A recurrence model for non-puerperal mastitis patients based on machine learning.

Objective: Non-puerperal mastitis (NPM) is an inflammatory breast disease affecting women during non-lactation periods, and it is prone to relapse after being cured. Accurate prediction of its recurrence is crucial for personalized adjuvant therapy, and pathological examination is the primary basis for the classification, diagnosis, and confirmation of non-puerperal mastitis. Currently, there is a lack of recurrence models for non-puerperal mastitis. The aim of this research is to create and validate a recurrence model using machine learning for patients with non-puerperal mastitis.

Methods: We retrospectively collected laboratory data from 120 NPM patients, dividing them into a non-recurrence group (n = 59) and a recurrence group (n = 61). Through random allocation, these individuals were split into a training cohort and a testing cohort in a 90%:10% ratio for the purpose of building the model. Additionally, data from 25 NPM patients from another center were collected to serve as an external validation cohort for the model. Univariate analysis was used to examine differential indicators, and variable selection was conducted through LASSO regression. A combination of four machine learning algorithms (XGBoost、Logistic Regression、Random Forest、AdaBoost) was employed to predict NPM recurrence, and the model with the highest Area Under the Curve (AUC) in the test set was selected as the best model. The finally selected model was interpreted and evaluated using Receiver Operating Characteristic (ROC) curves, calibration curves, Decision curve analysis (DCA), and Shapley Additive Explanations (SHAP) plots.

Results: The logistic regression model emerged as the optimal model for predicting recurrence of NPM with machine learning, primarily utilizing three variables: FIB, bacterial infection, and CD4+ T cell count. The model showed an AUC of 0.846 in the training cohort and 0.833 in the testing cohort. The calibration curve indicated excellent calibration of the model. DCA revealed that the model possessed favorable clinical utility. Furthermore, the model effectively achieved in the external validation group, with an AUC of 0.825.

Conclusion: The machine learning model developed in this study, serving as an effective tool for predicting NPM recurrence, aids doctors in making more individualized treatment decisions, thereby enhancing therapeutic efficacy and reducing the risk of recurrence.