{E}_{i}\mid \theta_{P}) = \operatorname{Dropout}\!\left(\operatorname{ReLU}(\mathbf{E}_{i}\mathbf{W}_{1}+\mathbf{b}_{1})\right)\mathbf{W}_{2}+\mathbf{b}_{2},$$

with $\theta_P=\{W_1,b_1,W_2,b_2\}$ (Mei et al., 22 Jul 2025). The paper states that the projector first performs dimensional alignment from the signal feature size to the LLM embedding dimension and then refines the representation nonlinearly.

An important training-stage distinction is that, during projector pretraining, the SpiroEncoder and LLM are frozen while only the SpiroProjector is trained. This stage is intended to learn a cross-modal mapping from physiological features toward textual semantics, especially the space of morphological descriptions (Mei et al., 22 Jul 2025). The full model then combines projected spirogram features with a structured textual prompt containing demographic variables and numerical PFT measurements. The prompt includes the COPD probability output from the encoder side, key PFT parameters extracted by SpiroUtils, and demographic information such as age, sex, smoking history, and height (Mei et al., 22 Jul 2025).

The LLM backbone is Llama3.1-8B, adapted using LoRA parameter-efficient fine-tuning. During training, the LLM backbone is frozen except for LoRA adapters, while the SpiroProjector is updated; optimization uses AdamW, cosine annealing with warmup, mixed bfloat16 precision, and early stopping (Mei et al., 22 Jul 2025). In effect, the pretrained model is adapted to read projected spirogram embeddings as part of a multimodal prompt rather than being retrained from scratch.

4. Report generation and evaluation protocol

The report-generation pipeline is explicitly multimodal. Projected respiratory-curve features are concatenated with structured textual information so that the LLM can generate a comprehensive diagnostic report rather than a label alone (Mei et al., 22 Jul 2025). The paper’s emphasis is that report generation should preserve both diagnostic conclusion and evidence traceability, particularly curve morphology and quantitative PFT interpretation.

Evaluation extends beyond surface-level text overlap. The authors use an "LLM-as-a-Judge" framework in which DeepSeek-V3 scores generated reports on six clinical dimensions: factual accuracy, completeness/coverage, logical evidence, medical terminology, medical safety, and curve description (Mei et al., 22 Jul 2025). Scores are mapped from 1–5 to a 0–100 scale. In addition, diagnostic performance is extracted by the judge model, which outputs a binary COPD decision and confidence score for each report, allowing AUROC, AUPRC, and F1-score computation (Mei et al., 22 Jul 2025).

The use of report-level scoring rather than conventional overlap metrics reflects the paper’s clinical framing. The target is not generic summarization quality, but whether the generated report is factually grounded, medically safe, diagnostically coherent, and morphologically faithful. This suggests that SpiroLLM is evaluated as a reporting system whose language is part of the clinical output, not as a detached classifier with explanatory postprocessing.

5. Quantitative performance and robustness

The main diagnostic result reported for SpiroLLM is an AUROC of 0.8980 with a 95% confidence interval of 0.8820–0.9132, together with an AUPRC of 0.9049, sensitivity of 0.8327, specificity of 0.6699, and F1-score of 0.7435 (Mei et al., 22 Jul 2025). Relative to the text-only or multimodal prompt variant SpiroLLM-pftonly, the full multimodal model has similar overall discrimination but improved sensitivity, from 0.7782 to 0.8327, which the authors describe as a 7% relative gain and as clinically meaningful for reducing missed COPD cases (Mei et al., 22 Jul 2025). Compared with the standalone DeepSpiro classifier, SpiroLLM improves AUROC from 0.8266 to 0.8980 and AUPRC from 0.8068 to 0.9049 (Mei et al., 22 Jul 2025). The unfine-tuned Llama3.1-8B baseline performs substantially worse, with AUROC 0.7690 and specificity 0.1248 (Mei et al., 22 Jul 2025).

On report quality, SpiroLLM receives scores of 78.36 for factual accuracy, 86.39 for completeness, 81.63 for logic/evidence, 95.62 for medical terminology, 89.03 for medical safety, and 85.76 for curve description (Mei et al., 22 Jul 2025). The related SpiroLLM-pftonly variant is slightly higher on factual accuracy and safety in the fully informed setting, but the paper argues that direct curve understanding does not degrade textual quality materially (Mei et al., 22 Jul 2025).

The strongest robustness evidence comes from missing-data evaluation. In a masking experiment where core numerical PFT values are removed, SpiroLLM-pftonly collapses: valid response rate falls from 100% to 13.4%, AUROC drops to 0.5575, and F1-score to 0.0048 (Mei et al., 22 Jul 2025). Under the same condition, SpiroLLM maintains a 100% valid response rate and still achieves AUROC 0.8688 with F1-score 0.6990 (Mei et al., 22 Jul 2025). The authors interpret this as showing that the SpiroEncoder contributes an independent signal channel rather than a redundant supplement.

Setting	Key result
Full SpiroLLM	AUROC 0.8980; AUPRC 0.9049
SpiroLLM-pftonly	Sensitivity 0.7782
Missing core PFT values, SpiroLLM	100% valid response rate; AUROC 0.8688
Missing core PFT values, SpiroLLM-pftonly	13.4% valid response rate; AUROC 0.5575

These findings situate SpiroLLM not only as a multimodal reporting model but as a robustness-oriented architecture in which physiological time series provide resilience when structured numeric inputs are incomplete.

6. Clinical validation, interpretability, and limitations

The study includes clinical validation beyond automated scoring. In a case study, the baseline Llama model incorrectly labeled a non-COPD case as COPD because it over-weighted "below predicted" values and failed to prioritize the FEV1/FVC ratio relative to LLN (Mei et al., 22 Jul 2025). SpiroLLM, by contrast, correctly ruled out COPD by combining numerical criteria with the visual morphology of the flow-volume curve (Mei et al., 22 Jul 2025). Senior pulmonary function experts then reviewed generated reports using the same six-dimensional rubric and generally found them high quality, especially noting accurate curve morphology description and correct application of GOLD-based diagnostic logic (Mei et al., 22 Jul 2025).

The same expert review also surfaced limitations. The paper notes terminology issues, including overly assertive phrasing about reversibility when bronchodilator comparison data were unavailable (Mei et al., 22 Jul 2025). This is a clinically consequential limitation because subtle wording in pulmonary reporting can imply evidence not actually present in the source measurements.

A common misconception would be to regard SpiroLLM as simply an LLM wrapper around spirometry classification. The paper instead frames it as an interpretable reporting system that integrates raw physiological signals, structured PFT values, and narrative reasoning (Mei et al., 22 Jul 2025). Another possible misconception would be to infer that multimodal input always improves every textual metric; the reported results are more specific. The full model remains highly competitive in report quality, but some text-only-with-PFT prompting scores are slightly higher in fully informed conditions, while the multimodal design becomes decisive under missing-data stress tests (Mei et al., 22 Jul 2025).

7. Position within related research and broader significance

SpiroLLM is situated at the intersection of physiological time-series modeling, multimodal alignment, and clinical LLM adaptation. Its central claim is that deeply fusing physiological signals with a pretrained LLM establishes a new paradigm for interpretable and reliable COPD decision support (Mei et al., 22 Jul 2025). Within the paper’s own comparison set, it improves on a classification-only baseline, DeepSpiro, and substantially outperforms an unfine-tuned general-purpose LLM, indicating that neither signal-only classification nor text-only generation alone is sufficient for the task (Mei et al., 22 Jul 2025).

The name should be distinguished from similarly titled but unrelated work in spiking-neural-network language modeling, such as "SpikeLM: Towards General Spike-Driven Language Modeling via Elastic Bi-Spiking Mechanisms" (Xing et al., 2024) and "SpikeLLM: Scaling up Spiking Neural Network to LLMs via Saliency-based Spiking" (Xing et al., 2024). Those papers concern spike-driven language modeling and spiking LLMs, whereas SpiroLLM concerns spirogram understanding for COPD reporting (Xing et al., 2024, Xing et al., 2024).

The broader significance claimed for SpiroLLM lies in automating draft COPD reports, improving efficiency and consistency, and extending standardized pulmonary interpretation to settings where specialist expertise is limited (Mei et al., 22 Jul 2025). This suggests a practical role in report drafting and triage rather than autonomous diagnosis. A plausible implication is that its multimodal design could generalize to other physiological signal domains in which waveform morphology and numerical measurements jointly determine the clinical narrative, but such an extension is not itself established in the reported study.