SpiroFusion-Net: Multimodal Deep Learning on Raw Spirograms & Linked EHR

Official open-source implementation and supplementary repository for the IEEE Journal of Biomedical and Health Informatics (J-BHI) study:

SpiroFusion-Net: Cross-Modal Deep Learning on Raw Spirograms and EHR for COPD Detection, Blow Salvage, and Survival Prediction Md Emran Hasan, Yue-Fang Wu, and Dong-Jun Yu. Repository: https://github.com/emrancub/SpiroFusion-Net

---

Key Highlights

• Dual Population Cohorts (N = 462,825 combined spirograms):
  • Primary Discovery Cohort: UK Biobank Application Number 419012 (N = 453,186 valid spirograms; 22,392 COPD cases; 47,997 deaths over 16.0 years; held-out test set N = 67,980).
  • Independent External Validation Cohort: US CDC National Health and Nutrition Examination Survey (NHANES 2007–2012, Cycles E, F, G; N = 9,639 multi-ethnic US adults; held-out test set N = 4,820).
• Physiologically Grounded 4-Channel Waveform Tensor (4 x 1,000) & 12 Small-Airway Curve Biomarkers:
  • Encodes absolute volume V(t), fractional volume V(t)/FVC, expiratory flow dV/dt, and flow acceleration d2V/dt2 alongside FEF25-75%, FEF50%, FEF75%, central/distal loop concavity indices, and AUFVL.
• QC-Aware Bidirectional Cross-Modal Attention:
  • Dynamically fuses a Multi-Scale 1D Squeeze-and-Excitation Residual Network (SE-ResNet-1D) with a Gated Residual Tabular Network (GRN), rescuing diagnostic performance in suboptimal/rejected blows (EXPFLOW / BELOW6SEC and NHANES QC Grades C & D).

Main Results Summary

Evaluation Benchmark	Clinical GOLD	LightGBM	1D-Transformer (PatchTST)	SpiroFusion-Net (Proposed)
UK Biobank Test AUROC (N = 67,980)	0.7821	0.8688	0.7988	0.8672
UK Biobank Test AUPRC	0.2125	0.4068	0.3055	0.4278
UK Biobank Suboptimal Blow Salvage AUROC (N = 11,996)	0.6864	0.8012	0.7640	0.8222
UK Biobank 16-Year Mortality C-Index (47,997 deaths)	0.5980	0.7070	0.6204	0.7454
External US CDC NHANES Overall AUROC (N = 4,820)	0.8932	0.9197	0.8931	0.9296
External US CDC NHANES AUPRC	0.8099	0.8482	0.7980	0.8810
External US CDC NHANES Post-Bronchodilator AUROC	0.9832	0.9717	0.9774	0.9946
External US CDC NHANES Suboptimal Blow Salvage AUROC	0.8985	0.9600	0.9073	0.9453

Repository Structure

• config.py: Central configuration for UK Biobank and US CDC NHANES schemas and hyperparameters.
• data/: 4-channel 10-ms waveform preprocessing (preprocessing.py), 12 small-airway curve biomarker extraction, and PyTorch Dataset loader (dataset.py).
• models/: SpiroFusionNet architecture (multimodal_net.py) and 11 clinical, classical ML, GBDT, and unimodal deep baselines (tabular_baselines.py).
• training/: Multi-task Focal-BCE + FP32 Cox proportional hazards trainer (trainer.py).
• evaluation/: Bootstrap 95% CIs, DeLong tests, C-index, Bayes prior calibration (metrics.py), and publication figure/table generators (publication_outputs.py).
• run_nhanes_external_validation.py: Automated downloader and PCHIP curve builder for US CDC NHANES 2007–2012 (.XPT modules).
• run_pipeline.py: End-to-end training, evaluation, external validation, and figure/table generation script.
• Supplementary_Material_IEEE_JBHI.pdf: Extended supporting tables (Table S1–Table S4, including SOTA comparison) and figures (Fig. S1–Fig. S8).
• outputs/figures/ & outputs/tables/: All 15 figures (.pdf and .png) and 10 CSV tables.

Data Availability & Acknowledgments

All source code, figures, and supplementary material required to reproduce the main results of this study are freely available at https://github.com/emrancub/SpiroFusion-Net. Individual-level primary biomedical datasets are available from the UK Biobank (https://www.ukbiobank.ac.uk/) upon formal research application and approval (UK Biobank Project Application Number 419012), and the external validation datasets are publicly available from the US CDC National Health and Nutrition Examination Survey (https://wwwn.cdc.gov/nchs/nhanes/).

📜 Citation

If you use this framework or ideas from this work, please cite:

The final BibTeX entry and author list will be updated after acceptance.

📧 Contact

For questions, feedback, or collaboration:

• Md Emran Hasan – mdemranhasan@njust.edu.cn or/and, Dong-Jun Yu – njyudj@njust.edu.cn

You are welcome to open an issue or pull request if you have suggestions or would like to contribute improvements after the journal’s decision.