SHAP-Prioritised Machine Learning for Diagnostic-Grade Prediction of Lung Function

Automated machine learning (ML) can streamline the characterisation and management of chronic airway conditions. With the advent of quantitative CT (qCT) imaging allowing precise extraction of structural features from scans, assessment of airway obstruction levels could be automated to compliment traditional testing. This “feature known” approach has the added potential benefit of identifying key structure-function relationships through explainability measures. We therefore aimed to develop inverse models to estimate spirometry parameters from high-dimensional quantitative data using these structural metrics as constraints. With the ATLANTIS (NCT02123667) dataset, this paper experiments with a selection of ML methods, specifically k-nearest neighbours (kNN), random forest (RF) and support vector machine (SVM), to predict spirometry values (Forced Expiratory Volume (FEV1), Forced Vital Capacity (FVC) and FEV1/FVC). The dynamic ratio FEV1/FVC was predicted better by all models than FEV1 or FVC. Results show effective counteraction to high-dimensionality through iterative feature refinement guided by SHapley Additive exPlanations (SHAP), and to limited training data through dynamic Gaussian noise (DGN). Diagnostic-grade prediction accuracy was achieved with DGN SHAP sequential feature selection (SFS)-kNN at 1.64% MRE with 37/76 features. A selection of typical variables including expiratory tissue density and lung volume, vasculature and airway geometries were seen to be important for prediction. This approach therefore can not only predict pulmonary function, but also extract useful structural information in a dynamic airway system through linking back to personalised abnormalities.