The value of machine learning inference to predict multidisciplinary discussion diagnosis of interstitial lung diseases.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, Division of Cardiothoracic Imaging, Perelman School of Medicine, University of Pennsylvania, 3400 Spruce Street, Ground Floor Founders Bldg, Philadelphia, Pennsylvania, USA. Electronic address: [email protected].
- Department of Radiology, Division of Cardiothoracic Imaging, Perelman School of Medicine, University of Pennsylvania, 3400 Spruce Street, Ground Floor Founders Bldg, Philadelphia, Pennsylvania, USA. Electronic address: [email protected].
- Department of Radiology, Division of Cardiothoracic Imaging, Perelman School of Medicine, University of Pennsylvania, 3400 Spruce Street, Ground Floor Founders Bldg, Philadelphia, Pennsylvania, USA. Electronic address: [email protected].
- Department of Radiology, Division of Cardiothoracic Imaging, Perelman School of Medicine, University of Pennsylvania, 3400 Spruce Street, Ground Floor Founders Bldg, Philadelphia, Pennsylvania, USA. Electronic address: [email protected].
Abstract
Multidisciplinary discussion (MDD) is the reference standard for diagnosing interstitial lung diseases (ILD), however, it requires considerable expertise, and up to 25% of patients remain unclassifiable. We evaluated whether machine learning (ML) models trained on multimodal data can accurately replicate MDD consensus ILD diagnoses. We retrospectively identified 428 patients with ILD confirmed by MDD consensus at a tertiary academic center from 2010 to 2021. Using comprehensive clinical and radiological data, we trained and cross-validated multiple ML models for three classifications: (1) three-level: connective tissue disease associated ILD (CTD-ILD), usual interstitial pneumonia (UIP), or unclassifiable; (2) UIP/IPF versus all other, and (3) unclassifiable versus all other. Logistic regression, random forest, and neural networks ML models were trained with five-fold cross-validation. Performance was compared via area under the curve (AUC) with 95% confidence intervals. MDD diagnoses were UIP/IPF in 131 patients (30.6%), CTD-ILD in 130 (30.4%), and unclassifiable ILD in 167 (39.0%). Clinical variables and CT imaging features were more predictive than PFTs. ML models achieved AUC values ranging from 0.66 to 0.92. For UIP/IPF vs. other, the neural network reached an AUC of 0.88. Notably, the Lasso regression model exhibited high specificity (0.966), suggesting utility as a high-confidence "rule-in" AI tool. ML classification closely approximated MDD diagnoses of ILD, especially for UIP/IPF, relying primarily on clinical and CT imaging features, even without pathology information, allowing expert-level diagnostic performance where MDD resources are not available.