To construct a prediction model of platelet concentration in platelet-rich plasma based on machine learning

Abstract Objective The insufficient accuracy of Platelet concentration prediction in the preparation of platelet-rich plasma (PRP) is a challenge in clinical application. Existing methods mainly rely on a single preapheresis platelet count and ignore other potential influencing factors. The purpose of this study is to build a multi-factor prediction model by machine learning technology to improve the prediction accuracy of platelet concentration of PRP and provide data support for the optimization of PRP preparation process. Methods A retrospective study design was used, and the data of 78 patients who received PRP treatment in a hospital from January to August 2022 were included. The min-max normalization method was used to preprocess the data, and the recursive feature elimination method was used to screen features. Multiple regression models such as linear regression, ridge regression, Lasso regression, elastic network, random forest and XGBoost were constructed and compared. Five-fold cross validation was used to evaluate the performance of the model, and SHAP value analysis was used to explain the prediction mechanism of the model. Results The data of 78 patients were included. The features of Pre-collection platelets (prePLT), width of platelet volume distribution (The width of platelet volume distribution, PDW), The yield of large platelets (pLCR), mean corpuscular volume (MCV) and mean corpuscular hemoglobin (MCH) were important features. In the model training, random forest and XGBoost performed well on the training set, and the linear regression generalization performance on the test set was the best (R² = 0.548), and the prediction error was stable (MSE=0.025 in the training set and MSE=0.022 in the test set). SHAP value analysis showed that among the hematological parameters, PDW and pLCR had a significant impact on the prediction results, and their SHAP values ranged from -0.2 to 0.4 and -0.4 to 0.2, respectively. prePLT also contributed significantly to the prediction results, which was consistent with clinical experience. Conclusions This study successfully constructed the PRP platelet concentration prediction model based on machine learning, which not only verified the importance of prePLT, but also found that platelet parameters such as PDW and pLCR had significant contributions to the prediction results. The model improves the prediction accuracy by integrating multi-dimensional hematological indicators, which provides a theoretical basis for the standardization and individualization of PRP preparation process. In clinical application, this model can be combined with traditional experience to further improve the quality of PRP preparation and the reliability of efficacy.