Introduction of a new approach to interpret pulmonary function tests (PFT) based on Machine learning and Game theory

Abstract

Objective: We introduce a new interpretation method for PFT data. Methods: Our new method consists of 3 steps: To build a diagnostic rule for a specific target from multi-PFT parameters. To estimate the Shapley-score, a game-theory based metric which measures the importance of each PFT index by comparing the model’s predictions with and without that index. To generate the interpretation, indicating the contribution level of each parameter to the positive/negative diagnosis. We applied this method to detect interstitial lung disease (ILD) in patients with systemic sclerosis. A machine learning diagnostic rule was developed from data of 300 patients undergoing 5 PFT techniques. Results: Validated on unseen data (n=100), the rule showed a good clinical performance (Sensitivity=0.86, Specificity=0.89, Positive and negative Likelihood-ratios of 7.86 and 0.16 respectively). At population level, the interpretation revealed that AV, TLCO-NO and FEV1 were the most important contributors to positive diagnosis. Personalized interpretations (Fig. 1) allow to verify the model’s plausibility in difficult cases with false negative/positive diagnosis and identify individual patterns of lung function impairments. Conclusion: Compared to the conventional approaches, our new method offers more advantages including ability of multivariate interpretation and valuable information for personalized treatment plan.