Bayesian Non-Parametric Classification with Tree-Based Feature Transformation for NIPPV Efficacy Prediction in COPD Patients

Abstract

Non-invasive positive pressure ventilation (NIPPV) is a life-saving approach which was developed to reduce the complications of endotracheal intubation and invasive ventilation in patients with chronic obstructive pulmonary disease (COPD). However, it has a certain probability of invalid. Failure of NIPPV will lead to an increase in mortality, which highlights the importance of rational diagnosis about the need for NIPPV therapy. In order to avoid delaying endotracheal intubation, we proposed a hybrid model which combine tree-based feature transformation with Bayesian non-parametric classification, to predict whether the patient should adopt NIPPV based on the their own physical condition. We delved into the feature importance and justified the rationality of using tree-based feature transformation. The proposed gaussian process classification (GPC) with gradient boosting decision tree (GBDT) feature transformation model has shown state-of-the-art results on both the NIPPV dataset and two simulated datasets with larger sample size. For critically ill COPD patients, the proposed method provides diagnostic assistance for physicians' decision making and avoids delaying endotracheal intubation or mechanical ventilation.