C73-23 A Multimodal Tabpfn-based Framework For Estimating Spirometric Pulmonary Function From Ct And Demographic Data

Rationale

Reliable estimation of pulmonary function from imaging and clinical information offers a noninvasive alternative for identifying chronic obstructive pulmonary disease (COPD). Nonetheless, many current methods fail to fully capture airway structural organization and regional functional variability. This work aims to design and assess a multimodal approach that enables accurate prediction of spirometric measures and robust classification of COPD.

Methods

We present a multimodal prediction framework that integrates three complementary feature sources: (1) airway structural information represented through an edge-convolution graph neural network (EdgeGNN); (2) spatial functional heterogeneity extracted using a three-dimensional convolutional neural network based on parametric response mapping (3D-cPRM); and (3) demographic and morphological variables modeled with an artificial neural network (ANN). The resulting features were concatenated, normalized using a robust scaling strategy, and fed into a transformer-based tabular prior-data fitted model (TabPFN) combined with a MultiOutputRegressor to simultaneously estimate forced expiratory volume in one second (FEV1), forced vital capacity (FVC), and the FEV1/FVC ratio. COPD status was determined by applying a 0.70 threshold to the predicted FEV1/FVC. Model interpretability was examined using Shapley additive explanations (SHAP).

Results

Using five-fold cross-validation, the proposed method showed robust performance for both regression and classification. The estimation of FEV1, FVC, and FEV1/FVC yielded R2 values of 0.7939, 0.6585, and 0.8478, respectively, demonstrating strong concordance with spirometric measurements, especially for the FEV1/FVC ratio. COPD identification achieved an accuracy of 97.53%. SHAP-based interpretability analysis revealed that airway structural information, regional ventilation characteristics, and clinical-demographic variables contributed in a complementary manner to pulmonary function prediction. Features derived from the ANN were the primary drivers for FEV1 and FVC estimation, with additional support from EdgeGNN-extracted airway topology, whereas prediction of FEV1/FVC showed a more balanced influence across modalities, including a notable contribution from PRM-based functional features.

Conclusions

By integrating structural, functional, and clinical data using EdgeGNN, 3D-cPRM, ANN, and TabPFN, and simultaneously predicting FEV1, FVC, and the FEV1/FVC ratio, this framework provides a strong alternative to traditional modeling strategies and supports noninvasive assessment of respiratory disorders. Key words: COPD, spirometry prediction, pulmonary function, graph neural networks, convolutional neural network, artificial neural network, TabPFN.

This abstract is funded by: the National Research Foundation of Korea (NRF) grants funded by the Korean government (MSIT) [RS-2023-NR077008, RS-2024-00407902], and by the Korea Ministry of Environment as “The Environmental Health Action Program” [2018001360004].