Predicting pulmonary function using deep-learning-generated synthetic data

Objective Pulmonary function test (PFT) is used to assess the risk of postoperative complications. However, in patients requiring emergency surgery, PFT is often not feasible. This study aimed to develop useful predictive models using synthetic data by Conditional Tabular Generative Adversarial Network (CTGAN). Methods Data were collected from patients who visited the respiratory medicine departments. Model performance was evaluated using R 2 , mean absolute error (MAE), and root mean square error (RMSE). Patients A total of 930 patient datasets were collected from two medium-sized hospitals. Results Prediction models for forced vital capacity (FVC) and forced expiratory volume in forced expiratory volume in 1 second (FEV₁) were constructed using multiple linear regression (MLR) and generalized linear models (GLM). After generating 100,000 synthetic datasets, neural network models (NNM) were developed. The prediction accuracy for FVC was higher than that of MLR in both males and females (R² = 0.63 for NNM, 0.53 for MLR in males, and 0.62 for MLR in females), whereas for FEV₁, the prediction accuracy was higher in males but lower in females (R² = 0.58 for NNM, 0.50 for MLR in males, and 0.65 for MLR in females). The prediction accuracy for FEV₁ was lower in females but higher in males (R² = 0.58 for NNM, 0.50 for MLR in males, and 0.65 for MLR in females). Conclusion NNM may be especially more effective than linear regression for patients with more severe and complex respiratory diseases and may therefore be useful for perioperative surgical risk assessment in such patients.