Abstract
Chronic obstructive pulmonary disease (COPD) remains one of the leading causes of morbidity and mortality worldwide, requiring accurate tools to predict and support clinical decisions. This study proposes a hybrid framework that combines a discrete Bayesian network and an NLP module to automatically extract key symptoms (shortness of breath (dyspnea), cough (cough), wheezing (wheezing), and phlegm (mucus)) from the texts of the MIMIC-IV-Note v2.2 statements. Structured demographic data and ICD-10 J44* codes have been combined with extracted text features and stored in Parquet for faster downloading. With a "naïve" BN structure (all signs → node copd) and the maximum likelihood method, the model showed ROC-AUC = 0.82 in a cohort of 12,528 hospitalizations. The integrated NLP module generates structured reports that include the likelihood of COPD, identified symptoms, and key risk factors, reducing documentation time and making decisions more informed. The proposed approach provides high predictive accuracy, transparency of conditional probabilities, and ease of implementation in clinical practice.