# Personalized Low-Invasive Approach to Chronic Endometritis Evaluation in Premenopausal Women: Machine Learning-Based Modeling

## TL;DR

This study uses machine learning to create models that predict chronic endometritis in premenopausal women using low-invasive markers, aiming to improve diagnosis and treatment.

## Contribution

The novel contribution is the development of machine learning-based models for CE prediction using low-invasive clinical and laboratory features.

## Key findings

- Five CE prediction models were developed with AUCs ranging from 0.673 to 0.769.
- Models 2 and 5 showed the best recall and precision, requiring only 5–7 risk markers.
- Serum adiponectin and SHBG were identified as key predictors of CE risk.

## Abstract

Background/Objectives: Chronic endometritis (CE) is a well-known risk factor for recurrent implantation failure. However, the traditional approach to CE diagnosis has several drawbacks. On the other hand, there is a lot of evidence that some clinical, instrumental, and/or laboratory parameters of patients are associated with CE. The aim of this study is to build a CE prediction model using machine learning tools based on low-invasive pathological features. Methods: The data of 108 women (44 with and 64 without CE) from a multicenter perspective cross-sectional study was included in this study. Basic characteristics, reproductive history, laboratory and ultrasound indicators, and immunohistochemistry results were collected. Binary feature selection was performed using forward stepwise selection with logistic regression as the evaluation criterion. For each feature configuration, a gradient-boosting model was trained on decision trees with a binary logistic loss function. The models were evaluated and compared on test data using standard metrics. Results: We built five comparable predictive models for CE. The models yielded the following AUCs (95% CI): Model 1 (seven indicators)—0.704 (0.5170, 0.8907), Model 2 (seven indicators)—0.673 (0.4716, 0.8745), Model 3 (nine indicators)—0.677 (0.4916, 0.8622), Model 4 (five indicators)—0.758 (0.5913, 0.9241), and Model 5 (five indicators)—0.769 (0.5913, 0.9241). Models 2 and 5 have the better recall and precision values, but the differences were not significant. SHAP values indicated that serum adiponectin level (Model 2) and SHBG (Model 5) had the greatest association with CE risks. Conclusions: Models 2 and 5 show the most promising potential for clinical application, as they demonstrate superior recall and precision metrics and require assessment of only 5–7 risk markers (with only a few being non-routine) for their implementation.