Related Experiment Videos
Interpretable prediction model for severe community-acquired pneumonia in adults: machine learning development and
Hefei Zha1,2, Qian Shi1,2, Chunyan Liu1,2
1Department of Medical Laboratory, Hospital of Xinjiang Production and Construction Corps, Second Affiliated Hospital of Shihezi University School of Medicine, Urumqi, China.
Background:
Community-acquired pneumonia (CAP) can lead to severe community-acquired pneumonia (SCAP), which is associated with an increased risk of complications and mortality. However, the clinical utility of existing SCAP assessment tools is limited by their insufficient timeliness and objectivity. Therefore, the early identification of patients at high risk of progression to SCAP is crucial for improving their outcomes. Here, we aimed to develop and validate a model using readily available laboratory parameters to predict SCAP in individuals diagnosed with CAP.
Methods:
Patients with CAP were enrolled from the Xinjiang Production and Construction Corps Hospital. Two cohorts were established: training (310 SCAP and 741 non-SCAP patients) and temporal validation (72 SCAP and 337 non-SCAP patients). Following feature selection via Least Absolute Shrinkage and Selection Operator regression, we developed and evaluated six machine learning models and identified the optimal model. Model interpretability was achieved through the application of the SHapley Additive exPlanations (SHAP) framework. To evaluate its robustness, the developed model was tested on a temporal validation cohort.
Results:
In the training cohort, the Light Gradient Boosting Machine (LightGBM) achieved an area under the receiver operating characteristic curve of 0.893 (95% confidence interval [CI]: 0.870-0.915) and calibration curve Brier score of 0.135 (95%CI: 0.127-0.143), indicating excellent predictive performance. The SHAP analysis identified the following top predictors of SCAP onset, in descending order of importance: albumin, age, D-dimer, lactate dehydrogenase, C-reactive protein, oxygen saturation, neutrophil count, and blood urea nitrogen. The model's robustness and generalizability were successfully validated in the temporal validation cohort, achieving an area under the curve of 0.879 (95%CI: 0.836-0.922) and Brier score of 0.131 (95%CI: 0.109-0.155).
Conclusions:
A robust early prediction model for SCAP, leveraging routine laboratory parameters and employing the LightGBM algorithm, was developed. The model exhibited strong performance and holds promise as an effective clinical decision-support tool to assist in early intervention strategies for SCAP.