Predicting the efficacy of recombinant human thrombopoietin in treating cancer therapy-related thrombocytopenia:...

Cancer treatment-induced thrombocytopenia (CTIT) is a common adverse effect of cancer therapy. CTIT increases the risk of bleeding, prolongs hospital stays, raises medical costs, and can negatively impact anti-tumor treatment outcomes, potentially leading to patient death. Therefore, it is crucial to initiate platelet-boosting therapy in a timely manner based on the individual circumstances of patients experiencing CTIT.

Patients who developed cancer treatment-induced thrombocytopenia and received Rh-TPO treatment from January 2023 and December 2023 were obtained for establishing the dataset. With absolute platelets increase as the outcome variable, univariate analysis was performed to screen out statistically significant factors, and 18 clinical variables were selected as initial features. The least absolute shrinkage and selection operator (LASSO) regression analysis was then used to identify the most important features. Based on this, a stacking ensemble model was constructed using cross-validation with out-of-fold predictions to prevent information leakage, and the predictive performance of the model was evaluated. Finally, the SHapley Additive exPlanations (SHAP) algorithm was used to explain the model, and a visual analysis of the features was conducted.

A total of 400 inpatients who developed cancer treatment-induced thrombocytopenia and received Rh-TPO treatment were included, of which 280 inpatients were assigned to the training set and 1,20 to the testing set. After LASSO regression screening, the study identified 7 key features: ethnicity, height, baseline serum creatinine, pre-chemotherapy platelet count, follow-up days after chemotherapy, platelet count before Rh-TPO, and duration of Rh-TPO treatment. We compared the performance of different machine learning models and selected the best three models as base models. Combined with Linear Regression as the meta-learner, we built a stacking ensemble model using 3-fold cross-validation with out-of-fold predictions. The stacking ensemble model showed best prediction ability compared to independent models with R² of 0.77 (training) and 0.74 (testing), MAE of 6.39 (training) and 8.34 (testing), MSE of 62.17 (training) and 97.70 (testing), RMSE of 7.88 (training) and 9.88 (testing), and MAPE of 0.09 (training) and 0.12 (testing). SHAP analysis showed that pre-chemotherapy PLT value and the follow-up days after chemotherapy were the important features affecting the prediction results.

The predictive model developed in this study could be beneficial for accurately predicting the improvement in platelet count in patients with CTIT who use Rh-TPO, facilitating timely assistance for patients in avoiding the risks caused by a drop in platelet count in a timely manner.