Machine Learning Analysis Was Conducted to Construct a Novel Model for Modulating the Clinical Manifestations and Predicting the Response to Glucocorticoids in Immune Thrombocytopenia

Abstract:

Aim: In this study, a novel prediction model was established using Recursive Feature Elimination (RFE) to predict the clinical manifestations of Immune Thrombocytopenia (ITP) and the response to glucocorticoids.

Methods: This study consisted of two parts: the establishment and validation of the prediction model. A total of 1019 ITP patients were included, and general information such as age, gender, medical history, and laboratory test results were recorded. RFE, SVM, and Logistic regression models were compared by ROC analysis, and the raw data were preprocessed for feature extraction. The RFE was then used to construct the prediction model, and patients were classified into high-risk group (n=464) and low-risk group (n=555) based on their risk levels. The model performance was evaluated using cross-validation, and the results were subsequently predicted.

Results: The results of this study showed that the RFE prediction model established had good predictive efficacy for the clinical manifestations of ITP patients. A total of 39 variables were extracted, and univariate analysis revealed that age, B lymphocyte surface antigen (CD19B), mean corpuscular volume (MCV), and hemoglobin (Hb) were the main influencing variables (P<0.05). After comparing the models, the ROC of RFE was 0.902 (0.884-0.920), SVM was 0.511 (0.501-0.522), and Logistic regression was 0.680 (0.646-0.715). Since RFE had the highest ROC, it was selected as the prediction model for this study. In the training cohort, the RFE model had a specificity of 0.789 and a sensitivity of 0.858. The threshold for high and low-risk classification was set at 0.33. In the prediction of clinical and laboratory differences, the threshold of the RFE model cutoff value was 16.5, with a specificity of 0.43 and a sensitivity of 0.70. In the validation cohort, the RFE model had a ROC of 0.962, which was higher than the other two models.

Conclusion: In this study, we successfully constructed a new model using machine learning techniques that can analyze and predict the clinical manifestations of patients with immune thrombocytopenia (ITP), as well as predict their response to glucocorticoid treatment. This model holds promise in providing more accurate treatment recommendations to clinical physicians and aiding in the development of individualized treatment strategies, thereby improving the treatment outcomes and quality of life for ITP patients. Further research can be conducted to expand the sample size and integrate other clinical data to further optimize and validate the performance of this model.

Keywords: Machine learning; Predictive model; Immune thrombocytopenia; Clinical manifestations; Glucocorticoids therapy