A Risk Prediction Model Based on Combined CBC and CPD Parameters for Triage of Non-Hodgkin Lymphoma Inpatients

This study aimed to construct and validate a machine learning triage model based on combined routine complete blood count (CBC) and cell population data (CPD) parameters, for stratifying non-Hodgkin lymphoma (NHL) risk among adult hematology inpatients at first admission.

A total of 339 NHL inpatients and 417 hematology inpatients with other hematological disorders (negative controls, NC) were retrospectively enrolled, split into training cohort (n=510) and temporal internal validation cohort (n=246). All laboratory data were routine blood test results obtained at the time of first hospital admission. Univariate analysis was performed to screen differential CBC and CPD parameters, followed by LASSO regression feature selection conducted exclusively on the training cohort to avoid feature selection leakage. Finally, 8 indicators were retained for modeling: Age, HGB, MONO, NEUT%, BASO%, NE-SSC, LY-WX, and MO-WZ. Five machine learning algorithms (logistic regression [LR], naive Bayes [NB], support vector machine [SVM], neural network [NNET], k-nearest neighbor [KNN]) were established. Model performance was evaluated via ROC curves, calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) analysis was applied to interpret feature importance of the optimal SVM model.

Sixteen differential parameters between NHL and NC groups were identified, and 8 independent predictive variables were retained after LASSO regularization. The SVM model constructed with the 8 combined CBC and CPD indicators demonstrated the optimal predictive efficacy and robustness across both cohorts, with AUC=0.894 (training, 95%CI: 0.858-0.923) and AUC=0.803 (validation, 95%CI: 0.748-0.850). SHAP analysis verified that three core CPD parameters (MO-WZ, NE-SSC, LY-WX) all ranked among the top six most influential predictors, providing independent predictive contribution beyond conventional CBC indicators. DCA demonstrated theoretical net clinical benefit for risk threshold probabilities between 0.1-0.8.

The SVM triage model integrating routine CBC and CPD parameters from first-admission blood tests achieves reliable risk stratification for NHL among hematology inpatients. The combined application of conventional blood count indicators and cell population morphological parameters can provide effective reference for inpatient pre-biopsy triage. Further multi-center prospective external validation is required to improve model generalizability.