Machine Learning-Based Classification and Symptom Forecasting Across the Menopausal Transition Using Longitudinal Self-Reported Data
Aadyaa Vijay* and Anu Prabha
ABSTRACT
Background: Menopause is a universal and natural biologically driven life transition that is highly heterogeneous in terms of symptom variability, its manifestation timing and extent of severity. Scalable and deployable models that utilize self-reported data to predict menopausal stage and short-term occurrence of symptoms could potentially enable earlier medical intervention and personalized treatment.
Objective: In this study we have utilized publicly available data from the Study of Women’s Health Across the Nation (SWAN); created and validated a scalable machine learning pipeline that does the following: (a) predicts menopausal stage (pre, peri, post) from self-reportable longitudinal variables and (b) generates probabilistic short-term symptom forecasts for hot flashes and mood changes based on self-reported last menstrual period (LMP) timing.
Methods: Rule-based feature selection (regular expression heuristics) was used to extract self-report variables from the SWAN dataset. Thirty-eight interpretable features were selected for this study that are self-reportable covering demographics, lifestyle, menstrual timing, vasomotor symptoms, mood symptoms, pain and general health indicators. Median imputation and one-hot encoding were implemented in reusable scikit-learn pipelines. Numerical features were standardized as needed. Two classifiers were used: Multinomial Logistic Regression and Random Forest (200 trees). These were trained on a stratified 80/20 split of the dataset and a five-fold cross validation was performed. Symptom forecasting is conducted via a rule-based Symptom CycleForecaster model that calculates cycle day from LMP and returns symptom probabilities.
