Option A: build a transparent baseline first
Use self-tracked or history-based data, build calendar / history baselines, and make the evaluation protocol unusually clear. This is often more defensible than a weak wearable model on noisy labels.
Data & Code
Data & Code
This page is not a flat resource dump. It helps you quickly decide whether you need public datasets, starter code, a baseline route, or a reproducibility checklist right now.
What do you need most right now?
I need public data
Start with the dataset chooser, then decide whether you need a tabular baseline, a wearable proof of concept, or a historical cycle dataset.
I need starter code
Go straight to the local Kaggle case study and the mcPHASES processing workflow to find an executable starting point.
I need a baseline route
Use the four realistic starter strategies to decide whether your best next step is a transparent baseline, a wearable proof of concept, or an evaluation paper.
I need a reproducibility checklist
If you have already started experiments, this section will help you catch the structural mistakes that are easiest to miss.
A realistic starter path for new researchers
Option B: wearable proof of concept
Use a dataset such as mcPHASES to test temperature and cardiovascular features, but avoid overclaiming broad generalization from a small cohort.
Option C: clinically anchored small cohort
If you can collect your own data, a modest but well-labelled cohort may be more scientifically valuable than a larger weakly labelled convenience sample.
Option D: evaluation-method paper
Work on irregularity subgrouping, metric standardization, or label-definition comparison is itself a meaningful contribution in this field.
Dataset chooser
| Dataset | Best for | What it can answer | What it should not answer | Main risk |
|---|---|---|---|---|
| mcPHASES | AI / DS researchers building wearable proofs of concept | How multimodal physiological signals align with menstrual-health labels | Large-scale product-level generalization | Small sample size and substantial preprocessing complexity |
| Kaggle menstrual cycle tables | Developers and data newcomers who need a fast baseline | Tabular cleaning, feature engineering, grouped evaluation | Strong clinical or wearable conclusions | Dataset provenance, label quality, and preprocessing need extra checking |
| Marquette Menstrual Cycle Data | Researchers interested in historical cycle structure and fertility-awareness records | Non-wearable cycle structure and longitudinal historical patterns | Modern consumer wearable validation | Collection context differs from modern app ecosystems |
Public access does not mean consistent quality. Always document label source, missingness, cohort filters, and whether the dataset resembles deployment reality or a more controlled study setting.
Kaggle and mcPHASES are two very different starting points
The Kaggle route
The mcPHASES route
Starter code and teaching-oriented entry points
Kaggle teaching case
A low-friction baseline case for understanding data cleaning, feature handling, and why GroupKFold matters in this area.
Note: this repo currently ships the teaching page only. The original notebook and local export files are not bundled with the website snapshot.
mcPHASES processing workflow
This is the more wearable-oriented route. The main challenge is not the model, but aligning temperature, HRV, hormone-linked events, and calendar data.
Curated open repositories
iurteaga/menstrual_cycle_analysis
Generative modelling that explicitly handles adherence and missingness. Useful for understanding real-world mHealth structure.
Linkcilab-ufersa/period_cycle_prediction
LSTM / GRU-based cycle-length forecasting. A useful reference for standard sequence-model baselines.
Linklenakmeth/MenstruatioNN
A straightforward RNN implementation for multi-step prediction, useful for medical readers learning how sequence models are framed.
Linknhaboudal/MenstrualCycle_SleepReadiness_Predictor
An example of using Oura Ring and Natural Cycles data to explore physiological shifts across cycle phases.
LinkWhy the most influential datasets are usually not fully open
Large app and consumer wearable studies have shaped the field, but the underlying longitudinal data are rarely available for open download. The usual reasons are privacy, commercial ownership, consent scope, and the sensitivity of reproductive-health data.
- App papers demonstrate scale and pattern, not necessarily reuse rights.
- Wearable papers often publish findings rather than raw longitudinal streams.
- Open resources such as mcPHASES are therefore especially valuable for method comparison and pipeline debugging.
- Some company-held datasets are not openly downloadable but may be accessible through research partnerships, for example at Clue or Oura. Clue Oura
Reproducibility checklist
For app developers, this is the minimum credibility bar. For AI / DS researchers, it is the minimum validity bar. For medical readers, it is the quickest way to judge whether an algorithmic claim is trustworthy.
- State the prediction target clearly: next bleed, ovulation day, fertile window, or phase.
- State the label rule clearly: for example LH onset, LH peak, or peak + 1 day.
- Specify what information is available at prediction time to avoid leakage.
- Prefer subject-wise split or GroupKFold when claiming generalization.
- Report both continuous error and window-based accuracy where appropriate.
- Describe missingness and adherence rather than silently removing hard samples.
- State whether the cohort is regular-only, mixed, or explicitly irregular.