Platelet function tests (PFTs) are essential for predicting bleeding tendencies and assessing the effectiveness of antiplatelet agents. The VerifyNow System is a quick and simple PFT, but warning messages (WMs) due to biological factors may reduce its effectiveness. This study aimed to quantify the frequency of WMs and evaluate the performance of machine learning (ML) models for predicting these WMs. Data were retrospectively collected from patients who underwent VeryNow System testing from October 2019 to April 2023. The patients were classified into WM-positive (WM_POS) and WM-negative (WM_NEG) groups. Significant variables between two groups were selected for feature analysis. Prediction models were developed using XGBoost, random forest (RF), support vector machine, and logistic regression algorithms. A receiver operating characteristic (ROC) curve analysis with five-fold cross-validation was performed using Python (Python Software Foundation, USA). A total of 6,998 data were collected from 6,438 patients, with 0.8% (55/6,998) classified as WM_POS. Significant differences were observed in sex, alanine transaminase levels, alkaline phosphatase levels, total bilirubin levels, creatinine levels, prothrombin time, white blood cell count, and hematocrit count between the two groups. The area under the ROC of the four models for predicting the WM_POS showed excellent or good (0.8–1.0) performance. Both XGBoost and RF models achieved accuracy, precision, recall, and F1 scores exceeding 0.99. Machine learning models were used to predict the WMs of PFT and showed good performance, potentially enhancing the efficiency of PFT. However, further research is needed to apply ML in clinical laboratories.

Keywords: Platelet function tests, VerifyNow, Machine learning, Performance