Background
Pertussis is a highly contagious respiratory illness that can cause severe complications, particularly in children. However, predicting which patients with pertussis are at risk of developing severe symptoms remains a challenge in clinical practice. In this study, we developed a two-step cascading machine learning pipeline to predict severe pertussis in patients with the disease.
Study Design
This work retrospectively involved both outpatients and inpatients with pertussis between March 2011 and December 2018 from Capital Institute of Pediatrics. We developed a two-step cascading prediction pipeline for severe pertussis in both outpatients and inpatients based on logistic regression models and nomograms. The pipeline selected potential severe pertussis patients with high sensitivity in the first step and predicted severe pertussis patients with high specificity in the second step using additional information. Further, the second step of pipeline was tested on a held-out test set of 17 inpatients with pertussis patients enrolled since December 2018.
Results
This work split the enrolled patients into training cohort and validation cohort in a 7:3 ratio. The first-step machine learning models included onset age (months) as a predictor and achieved a sensitivity of 0.923 (95% CI, 0.778-1.00) in the validation cohort. The second-step machine learning model consisted of five variables: white blood cell count, globulin usage, fever, paroxysmal cyanosis, and increased leukocyte levels. The validation cohort showed an accuracy of 0.797, AUC of 0.920 (95% CI, 0.832-1.000), sensitivity of 0.647 (95% CI, 0.413-0.827), and specificity of
0.846 (95% CI, 0.725-0.920). Nomograms were developed based on the machine learning models, and the calibration curves indicated a good fit.
Conclusion
This study developed a severe pertussis prediction pipeline using machine learning models, which demonstrated high sensitivity in the first step and high accuracy in the second step. The pipeline provides a useful screening and prediction workflow for severe pertussis patients, and the inclusion of nomograms enhances its practicality for clinical use. Overall, this pipeline has potential to improve the early identification and management of severe pertussis cases.