For app developers
- Read the task families and evaluation pitfalls in the Field Guide.
- Then go to Data & Code for starter paths and dataset tradeoffs.
- Finish with Current Research to understand what you should not overpromise.
Open Educational Resource
Start with your role, then enter menstrual prediction research
This OER is built for cross-disciplinary readers. Instead of dropping every visitor into the same wall of papers and terminology, it first helps you locate yourself: are you building an app, approaching the field from AI, or entering from medicine?
Which entry point fits you best right now?
Persona 01
I am an app developer
I need to know what should actually be predicted, which route is realistic first, what public data can support an MVP, and where product claims become scientifically unsafe.
Persona 02I come from AI / data science
I know how to build models, but I need the physiological constraints, label realities, task definitions, and the reasons papers in this field are often not directly comparable.
Persona 03I come from medicine / clinical research
I understand the cycle clinically, but I need to quickly understand data processing, model families, subject-wise splits, and what AI papers are actually comparing.
The shortest route for each audience
For AI / DS researchers
- Learn the cycle basics first.
- Then read how menstrual questions become AI tasks.
- Then move into evaluation design and data resources.
For medical / clinical researchers
- Read the AI task translation and method families first.
- Then see how wearable signals and model families support different questions.
- Then move to the Literature Map and the baseline case study.
Shared minimum path
- Understand that menstrual prediction is not a single task.
- Separate calendar, wearable, and clinical routes.
- Understand why average accuracy is not enough.
Do not default to a 28-day template
Real-world cycle length, ovulation timing, and variability differ widely. Fixed templates will hide biology inside apparent model error.
Do not start with the most complex model
In this field, label definition, splits, and deployment constraints often matter more than architecture choice.
Do not treat wearable signals as ground truth
Temperature, heart rate, and HRV are valuable proxies, but they vary with device, wear site, and cohort.
Do not trust only the average score
Regular cycles, irregular cycles, long cycles, and highly variable cycles may be fundamentally different prediction problems.
Calendar, wearables, and clinical routes answer related but non-identical questions
Calendar / history route
Scalable and app-friendly with low user burden, but fragile for irregular cycles and sparse history.
Wearable physiology route
Can expose biological structure beyond dates alone, but carries noise, device differences, and weak-label problems.
Clinical / biomarker route
Closest to biological ground truth, but costly, smaller in scale, and harder to map onto daily deployment.
Many studies mix wearable and clinical routes: clinical signals define labels, but those labels are unavailable at deployment time.
Why this research matters
The goal is no longer only to improve next-date prediction by a small average margin. The stronger goal is to understand different cycle dynamics, identify under-served groups, and support more specific forms of personalization.
Average accuracy can hide the real problem
Real-world cycles vary substantially in length, follicular timing, and ovulation timing, so one summary score can hide who the model actually works for [1] [2].
Different cycle types create different tasks
Short-stable, long, and highly variable cycles are not the same prediction problem. Subgroup-aware modelling can be more meaningful than a single pooled model [2].
Wearable signals add biological structure
Temperature, heart rate, and HRV matter not only for prediction gains, but because they can reveal structure that date history alone cannot [5].
Better progress means better stratification
A stronger method is not just one with lower average error, but one that explains where failure happens and for whom [4].
Why this area is easy to misread
- Menstrual prediction is not one forecasting problem. It is a person-specific time-series problem with explicit biological constraints.
- Ovulation does not reliably happen at the midpoint, so midpoint heuristics are structurally weak.
- Wearable temperature and cardiovascular signals vary by device and body site and should not be treated as interchangeable.
- Irregular cycles are often collapsed into one coarse category, even though long, high-variability, and shifted-but-stable patterns create different tasks.
- Labels, splits, and metrics vary widely across studies, which weakens paper-to-paper comparison.
Current research gaps
Irregular cycles are under-served
Many methods are developed or reported on relatively regular cohorts, so performance often looks better than it is for the people most likely to need support.
No shared evaluation standard
Studies define ovulation differently, report different windows, and use different metrics, which makes cross-paper comparison weak.
Open data remain scarce
The most influential app and device datasets are often private, so public data are still mainly used for baselines and method validation.
Deployment constraints are often ignored
A model that works retrospectively may still fail when limited to information that would exist in real product use.
Selected references
- Bull JR, Rowland SP, Berglund Scherwitzl E, et al. Real-world menstrual cycle characteristics of more than 600,000 menstrual cycles. npj Digital Medicine. 2019;2:83. Article
- Li K, Urteaga I, Wiggins CH, et al. Characterizing physiological and symptomatic variation in menstrual cycles using self-tracked mobile-health data. npj Digital Medicine. 2020;3:79. PubMed
- Yu JL, Su YF, Ren YY, et al. Tracking of menstrual cycles and prediction of the fertile window via measurements of basal body temperature and heart rate as well as machine-learning algorithms. Reproductive Biology and Endocrinology. 2022;20:118. Article
- Luo C, Su YF, Ren YY, et al. Prediction of the fertile window and menstruation with a wearable device via machine-learning algorithms. Reproductive BioMedicine Online. 2025;51(1):104795. PubMed
- Jasinski SR, Presby DM, Grosicki GJ, et al. A novel method for quantifying fluctuations in wearable derived daily cardiovascular parameters across the menstrual cycle. npj Digital Medicine. 2024;7:373. Article