Self-Reported Symptoms Enable Four-Phase Menstrual Cycle Classification with Hormonally Validated Labels

This study demonstrates that a hybrid machine learning framework combining gradient boosting and Hidden Semi-Markov Models can accurately classify four menstrual cycle phases using only self-reported symptoms, achieving 67.6% accuracy and establishing symptom dynamics as a scalable, device-free digital biomarker for reproductive health.

The Gist

Imagine your body is like a four-season garden. Just as nature cycles through Spring, Summer, Autumn, and Winter, your menstrual cycle goes through four distinct biological phases: Menstrual, Follicular, Fertility (Ovulation), and Luteal.

For years, scientists have tried to predict which "season" you are in. Some use expensive, high-tech weather stations (wearable sensors like smartwatches) that measure your heart rate and temperature. Others use a simple calendar, guessing based on the last time it rained (your period).

This paper asks a bold question: Can we predict the season just by asking you how you feel?

Here is the story of how they tried to answer that, using a mix of "feeling" and "logic."

The Problem: The "Noisy" Diary

Imagine you ask 41 people to keep a daily diary of their symptoms: "How tired are you?" "Do you have a headache?" "How's your mood?"

The problem is that human diaries are messy.

• Subjectivity: One person's "mild headache" is another person's "severe migraine."
• Forgetfulness: People forget to write things down.
• Variability: Some people get cramps every month; others never do.

Because of this noise, scientists thought, "Maybe we can't trust these diaries to tell us exactly what's happening biologically." They thought the signal was too weak.

The Solution: A Two-Step Detective Team

The researchers built a smart computer system to read these diaries. They didn't just use one tool; they used a two-part detective team:

1. The "Feeling" Detective (CatBoost):
   This is a super-smart AI that looks at your daily symptoms. It's great at spotting patterns. If you say, "I'm tired, my breasts are sore, and I'm stressed," it guesses, "Ah, you're probably in the Luteal phase!"

   • But it has a flaw: It looks at each day in isolation. It might guess "Luteal" on Tuesday, then "Menstrual" on Wednesday, then "Luteal" again on Thursday. Biologically, that's impossible! You can't skip seasons.

2. The "Logic" Detective (HSMM):
   This is the rule-enforcer. It knows the laws of nature: You must go Menstrual → Follicular → Fertility → Luteal → Menstrual. It also knows how long each season usually lasts (e.g., the Fertility season is very short, like a brief summer storm, while the Luteal season is long).

   • What it does: It takes the "Feeling" Detective's guesses and smooths them out. If the first detective makes a weird jump, the Logic Detective says, "Wait, that doesn't make sense. Let's adjust the timeline to fit the rules of biology."

The Big Discovery: It's About the Change, Not the Level

Here is the most surprising part of the paper.

When the AI looked at the data, it didn't care much about how bad your symptoms were.

• Example: It didn't matter if you rated your cramps a "2" or a "6" on a scale of 10.
• What mattered: It cared about how much your symptoms changed from day to day.

The Analogy: Imagine driving a car.

• Absolute Level: Driving at a steady 60 mph.
• Variability: Suddenly hitting the brakes, then accelerating, then swerving.

The AI realized that the "seasons" of your cycle are defined by swings and shifts, not by how "bad" you feel. If your mood or energy starts fluctuating wildly, that's a stronger signal that a phase change is coming than just feeling "very tired" consistently.

The Results: A Surprising Success

They tested this system on 41 women, using a strict rule: The AI had to predict the cycle for a woman it had never seen before. (This is like a chef trying to cook a perfect meal for a stranger without tasting their food first).

• The Score: The system got it right 67.6% of the time.
• The Comparison:
  • A simple calendar guess (just counting days) got about 63%.
  • A random guess got about 35%.
  • This "Diary + Logic" system beat the calendar and came close to the accuracy of expensive wearable sensors (which usually get around 63-65% for this specific 4-phase task).

Why This Matters

1. No Gadgets Needed: You don't need to buy a $300 smart ring or wear a patch. You just need a phone app to log how you feel. This makes cycle tracking accessible to anyone, anywhere, even in places without high-tech devices.
2. Better Questions: Since the AI cares about changes in symptoms, future apps might ask, "How is your energy today compared to yesterday?" rather than just "How tired are you?"
3. The Weak Spot: The system still struggles with the Fertility Phase (the few days you can get pregnant). It's like trying to spot a fleeting cloud in a blue sky; it's very short and hard to catch. The AI got this right less than half the time.

The Bottom Line

This paper proves that your subjective feelings contain a hidden biological map. Even though our diaries are messy and personal, if we look at the patterns of change rather than just the raw numbers, and we apply some biological common sense, we can accurately predict where you are in your cycle.

It's a step toward a future where your phone knows your body's rhythm just as well as your smartwatch does, but without needing to wear anything at all.

Technical Summary

1. Problem Statement

The menstrual cycle is governed by the hypothalamic-pituitary-ovarian axis, creating four distinct physiological phases: Menstrual, Follicular, Fertility (Ovulatory), and Luteal. Accurate tracking of these phases is critical for fertility planning, contraception, and managing cycle-related conditions.

Current approaches face significant limitations:

• Calendar-based methods: Rely solely on cycle length history, achieving low accuracy (~21% for ovulation prediction) and failing to account for individual variability.
• Wearable sensors: Passive sensors (temperature, heart rate) achieve high accuracy (88–90%) but require specific hardware, creating barriers for resource-constrained populations.
• Self-reported symptoms: While widely collected via apps, it remains unproven whether subjective symptom data alone contains sufficient signal to distinguish all four phases prospectively across different individuals (cross-subject generalization), especially given the noise, recall bias, and high inter-individual variability of symptom reporting.

The Core Challenge: Can a machine learning model infer four distinct menstrual phases using only self-reported symptom data, validated against hormonal ground truth, without relying on wearable sensors or calendar assumptions?

2. Methodology

Dataset

• Source: The mcPHASES dataset, a longitudinal study of 42 participants (aged 18–29) in Toronto, Canada.
• Ground Truth: Phases were labeled based on daily hormonal assays (LH, Estradiol, Progesterone) via urine analysis.
• Input Data: Daily self-reports of 14 variables: flow volume (8-point scale), flow color (9 categories), and 12 symptom scales (e.g., fatigue, mood swings, cramps) rated on a 6-point Likert scale.
• Preprocessing:
  • Excluded 2024 data due to missing symptom entries.
  • Removed cycles with ≥5 consecutive days of missing data to ensure rolling statistics were reliable.
  • Final cohort: 41 participants with 3,557 participant-days (2,983 after a 14-day warmup period).

Feature Engineering

The authors constructed 97 features (reduced to ~83 after correlation filtering) from the 14 raw variables, categorized into:

1. Cycle Position (13 features): Days-since-bleed counters with varying thresholds, linear progress, and cyclic encoding ($\sin, \cos$) to handle boundary continuity.
2. Rolling Standard Deviations (42 features): Variability of symptoms over 5-, 7-, and 14-day windows. Hypothesis: Fluctuation is more informative than absolute levels.
3. Rolling Means (28 features): Medium-to-long-term trends.
4. Raw Values (14 features): Same-day symptom levels.

Model Architecture: Hybrid Discriminative-Generative

The proposed framework combines a gradient-boosted classifier with a temporal model:

1. Stage 1: CatBoost Classifier

   • A gradient-boosted decision tree algorithm selected for its robustness to missing values and ability to handle non-linear feature interactions.
   • Outputs per-day probabilities for the four phases.
   • Limitation: Treats days independently; cannot enforce biological constraints (e.g., preventing a Luteal phase from immediately following a Menstrual phase without a Follicular phase).

2. Stage 2: Hidden Semi-Markov Model (HSMM)

   • Acts as a temporal regularizer over the CatBoost outputs.
   • Discriminative Emissions: Converts CatBoost probabilities into log-emission scores.
   • Duration Priors: Uses empirical Gaussian distributions for phase lengths (derived from training data) to enforce physiologically plausible durations (e.g., preventing a 1-day Luteal phase).
   • Transition Constraints: Enforces strict cyclic ordering: Menstrual $\to$ Follicular $\to$ Fertility $\to$ Luteal $\to$ Menstrual.
   • Decoding: Uses the Viterbi algorithm to find the most probable state sequence.

Evaluation Protocol

• Leave-One-Subject-Out (LOSO) Cross-Validation: The model is trained on 40 participants and tested on the held-out 41st. This is the strictest test for generalizability, ensuring the model learns phase signatures rather than individual reporting styles.
• Baselines: Compared against Majority Class, Bleed-only Logistic Regression, Calendar-only Logistic Regression, standard Logistic Regression, Random Forest, and CatBoost alone.

3. Key Results

• Overall Performance: The hybrid CatBoost+HSMM model achieved 67.6% accuracy and a Macro F1 score of 0.662.
  • This outperformed all baselines, including CatBoost alone (65.1% accuracy) and Random Forest (65.3%).
  • The HSMM temporal smoothing provided a statistically significant improvement (+2.4 percentage points accuracy, +3.2 F1 points).
• Per-Phase Performance:
  • Menstrual: High performance (F1 0.774) due to clear flow signals.
  • Luteal: Strong performance (F1 0.750).
  • Follicular: Moderate performance (F1 0.661).
  • Fertility: Lowest performance (F1 0.462). This phase is difficult due to its short duration (~3 days) and subtle symptom shifts. 57% of Fertility days were misclassified as Follicular or Luteal.
• Feature Importance (SHAP Analysis):
  • Rolling Standard Deviations were the dominant predictive signal, accounting for ~45% of feature importance.
  • Cycle Position features were the second most important.
  • Raw symptom levels had the least impact.
  • Conclusion: Within-person symptom variability (fluctuation) is a more generalizable signal than absolute symptom intensity.

4. Key Contributions

1. First Cross-Subject Benchmark: To the authors' knowledge, this is the first study to achieve prospective, cross-subject four-phase classification using only self-reported data with hormonally validated ground truth.
2. Hybrid Architecture: Demonstrated that combining a discriminative classifier (CatBoost) with a generative temporal model (HSMM) effectively enforces biological constraints, improving robustness over standard classifiers.
3. Discovery of Variability Signal: Established that symptom variability (captured by rolling standard deviations) is a stronger predictor of cycle phase than symptom intensity. This suggests that the rate of change in symptoms, rather than their absolute level, carries the physiological signal.
4. Low-Burden Feasibility: Proved that device-free, low-cost monitoring is feasible, offering a pathway for menstrual health tracking in populations underserved by wearable technology.

5. Significance and Implications

• Scientific Validation: Confirms that subjective symptom reports contain sufficient physiological signal to distinguish complex cycle phases, challenging the notion that they are merely "noisy" or uninformative.
• Clinical & Public Health: Enables scalable digital biomarkers for reproductive health in resource-constrained settings where wearables are unavailable.
• Future System Design: Suggests that future symptom-based apps should focus on capturing fluctuation (e.g., asking "how has your mood changed since yesterday?" rather than just "how is your mood?") and can reduce reporting burden by focusing on variability rather than absolute levels.
• Limitations & Future Work:
  • Fertility Phase: The model struggles with the Fertility phase (F1 0.462), limiting its immediate utility for fertility tracking.
  • Flow Dependence: Performance degrades significantly for users who do not consistently report flow data, as cycle position features are critical.
  • Demographics: The study cohort was young (18–29) and relatively homogenous; validation in diverse age groups and those with conditions like PCOS is needed.
  • Next Steps: The authors propose multimodal fusion (combining self-reports with passive wearable data like temperature) to bridge the gap in Fertility phase detection.