Construction and Validation of a Machine Learning Model Based on Clinical and Microbiomic Features for Predicting High Mucus Secretion in COPD.
To evaluate clinical and airway microbiome features of excessive mucus secretion (CMH) in COPD progression and apply machine learning for CMH status identification.
A total of 319 COPD patients from Changzhi People's Hospital (May 2020-March 2024) were consecutively enrolled and divided by sputum volume and characteristics into a high mucus secretion group (n=173) and a non-high mucus secretion group (n=146). Patients were randomly assigned to training (80%) and testing (20%) sets. Airway microbiome structure was analyzed via 16S rRNA sequencing. From clinical and microbiome data, 70 features were extracted. Six machine learning algorithms (SVM, KNN, RF, BN, GBDT, NN) were used to build classification models. Feature selection employed filtering methods, and hyperparameters were optimized by 10-fold cross-validation. Model performance was assessed using sensitivity, specificity, accuracy, and AUC.
The CMH group and the non-CMH group differed significantly in a number of factors, including age, the length of the disease, and pulmonary function indices, according to a comparison of baseline patient data. Analysis of airway microbiome characteristics revealed that the CMH group had significantly lower observed ASVs and Shannon indices (p<0.001), along with significant enrichment of potentially pathogenic bacterial genera such as Haemophilus and Pseudomonas. Following feature selection, disease duration, Haemophilus abundance, history of AECOPD, Pseudomonas abundance, and predicted FEV1% were identified as significant predictive factors. With a sensitivity of 0.867, specificity of 0.789, PPV of 0.805, NPV of 0.855, and AUC of 0.911, the Bayesian Network (BN) model outperformed the other six machine learning models on the testing sets; its generalization ability was significantly superior to other algorithms such as SVM and RF.
CMH in COPD is linked to airway dysbiosis and pathogen enrichment. The BN model effectively identifies this phenotype with strong generalization ability.
A total of 319 COPD patients from Changzhi People's Hospital (May 2020-March 2024) were consecutively enrolled and divided by sputum volume and characteristics into a high mucus secretion group (n=173) and a non-high mucus secretion group (n=146). Patients were randomly assigned to training (80%) and testing (20%) sets. Airway microbiome structure was analyzed via 16S rRNA sequencing. From clinical and microbiome data, 70 features were extracted. Six machine learning algorithms (SVM, KNN, RF, BN, GBDT, NN) were used to build classification models. Feature selection employed filtering methods, and hyperparameters were optimized by 10-fold cross-validation. Model performance was assessed using sensitivity, specificity, accuracy, and AUC.
The CMH group and the non-CMH group differed significantly in a number of factors, including age, the length of the disease, and pulmonary function indices, according to a comparison of baseline patient data. Analysis of airway microbiome characteristics revealed that the CMH group had significantly lower observed ASVs and Shannon indices (p<0.001), along with significant enrichment of potentially pathogenic bacterial genera such as Haemophilus and Pseudomonas. Following feature selection, disease duration, Haemophilus abundance, history of AECOPD, Pseudomonas abundance, and predicted FEV1% were identified as significant predictive factors. With a sensitivity of 0.867, specificity of 0.789, PPV of 0.805, NPV of 0.855, and AUC of 0.911, the Bayesian Network (BN) model outperformed the other six machine learning models on the testing sets; its generalization ability was significantly superior to other algorithms such as SVM and RF.
CMH in COPD is linked to airway dysbiosis and pathogen enrichment. The BN model effectively identifies this phenotype with strong generalization ability.