Machine learning for classification of uterine activity outside pregnancy

Area of Science:

• Biomedical Engineering
• Machine Learning
• Women's Health

Background:

• Uterine activity varies significantly throughout the menstrual cycle.
• Noninvasive methods for monitoring uterine function are crucial for reproductive health.
• Accurate discrimination of uterine phases can aid in diagnosing gynecological conditions.

Purpose of the Study:

• To evaluate machine learning classifiers for differentiating uterine activity during the four menstrual cycle phases.
• To compare the performance of Support Vector Machine (SVM), K-nearest neighbors (KNN), Gaussian mixture model (GMM), and Naïve Bayes classifiers.
• To identify the most effective classification method for uterine activity analysis.

Main Methods:

• Extracted amplitude- and frequency-based features from electrohysterography and ultrasound speckle tracking signals.
• Trained four machine learning classifiers (SVM, KNN, GMM, Naïve Bayes) using all feature combinations.
• Validated the classifiers on a database of 24 measurements from different menstrual cycle phases.

Main Results:

• The Support Vector Machine (SVM) classifier demonstrated superior performance in discriminating menstrual phases.
• SVM achieved a classification accuracy of 90%, with 79% sensitivity and 93% specificity.
• Other classifiers showed varying degrees of success but were outperformed by SVM.

Conclusions:

• Machine learning, particularly SVM, is effective for classifying uterine activity during the menstrual cycle.
• This approach holds promise for future development of diagnostic tools for infertility and uterine diseases like endometriosis.
• Noninvasive signal analysis combined with machine learning offers a viable pathway for objective assessment of uterine function.