Construction and validation of an intelligent pertussis identification model based on routine blood parameters and machine learning algorithms

Abstract Background Pertussis is a highly contagious respiratory disease whose early symptoms often resemble those of common respiratory tract infections, posing challenges for timely diagnosis. This study aimed to develop machine learning models based on routine blood parameters and evaluate their discrimination between pertussis-positive and symptomatic pertussis-negative children, with an additional exploratory comparison against healthy controls. Methods A total of 1132 paediatric cases were retrospectively collected from The Affiliated Dazu’s Hospital of Chongqing Medical University, including children with positive pertussis nucleic acid test results, symptomatic but pertussis-negative cases, and healthy controls. All cases had complete routine blood test results, covering 24 routine blood parameters. Nine machine learning (ML) algorithms were evaluated in two separate binary classification tasks: pertussis-positive versus symptomatic pertussis-negative children, and pertussis-positive children versus healthy controls. Performance was assessed using seven assessment metrics, including accuracy, sensitivity, specificity, F1 score, and area under the receiver operating characteristic curve (AUC). A separate temporal cohort was further used for calibration assessment. The selected models were incorporated into a prototype webpage. Results On the held-out test set within the modelling subset, XGBoost achieved an accuracy of 0.873 (95% CI: 0.836–0.909) and an F1 score of 0.867 (95% CI: 0.828–0.913) for distinguishing pertussis-positive from symptomatic pertussis-negative cases. For the exploratory positive-versus-healthy comparison, the CatBoost model achieved an accuracy of 0.938 (95% CI: 0.918–0.964), an F1 score of 0.932 (95% CI: 0.908–0.962), and an AUC value of 0.988 (95% CI: 0.975-1.000). SHAP analysis identified that Baso%, Mono#, and CRP exerted the most significant influence as the leading features in the CatBoost positive-versus-healthy model, and Eos#, Mono%, and PCT (plateletcrit) as the leading features in the XGBoost positive-versus-negative model. Conclusions Machine learning models based on routine blood parameters demonstrated the ability to discriminate between pertussis-positive and symptomatic pertussis-negative children. The additional positive-versus-healthy comparison provided exploratory evidence of discrimination between pertussis cases and healthy controls. These findings highlight the potential value of routine blood parameters for machine learning-assisted pertussis identification.