Diagnostic grade SHAP-prioritised machine learning for spirometry from high-dimensional qCT

Résumé fourni par la source

Background – Lung function characterisation with AI can streamline management of chronic airway diseases. Objective – Develop inverse models to estimate lung functional parameters from high-dimensional quantitative computed tomography (qCT) data, using structural metrics as constraints. Methods – qCT-derived features were extracted from (Assessment of Small Airways Involvement in Asthma) ATLANTIS ( NCT02123667 ) CT scans of 304 asthma and 50 control patients with the MI-TAP algorithm. Shallow learning models with Bayesian hyperparameter optimisation were trained to FEV1, FVC or FEV1/FVC labels. Recursive Feature Elimination (RFE)/Sequential Feature Selection (SFS) were applied to optimised models. SHapley Additive exPlanations (SHAP) linked predictions to structure, guiding RFE/SFS via prioritisation and dynamic Gaussian noise added training data variation. Results – Models predicted FEV1/FVC better than FEV1 or FVC. Combining RFE/SFS and SHAP prioritisation was more accurate than RFE/SFS alone - RFE-RF (10.32% MRE without vs 8.87% with SHAP) and SFS-SVM (9.48% vs 7.89%). Adding demographic/clinical data improved only SFS-kNN (8.95% MRE, 37/76 features). Key SHAP features were inspiratory/expiratory tissue volumes/densities, Pi10 and PRM4 (emphysema %). Dynamic gaussian noise improved SHAP SFS-SVM (5.16% MRE, 44/67 features) and SHAP SFS-kNN (1.65% MRE, 37/76 features), both outperforming benchmarks (p<0.01). Conclusion – Dynamic Gaussian noise and SHAP-prioritised feature selection enable ML models to accurately predict FEV1/FVC values from qCT data alone. A larger, more diverse training set could further optimise performance, whilst external validation of diagnostic grade MRE ±2.5% is needed.