Daily symptom reports enable four-phase menstrual cycle tracking without wearables: Implications for equitable digital health

Technical Summary: Daily Symptom Reports Enable Four-Phase Menstrual Cycle Tracking Without Wearables

Problem Statement
While millions of individuals use digital applications to track menstrual cycles, it remains unverified whether self-reported daily symptoms contain sufficient physiological signal to infer cycle phases across different individuals. Current non-hormonal tracking methods, such as calendar-based approaches, exhibit low predictive reliability (e.g., ~21% accuracy for ovulation prediction). Conversely, while wearable sensors achieve high accuracy for binary fertile-window detection, their performance drops to 63–65% when extended to four-phase classification under rigorous leave-one-subject-out (LOSO) evaluation. Furthermore, wearable devices are often inaccessible, unaffordable, or impractical for many populations. The core challenge lies in the subjective, heterogeneous nature of self-report data, where inter-individual variability in baseline symptom expression and reporting behavior often obscures the underlying physiological signal. Additionally, standard classifiers struggle to enforce biologically plausible temporal constraints, such as cyclic phase ordering and realistic phase durations.

Methodology
The authors developed a hybrid machine learning architecture to classify menstrual cycles into four phases (Menstrual, Follicular, Fertile, and Luteal) using exclusively self-reported symptom data.

• Dataset: The study utilized the mcPHASES dataset, a publicly available longitudinal dataset containing daily self-reported symptoms and gold-standard hormonal validation (LH, estrone-3-glucuronide, and pregnanediol glucuronide) from 41 participants. The analysis focused on 14 daily variables: flow volume, flow color, and 12 symptom scales (e.g., fatigue, mood swings, cramps).
• Feature Engineering: From the 14 raw variables, 97 features were constructed, reduced to approximately 83 per fold after correlation filtering. These features were categorized into:
  • Cycle Position: Days-since-bleed counters, linear progress, and cyclic encoding.
  • Rolling Standard Deviations: Capturing within-person symptom variability over 5-, 7-, and 14-day windows.
  • Rolling Means: Representing medium- to long-term trends.
  • Raw Values: Same-day symptom levels.
  • Key Design Choice: The feature set prioritizes within-person variability (rolling standard deviations) over absolute symptom intensity.
• Model Architecture:
  1. Discriminative Stage: A CatBoost gradient-boosted classifier generates daily phase probability estimates from the engineered features. CatBoost was selected for its ability to handle missing values and reduce prediction shift on small sample sizes.
  2. Generative/Temporal Stage: A Hidden Semi-Markov Model (HSMM) overlays temporal structure on the CatBoost outputs. The HSMM enforces two biological constraints:
     • Strict Cyclic Ordering: Menstrual → Follicular → Fertile → Luteal → Menstrual.
     • Gaussian Duration Priors: Phase lengths are constrained to physiologically realistic ranges observed in the training data (e.g., preventing a 1-day Luteal phase).
  3. Optimization: The Viterbi algorithm jointly optimizes the state sequence to maximize the total score, combining discriminative emission scores (from CatBoost) with duration priors and transition constraints.
• Evaluation Protocol: The model was evaluated using Leave-One-Subject-Out (LOSO) cross-validation. For each of the 41 participants, the model was trained on the remaining 40 and tested on the held-out participant. This ensures the model learns generalizable phase signatures rather than memorizing subject-specific baselines.

Key Results

• Performance: The hybrid CatBoost+HSMM model achieved 67.6% accuracy and a macro F1 score of 0.662.
• Benchmarking: This performance outperformed all baselines, including:
  • A calendar-only model (63.0% accuracy).
  • A random forest using all features (65.3% accuracy).
  • CatBoost alone (65.1% accuracy).
  • Previously reported wearable-sensor benchmarks for four-phase classification (63–65% accuracy).
• Feature Importance: Analysis revealed that rolling standard deviations (capturing within-person symptom variability) accounted for 47.7% of global feature importance. This substantially outperformed raw symptom levels, which contributed minimally to the model's predictive power.
• Temporal Smoothing: The HSMM component provided a statistically significant improvement over CatBoost alone (+2.4 percentage points accuracy, +3.2 percentage points macro F1), with the most significant gains observed in the Fertile phase (F1 increased from 0.398 to 0.462).
• Phase-Specific Performance: The model performed best on the Menstrual (F1 0.774) and Luteal (F1 0.750) phases. The Fertile phase remained the most challenging (F1 0.462), with 57% of fertile days misclassified as Follicular or Luteal.

Significance and Claims
The paper claims three primary contributions:

1. Feasibility of Device-Free Monitoring: It establishes that self-reported symptom dynamics alone are sufficient for cross-subject four-phase classification, achieving accuracy levels that match or exceed comparable wearable-sensor benchmarks. This suggests that device-free reproductive health monitoring is technically viable for populations where wearables are inaccessible.
2. The Primacy of Variability: The study identifies within-person symptom variability (rather than absolute severity) as the dominant predictive signal. The authors argue that rolling standard deviations act as "reporting-invariant digital biomarkers" that generalize across individuals more effectively than raw intensity scores, as they are invariant to individual reporting baselines.
3. Reproducible Benchmark: The work provides a reproducible benchmark on a publicly available, hormonally annotated dataset, establishing a reference point for future multimodal systems and digital health research.

Limitations and Future Directions
The authors explicitly note several limitations:

• Fertile Phase Detection: Accuracy for the fertile phase remains insufficient for high-stakes clinical decisions (e.g., fertility treatment), likely due to the transient nature of the physiological signal.
• Flow Data Dependency: The model relies heavily on self-reported flow data for cycle anchoring; performance degrades significantly for users with sparse or absent flow records.
• Demographic Narrowness: The cohort was small (41 participants), young (mean age 20.9), and lacked diversity in terms of hormonal contraception use or conditions like PCOS.
• Label Uncertainty: Ground-truth labels derived from hormonal assays involve threshold decisions near biologically continuous transitions, placing a ceiling on achievable classification accuracy.

The authors conclude that while the current system is not a replacement for high-stakes clinical tools, it demonstrates that accurate cycle-phase inference does not require hardware. They suggest future work should focus on multimodal fusion (combining symptoms with passive sensors like skin temperature), personalization strategies, and designing reduced questionnaires that target high-importance variability features to reduce user burden.