A machine learning (ML)-based model may aid in-hospital community-acquired pneumonia (CAP) mortality prediction, according to study findings published in Respiratory Medicine.
Researchers conducted a retrospective chart review to evaluate adherence to the 2019 American Thoracic Society/Infectious Diseases Society of America (ATS/IDSA) CAP guidelines; the researchers also developed and tested a machine learning-based CAP mortality prediction model, PneumoMLPred.
The study enrolled inpatients older than 13 years of age with CAP between 2016 and 2021 from a tertiary training hospital in the United Arab Emirates. Data were obtained on the patients’ demographics, comorbidities, preventive practices and counselling, vitals and measurements, laboratory results, blood gases, imaging results, blood cultures, overall patient outcome, and other variables.
PneumoMLPred, an explainable ML-based pneumonia prognostication model, has shown excellent accuracy and specificity as well as very good discriminatory ability for mortality prediction.
Imputation was performed with means for continuous variables and mode for categorical ones for the machine learning modeling. Base classifiers and ensemble techniques were used to identify which model best represented the underlying data and predicted in-hospital mortality.
The analysis included 783 pneumonia cases (64% male; 37% involving individuals aged ≥65 years; 16% involving current smokers). About 97% of participants initiated broad-spectrum antibiotics, 15.24% switched antibiotics during the hospital course, and 18% transitioned to an oral type.
Dosages were a major deterrent to guideline adherence, as 30% of fluoroquinolones and 46% of antipseudomonal drugs were prescribed at the correct doses. About 21% of all pneumonia management plans met the criteria for nonsevere CAP and 16% met severe CAP recommendations. Guideline adherence was not significantly associated with a decrease for in-hospital mortality (P >.05).
During the hospital course, 16% (n=123/781) of patients died. Pneumonia Severity Index (PSI) and C-reactive protein values were the most important variables for mortality prediction and were followed by glucose, sodium, and urea levels. Performance of the Random Forest model was best with 32 features, with a maximum area under the receiver operating curve (AUROC) of 0.864. Among the models, support vector classifier had the best performance (AUROC, 0.87, although it was nonsuperior to other ensemble methods).
CAP mortality was predicted by increased PSI, C-reactive protein levels, respiratory rates, sodium values, diastolic values, kidney disease, intensive care unit admission, vasopressor and/or ventilator use, confusion, and South Asian ethnicity. Increased systolic blood pressure, creatinine values, SpO2, glucose and hematocrit levels, and CURB-65 score of 1 favored CAP survival.
In addition to the single-center design, limitations included lack of assessment of the duration of antibiotic therapy and that pneumonia severity was not assessed per the ATS/IDSA guidelines. Also, no external validation was conducted, and no decision curve analysis was performed to help determine the model’s clinical use.
“PneumoMLPred, an explainable ML-based pneumonia prognostication model, has shown excellent accuracy and specificity as well as very good discriminatory ability for mortality prediction, pointing towards a potential role in assisting with CAP management locally,” the study authors concluded.
This article originally appeared on Pulmonology Advisor
Leave a comment