Interpretable machine-learning model for Predicting the Convalescent COVID-19 patients with pulmonary diffusing...

Area of Science:

• Pulmonary Medicine
• Medical Informatics
• Machine Learning in Healthcare

Background:

• COVID-19 survivors often experience pulmonary diffusing capacity impairment (PDCI).
• Predicting PDCI is crucial for assessing long-term pulmonary function in COVID-19 survivors.
• Current predictive models for PDCI in this population are limited.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting PDCI in COVID-19 convalescent patients.
• To utilize routinely available clinical data for PDCI prediction.
• To aid in the clinical diagnosis and management of post-COVID-19 pulmonary complications.

Main Methods:

• A cohort of 221 COVID-19 survivors was studied 18 months post-discharge.
• Data were randomly split into training (80%) and validation (20%) sets.
• Six ML models were evaluated, with feature selection and data balancing techniques employed.

Main Results:

• The XGBoost model demonstrated optimal performance with an AUC of 0.755 and 78.01% accuracy.
• Hemoglobin (Hb), maximal voluntary ventilation (MVV), illness severity, platelet count (PLT), uric acid (UA), and blood urea nitrogen (BUN) were identified as key predictors.
• SHAP analysis highlighted Hb and MVV as the most influential factors.

Conclusions:

• The developed XGBoost model shows strong prognostic capability for PDCI in COVID-19 survivors.
• Clinical factors, particularly Hb and MVV, are significant predictors of long-term pulmonary function after COVID-19.
• This ML approach can assist clinicians in identifying survivors at risk for PDCI.