A Machine Learning-Based Multimodal Model Integrating Radiomics and Clinical Features for Predicting Interstitial Lung Disease: Development and External Validation.
Authors
Affiliations (5)
Affiliations (5)
- Department of Pulmonary and Critical Care Medicine, The Affiliated Huai'an Hospital of Xuzhou Medical University, Huai'an, China (Y.W., X.D., R.F.).
- The Affiliated Huai'an Hospital of Xuzhou Medical University, Huai'an, China (Z.Z.).
- Department of Radiology, Huai'an Hospital Affiliated to Yangzhou University, the Fifth People's Hospital of Huai'an, Huai'an, China (B.S.).
- Information & Statistical Center, The Affiliated Huai'an Hospital of Xuzhou Medical University, Huai'an, China (Y.Z.).
- Department of Pulmonary and Critical Care Medicine, The Affiliated Huai'an Hospital of Xuzhou Medical University, Huai'an, China (Y.W., X.D., R.F.). Electronic address: [email protected].
Abstract
To develop and validate a multimodal model for predicting the risk of interstitial lung disease (ILD) by integrating radiomics features with clinical variables using machine learning (ML). In this multicenter retrospective study, 2456 subjects at risk for ILD from three centers between January 2018 and January 2023 were included. A development cohort (Centers A and B, n = 1842) was used for model training and internal validation, while an independent cohort (Center C, n = 614) was used for external validation. Radiomics features extracted from baseline chest CT images were used to construct a radiomics score (Rad-score) using least absolute shrinkage and selection operator regression. Clinical, radiomics, and multimodal fusion models were developed. Additional ML algorithms were explored for comparative analysis. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration, and decision curve analysis. The multimodal fusion model achieved the best discriminative performance in the training cohort (AUC: 0.738, 95% CI: 0.703-0.772), with stable performance in the internal validation cohort (AUC: 0.733, 95% CI: 0.680-0.786) and external validation cohort (AUC: 0.715, 95% CI: 0.661-0.770). The multimodal model outperformed single-modality models and showed favorable calibration and clinical utility. Comparative analyses demonstrated no substantial improvement in external validation performance with more complex ML algorithms. An ML-based multimodal model integrating CT radiomics and clinical features demonstrated stable performance for ILD risk prediction across multicenter datasets, suggesting potential value for early ILD risk stratification.