Machine Learning Models for Prediction of Severe Pneumocystis carinii Pneumonia after Kidney Transplantation: A Single-Center Retrospective Study

Abstract

Background: The objective of this study was to formulate and validate a prognostic model for postoperative severe Pneumocystis carinii pneumonia (SPCP) in kidney transplant recipients utilizing machine learning algorithms, and to compare the performance of various models. Methods: Clinical manifestations and laboratory test results upon admission were gathered as variables for 88 patients who experienced PCP following kidney transplantation. The most discriminative variables were identified, and subsequently, Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), K-Nearest Neighbor (KNN), Light Gradient Boosting Machine (LGBM), and eXtreme Gradient Boosting (XGB) models were constructed. Finally, the models’ predictive capabilities were assessed through ROC curves, sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and F1-scores. The Shapley additive explanations (SHAP) algorithm was employed to elucidate the contributions of the most effective model’s variables. Results: Through lasso regression, five features—hemoglobin (Hb), Procalcitonin (PCT), C-reactive protein (CRP), progressive dyspnea, and Albumin (ALB)—were identified, and six machine learning models were developed using these variables after evaluating their correlation and multicollinearity. In the validation cohort, the RF model demonstrated the highest AUC (0.920 (0.810–1.000), F1-Score (0.8), accuracy (0.885), sensitivity (0.818), PPV (0.667), and NPV (0.913) among the six models, while the XGB and KNN models exhibited the highest specificity (0.909) among the six models. Notably, CRP exerted a significant influence on the models, as revealed by SHAP and feature importance rankings. Conclusions: Machine learning algorithms offer a viable approach for constructing prognostic models to predict the development of severe disease following PCP in kidney transplant recipients, with potential practical applications.

Discriminating insulin resistance in middle-aged nondiabetic women using machine learning approaches

“…Previous studies used different numbers of top features, ranging from 5 to 20, to predict diseases of interest [12] , [35] . However, we observed that when employing the top five predictors to discriminate IR, the models' performance metrics did not significantly decline compared to the 48 predictors.…”

“…The training dataset was divided into five groups for the 5-fold cross-validation, with one group as the internal validation set and four as the internal training dataset. The average performance was computed using a grid search, and the hyperparameters were optimized to maximize the AUC of receiver operating characteristic (ROC) for the internal validation set [35] . Following the completion of model training, we utilized the testing dataset for validation.…”

“…The F1 score is a metric to quantify accuracy by precision and sensitivity [37] . Finally, we used the SHapley Additive exPlanations (SHAP) framework to illustrate the predictors' contributions and identify the threshold values of the predictors [38] .…”