A Foundation Model for Capturing Complexity of Menstrual Health Data

Abstract

Despite its centrality to women’s health, the menstrual cycle remains understudied in computational health research due to its complexity, variability, and limited data availability. Recent advances in generative artificial intelligence (AI) offer new opportunities for modeling large-scale, user-generated menstrual health data. We introduce and evaluate a generative foundation model trained on self-tracked data from over 1.2 million users of a widely used menstrual tracking app. We assess the model’s ability to generate physiologically plausible synthetic cycles and realistic tracking behaviors, examine whether learned representations capture meaningful temporal and symptomatic patterns, and evaluate privacy risks. Results show that the model produces high-fidelity synthetic data closely mirroring real-world users, with no evidence of data leakage, while learned representations consistently outperform baseline methods on downstream forecasting tasks. These findings highlight generative AI’s potential to advance menstrual health forecasting, support privacy-sensitive data sharing, and enable scientific inquiry in women’s health research.

Introduction

The menstrual cycle is a critical yet underexplored indicator of overall health^1,2. It consists of a series of cyclical hormone-driven processes that manifest in externally observable signs, such as bleeding, pain, and mood changes^1,3. While central to reproductive function, the menstrual cycle also acts as an overall health indicator and can signal risk for chronic conditions such as cardiovascular disease, osteoporosis, and certain cancers^4–8. Its clinical importance is increasingly recognized, with some researchers calling it “the fifth vital sign”¹. Despite its clinical importance, menstruation remains one of the least understood aspects of human biology⁹. Common conditions with strong links to menstruation, such as endometriosis, polycystic ovary syndrome (PCOS), and fibroids, are highly prevalent yet vastly under-researched^10,11. This contributes to persistent gaps in both clinical care and biomedical understanding^12–14. These disparities are emblematic of broader systemic challenges in women’s health research, where access to comprehensive menstrual datasets is limited and analytical tools capable of capturing the complexity of this data remain underdeveloped^9,15.

Over the past decade, menstrual tracking mobile apps have enabled millions of users globally to track menstrual cycles and related behaviors in real time^16,17. Platforms like the Clue app (https://helloclue.com/) have generated large-scale, self-tracked datasets that include daily entries on mood, symptoms, bleeding, sexual activity, and other lifestyle factors^18–23. The resulting data offer a unique longitudinal perspective on individual menstrual experiences outside clinical settings. Momentum is also growing around the use of wearable technologies, such as the Apple Watch, Oura Ring, and Fitbit in menstrual health research^24–29. As these user-generated menstrual datasets continue to grow in scale and resolution, researchers are increasingly focused on developing methods that leverage their richness to generate insights beyond conventional clinical settings.

This growing body of user-tracked data has begun to inform menstrual health research in meaningful ways. Although not collected through traditional research protocols, app-based menstrual data has proven valuable for investigating individual and population-level variability, symptom trajectories across cycles, and changes across life stages^18,30–32. Researchers have even used these datasets to study the effects of stress and environmental exposures on menstrual health^31,33–35. For example, studies have identified cycle disruptions during the COVID-19 pandemic, offering new insight into reproductive responses to physiological and psychological stressors^33–35. Others have investigated associations between air pollution exposure and menstrual cycle irregularities³¹. Together, these studies highlight the opportunity that with appropriate modeling approaches and access to these datasets, menstrual tracking data can generate actionable insights relevant to both individual health and population level health policy.

Despite the significant promise, leveraging the potential benefits from these self-tracked menstrual datasets presents several challenges. The data is noisy, sparse, and temporally irregular, which traditional modeling methods often struggle to capture³⁰. Privacy concerns are especially acute in the context of reproductive health, given rising concerns around surveillance and the politicization of fertility data³⁶. Moreover, prior work shows that de-identification alone may not be sufficient to prevent privacy leakage, motivating the use of synthetic data generation together with explicit, attack-aware privacy analysis for sensitive health datasets³⁷. Compounding these issues, these datasets are shaped as much by logging behavior as by biology, with structured missingness data and user-induced behavioral artifacts emerging from logging habits^18,38. While these factors complicate analysis, they also underscore the need for flexible, data-efficient approaches that can support privacy-aware modeling and enable broader scientific discovery.

To fully realize the potential of self-tracked menstrual data, we need modeling tools that can learn from and leverage irregular and noisy input while preserving user privacy to support scalable health research. In this work, we use the term foundation model to refer to a pretrained model that captures critical structure in a large, heterogeneous dataset and can be flexibly adapted via prompting, learned representations, or fine-tuning to support diverse downstream tasks and cohort construction^39–42. Generative foundation models are well-suited to meet these demands, but adapting them to self-tracked menstrual health data introduces non-trivial challenges. Behavioral sparsity, user-driven structure, and irregular tracking limit the effectiveness of generative tabular and time series foundation modeling approaches, which require designs that can flexibly represent sequential and heterogeneous inputs^43–45. Generative language model–based architectures offer a strong fit for this setting^40,41. Their ability to learn rich, transferable representations from large-scale data enables them to capture nuanced temporal and symptom-based patterns across individuals and cycles⁴⁶. At the same time, generative language models can be used to simulate realistic synthetic data that reflects underlying population and subgroup dynamics without exposing sensitive user-level information^40–42. This dual capacity, data generation and representation learning, makes them a promising tool for advancing menstrual health research. By training such models on large, diverse, self-tracked datasets, we can accelerate personalized predictions, facilitate rapid hypothesis testing, and democratize access to high-fidelity menstrual health data for researchers and developers alike.

In this study, we introduce a generative foundation model for menstrual health data, trained on large-scale, self-tracked records from the Clue app (Figure 1). Clue, developed by BioWink GmbH, is one of the most widely used menstrual tracking apps globally. Users can log information across more than 20 symptom categories, capturing both physiological dynamics (e.g., cycle and period length) and daily experiences (e.g., pain, mood) throughout their menstrual cycle as seen in Figure 2a. We use a de-identified snapshot of this dataset to build a model designed to serve as a reusable foundation model for menstrual health research.

Figure 1. Clue app screenshot.

Screenshot of the Clue menstrual tracking application illustrating cycle visualization and symptom tracking features. Reproduced with permission from Clue by BioWink GmbH.

Figure 2. A foundation model for menstrual data.

(a) A schematic of raw user-tracked data, illustrating how users log a variety of symptoms and experiences (e.g., period flow, energy levels, cravings) across dates. A menstrual cycle is defined as the span from the onset of bleeding greater than spotting to the day before the next such bleeding episode. These events are grouped into discrete cycles (e.g., Cycle 1, Cycle 2), forming the basis for longitudinal modeling. For each cycle, we show: (cycle length in days, number of unique days with tracked events, total number of tracked events). (b) Our foundation model generates realistic synthetic menstrual cycles conditioned on a combination of user demographics and previously tracked cycles. The number of input cycles can be varied, enabling flexible model prompting.

Our model supports two complementary goals: (1) generating realistic synthetic menstrual data to broaden data accessibility and support subgroup-specific analyses, and (2) learning user-level representations that encode temporal and symptom-based patterns across cycles, which transfer effectively to diverse downstream forecasting tasks. We develop a modeling backbone that accommodates the complexity and sensitivity of self-tracked data, enabling privacy-aware analysis through the integration of recent advances in generative modeling with the unique challenges of menstrual health research. It also opens new avenues for individualized predictions, population-level insights, and scientific discovery in menstrual health. Our key contributions are as follows:

• We introduce a scalable framework for encoding self-tracked menstrual data as token sequences, enabling the use of generative language models.

• We leverage these encoded sequences to train the first generative foundation model on menstrual health data.

• We propose an evaluation strategy that integrates generative and predictive performance to benchmark the real-world utility of our foundation model in menstrual health applications.

• We generate and evaluate a synthetic dataset that has realistic menstrual cycle trajectories, demonstrating its ability to produce physiologically plausible and behaviorally realistic synthetic menstrual cycles.

• We assess the utility of the learned representations and the impact of further fine-tuning the foundation model across a range of forecasting tasks, including both discrete and continuous outcomes, demonstrating consistent improvements over baseline representations.

• We utilized an established privacy evaluation framework proposed by Yan et al.⁴⁷ to assess privacy risks in synthetic menstrual health data and found no empirical evidence of elevated individual-level privacy leakage.

Results

Study Population

As summarized in Table 1, the study cohort consists of 1,206,919 unique users between the ages of 13–58 years. The population of users skews young, with 87% of users under 33. 66% of users have tracked using birth control at least once while using the Clue app. We divided the cohort into two non-overlapping subsets for training and evaluation. About 70% users were assigned to the training cohort used to develop our generative menstrual foundation model. The remaining 30% of users make up the evaluation cohort. All results reported in this study are derived from the evaluation cohort.

Table 1. Summary statistics of the study cohort.

High-level demographics and app usage statistics for the full cohort of menstruators included in this study. The cohort is divided into a training cohort, used to develop the generative menstrual foundation model, and an evaluation cohort, used to assess model performance.

	Metric	Study cohort	Training cohort	Evaluation cohort
	Number of users	1,206,919	862,675	344,244
	Number of cycles	23,636,169	16,894,829	6,741,340
	Number of tracked events	405,461,127	290,230,026	115,231,101
Age (years)	< 21	41.93% (506,004)	41.92% (361,661)	41.93% (144,343)
	21–33	44.88% (541,659)	44.88% (387,126)	44.89% (154,533)
	> 33	13.20% (159,256)	13.20% (113,888)	13.18% (45,368)
Birth Control	False	66.50% (802,541)	66.53% (573,928)	66.41% (228,613)
	True	33.50% (404,378)	33.47% (288,747)	33.59% (115,631)

Generating Realistic Synthetic Data

To evaluate whether our foundation model creates realistic and representative menstrual data, we generated a synthetic evaluation cohort and compared it directly to real users. Specifically, we prompted the model with real user demographics and cycle histories in the held-out evaluation cohort (Figure 2b) and assessed how well the generated synthetic cohort matches the statistical and behavioral patterns of the real users. As detailed in the results section below, across metrics that reflect physiological, symptomatic, and behavioral variation, we find strong alignment between the synthetic and real data distributions. Throughout this work, we refer to data generated by our menstrual health foundation model as synthetic data.

Menstrual Cycle Metrics.

Cycle length and period length are closely matched between real and synthetic cycles, indicating that the model accurately reproduces core menstrual timing characteristics (Supplementary Table 1). As shown in Figure 3, the distributions of average cycle length and average period length per user are statistically indistinguishable based on the Mann–Whitney U test (p > 0.05; see Supplementary Table 1) and exhibit low divergence under both Kullback–Leibler (KL) and Jensen–Shannon (JS) metrics (Cycle length: KL = 0.002, JS = 0.023; Period length: KL = 0.012, JS = 0.054). These metrics form the basis of many physiological and clinical assessments, with cycle length defined as the number of days from the first day of one period to the day before the next, and period length defined as the number of consecutive days of bleeding within a cycle^1,9,48.

Figure 3. Comparison of real and synthetic user cycle and period length.

We compare the distribution of per-user (a) average cycle lengths and (b) average period lengths between real and synthetic cohorts. Both distributions are statistically indistinguishable based on the Mann–Whitney U test (see Supplementary Table 1), and exhibit low divergence by both Kullback–Leibler (KL) and Jensen–Shannon (JS) measures (Cycle: KL = 0.002, JS = 0.023; Period: KL = 0.012, JS = 0.054). These results suggest that the generative model captures the cohort-level distributional structure of high-level menstrual metrics.

Temporal structure in these metrics is likewise preserved in the synthetic data. Average cycle length and average period length remain stable across successive cycles within users, with no evidence of systematic drift over time in either cohort or in any variability group (Supplementary Figure 1). Age-stratified analyses across users aged 21–33 further reveal no systematic differences between real and synthetic cohorts within any age stratum (Supplementary Tables 2–5), indicating that the model preserves age-conditional structure for these core menstrual metrics.

In addition to marginal cycle and period lengths, intra-user variability is accurately captured in the synthetic cohort. We evaluate cycle length difference (CLD), defined as the absolute difference between successive cycle lengths¹⁸, and find that both median and mean CLD are closely matched between real and synthetic cohorts (median CLD: 3.0 days in both cohorts; mean CLD: 4.80 days in real vs. 4.78 days in synthetic). Percentile-based tail analyses further show close alignment between real and synthetic data across both lower and upper tails of the cycle length, period length, and CLD distributions (Supplementary Figure 2), demonstrating faithful reproduction of extreme but physiologically plausible values in the full cohort.

Cycle Variability Patterns.

The synthetic data accurately reproduces clinically meaningful heterogeneity in menstrual cycle variability subgroups. Users were stratified into consistently highly variable and consistently not highly variable subcohorts based on cycle length difference (CLD), following the definition of Li et al.¹⁸ (median CLD ≥ 9 days). Using this criterion, the prevalence of consistently highly variable users is closely matched between real and synthetic cohorts (Real: 8.63%; Synthetic: 8.42%).

Within these variability-defined subcohorts, cycle characteristics differ in expected ways between high- and low-variability users but are closely aligned between real and synthetic data. As summarized in Supplementary Table 6, consistently highly variable users exhibit longer cycle lengths, longer period lengths, and substantially higher median and maximum CLD than consistently not highly variable users, reflecting greater cycle-to-cycle irregularity. For each metric, real and synthetic cohorts show near-identical mean and median values within the corresponding subgroups. Consistent with these findings, the distributions of per-user average cycle length within each variability subgroup are nearly identical between real and synthetic cohorts (Supplementary Figure 3).

Extreme variability behavior within each subgroup is also faithfully reproduced. Percentile-based tail analyses show that the synthetic data closely tracks the real data across both lower (p1–p10) and upper (p90–p99) tails for cycle length, period length, median CLD, and maximum CLD within each variability-defined subpopulation (Supplementary Figure 2). Age-stratified analyses within both variability subgroups further reveal stable cycle length and period length distributions across users aged 21–33, with no systematic differences between real and synthetic cohorts at any age (Supplementary Tables 2–5). Together, these results demonstrate that the synthetic cohort accurately reproduces both the prevalence and internal structure of menstrual cycle variability observed in real users, including conditional patterns across age strata and tail behavior within clinically relevant subpopulations.

Symptom Category Distribution.

Symptom tracking frequencies across 19 categories (described in Table 2) are not significantly different between real and synthetic users, demonstrating that the synthetic data preserves the richness and diversity of self-tracked experiences. As shown in Table 2, category frequencies show no statistically significant differences (p > 0.05 for all categories, chi-square test), including both common categories like Period (approximately 48%) and rare ones like Ailment (0.04%). Importantly, synthetic users track symptoms at comparable rates to real users, indicating that the model maintains not only the structural coherence of menstrual cycles but also the behavioral patterns of symptom reporting.

Table 2. Comparison of tracked symptoms between real and synthetic cohorts.

We compare the distribution of logged symptom categories in the real and synthetic datasets, reporting the percentage of all tracked entries belonging to each category. The distributions are closely aligned, with no statistically significant differences across categories based on chi-squared tests (p > 0.05). This suggests that the generative model captures the overall structure of symptom tracking behavior, including the relative frequency of common (e.g., Period, Pain, Emotion) and rare (e.g., Medication, Ailment) categories.

Tracked symptom category	Symptom description	Evaluation subcohort (%)	Synthetic subcohort (%)
Period	Period flow	48.35	47.78
Pain	Type of pain experienced	11.86	12.74
Emotion	Emotional state	9.61	9.41
Energy	Energy level	6.49	6.32
Sex	Sexual health	5.51	5.80
Skin	Skin health	1.63	1.64
Craving	Food craving	1.59	1.72
Mental	Mental health	1.40	1.52
Digestion	Digestive health	1.38	1.34
Motivation	Motivation level	1.28	1.40
Poop	Stool health	1.02	0.99
Social	Social behavior	0.94	1.05
Hair	Hair health	0.88	0.84
Collection method	Method for period collection	0.65	0.65
Fluid	Vaginal discharge type	0.57	0.50
Exercise	Physical exercise	0.20	0.19
Party	Party-related experiences	0.10	0.09
Medication	Type of medication taken	0.06	0.07
Ailment	Physical maladies	0.04	0.04

User Engagement Behavior.

The synthetic data captures realistic patterns of user engagement, closely matching the frequency, volume, and structure of logging observed in the real cohort. Users in both cohorts exhibited similar daily and weekly engagement frequencies. Following the user engagement analysis of Elhadad et al.⁴⁹ we see that over a short term usage window (12 weeks), real users logged data on 15.9% of days on average, compared to 16.5% for synthetic users. Over a long term usage window (36 weeks), real users logged data on 27.4% of weeks on average, compared to 27.8% for synthetic users. The average number of interactions per user was also comparable, with synthetic users logging slightly more events than real users across both daily (29.5 vs. 26.3) and weekly (61.3 vs. 55.1) windows. Beyond these aggregate patterns, both cohorts exhibited similar engagement archetypes (Supplementary Figure 4). Clustering users by their daily binary app usage revealed three dominant behavioral modes: users with sporadic or irregular logging patterns (engagement archetype 0), and users with cyclic logging that aligns with menstruation, split into subgroups reflecting different average cycle lengths (engagement archetype 1 and 2). These archetypes reflect common real-world engagement behaviors, including short-term curiosity-driven use, periodic tracking anchored to menstrual events, and more consistent long-term use. Together, these findings indicate that the model captures not only physiological signals but also realistic user interaction behavior, an essential property for generating behaviorally faithful synthetic datasets.

Predictive Performance on Downstream Forecasting Tasks

To evaluate whether our foundation model learns meaningful user representations that support downstream menstrual health forecasting, we tested six relevant forecasting tasks. These tasks included four discrete forecasting tasks—classifying Period Flow Intensity, Cramps During Menstrual Phase, Spotting After Menstrual Phase, and Cramps After Menstrual Phase (Table 3)—and two continuous forecasting tasks—predicting Next Cycle Length and Days Until First Emotion Symptom Logged (Table 4). To further assess the generalizability and utility of the learned representations, we compared them to a count-based baseline embedding derived from simple symptom frequencies. For each task, we trained lightweight task-specific models using frozen representations from the foundation model, with and without dimensionality reduction via PCA. We also evaluated a fine-tuned variant of the foundation model optimized end-to-end for each prediction task. We find that models using learned representations from the pretrained foundation model consistently outperform baseline embeddings across all downstream menstrual health forecasting tasks. This includes both frozen representations (with or without being projected) and fine-tuned model variants, with the greatest relative gains observed in low-data settings. These results demonstrate the transferability and practical utility of learned representations for personalized prediction. Supplementary Note 1 summarizes the demographic characteristics of the evaluation cohort split into train/test subcohorts which are used in all forecasting tasks.

Table 3. Discrete outcome forecasting tasks.

Performance metrics across four symptom forecasting tasks, where predictive models are trained to predict outcomes in cycle 3 given user demographics and tracked data from cycles 1 and 2. All models were trained with 1,000 examples. We report performance for the task specific fine-tuned foundation model along with three supervised logistic regression classifiers trained on different input representations. Higher values indicate better performance for AUC and F1 scores. Across all discrete forecasting tasks, both the fine-tuned model and task-specific models trained on PCA-reduced learned representations outperform those using the baseline embedding, demonstrating the added value of learned representations.

Forecasting task	Label distribution (%)	Baseline performance	Frozen representation	PCA-reduced representation	Fine-tuned model performance
(1) Period flow intensity on first day of next cycle	Light = 42% Medium = 37% Heavy = 21%	AUC = 0.55 Macro F1 = 0.38 Micro F1 = 0.39	AUC = 0.59 Macro F1 = 0.41 Micro F1 = 0.43	AUC = 0.62 Macro F1 = 0.44 Micro F1 = 0.45	AUC = 0.56 Macro F1 = 0.44 Micro F1 = 0.49
(2) Cramps during menstrual phase of next cycle	False = 66% True = 34%	AUC = 0.71 Macro F1 = 0.67 Micro F1 = 0.69	AUC = 0.74 Macro F1 = 0.68 Micro F1 = 0.70	AUC = 0.78 Macro F1 = 0.71 Micro F1 = 0.73	AUC = 0.81 Macro F1 = 0.69 Micro F1 = 0.75
(3) Spotting after menstrual phase of next cycle	False = 72% True = 28%	AUC = 0.60 Macro F1 = 0.58 Micro F1 = 0.63	AUC = 0.61 Macro F1 = 0.58 Micro F1 = 0.64	AUC = 0.66 Macro F1 = 0.60 Micro F1 = 0.65	AUC = 0.71 Macro F1 = 0.57 Micro F1 = 0.73
(4) Cramps after menstrual phase of next cycle	False = 92% True = 8%	AUC = 0.59 Macro F1 = 0.54 Micro F1 = 0.80	AUC = 0.65 Macro F1 = 0.57 Micro F1 = 0.85	AUC = 0.67 Macro F1 = 0.56 Micro F1 = 0.78	AUC = 0.75 Macro F1 = 0.52 Micro F1 = 0.92

Table 4. Continuous outcome forecasting tasks.

Performance metrics across two continuous forecasting tasks, where predictive models are trained to predict outcomes in cycle 3 given user demographics and tracked data from cycles 1 and 2. All models were trained with 1,000 examples. We report performance for the task specific fine-tuned foundation model along with three supervised linear regression classifiers trained on different input representations. Lower values indicate better performance RMSE (days) and MAE (days). The PCA-based representation outperforms the count vector baseline across both continuous forecasting tasks.

Forecasting task	Label distribution (days)	Baseline performance	Frozen representation	PCA-reduced representation	Fine-tuned model performance
(5) Next cycle length	Median: 29 IQR: 26–33 P95: 73	RMSE = 27.04 MAE = 17.18	RMSE = 47.24 MAE = 32.50	RMSE = 15.07 MAE = 11.61	RMSE = 14.18 MAE = 8.90
(6) Days until emotion tracked in next cycle	Median: 0 IQR: 0–1 P95: 24	RMSE= 11.14 MAE = 2.34	RMSE = 19.92 MAE = 4.79	RMSE = 3.55 MAE = 1.36	RMSE = 10.17 MAE = 7.98

Forecasting Discrete Menstrual Outcomes.

Fine-tuned and learned representation-based models outperform the unlearned user embedding baseline across all discrete forecasting tasks, with the fine-tuned model generally achieving the highest AUC and F1 scores (Table 3). While the Period Flow on First Day of Next Cycle task is best handled by the PCA-reduced representation based models (AUC = 0.62, F1 = 0.44), the fine-tuned model leads on other tasks, especially in imbalanced cases like Cramps After Menstrual Phase of Next Cycle (True = 8%) where it achieves +0.16 AUC over baseline. These findings highlight the effectiveness of pretrained representations in both balanced and skewed discrete forecasting settings. The strong performance of these models across forecasting tasks targeting different phases of the next cycle suggests that the user-level representations effectively capture temporal dynamics and symptom-based patterns across cycles, enabling flexible and accurate prediction of future outcomes.

Forecasting Continuous Menstrual Outcomes.

PCA-reduced representation models consistently outperform the unlearned user embedding baseline across both continuous forecasting tasks, demonstrating that pretrained user-level representations encode transferable temporal structure (Table 4). While the fine-tuned model achieves the best performance for Next Cycle Length prediction, with root mean square error (RMSE) of 14.18 and mean absolute error (MAE) of 8.90, the PCA-reduced representation leads on the Days Until Emotion Tracked in Next Cycle task (RMSE = 3.55, MAE = 1.36). Notably, fine-tuning underperforms PCA-reduced frozen representations on Task 6, which we attribute to the extreme zero-inflation and long-tailed label distribution interacting with the task-specific optimization strategy, rather than to deficiencies in the underlying learned representations. The baseline model performs worst overall, particularly on the skewed cycle length task (RMSE = 27.04, MAE = 17.18), underscoring the limitations of unlearned user embeddings in capturing nuanced temporal durations compared to learned representations. We additionally evaluated simple per-user statistical baselines (historical mean and median cycle length), which performed substantially worse under our task definition (see Supplementary Note 2), reflecting the heterogeneity and heavy-tailed nature of self-tracked cycle data.

Impact of Experimental Setting.

Learned user representations consistently outperform baseline embeddings across all training sizes and user history lengths, with the largest performance gains observed in low-data and short-history settings. This highlights their strong generalizability and efficiency for menstrual health forecasting. We first examined how performance changes with the task specific training set size. As shown in Figure 4a, learned representation-based models outperform the unlearned baseline count vector across all data regimes. Gains are especially pronounced in low-data settings, where learned embeddings provide a strong advantage. Next, we assess the impact of input history by varying the number of prior user cycles provided as input. As shown in Figure 4b, we see that longer user histories consistently lead to a higher AUC. Notably, the baseline model lags behind across all history lengths, reinforcing its limitations in capturing temporal structure even when additional input is available. Together, these findings demonstrate that learned representations not only offer strong performance with limited training data but also scale effectively with increased user history. Additional breakdowns of the impact of experimental setup across all tasks are provided in Supplementary Note 3.

Figure 4. Impact of experimental setup on forecasting performance.

We evaluate AUC (ROC) for the Cramps During Menstrual Phase task while varying both training set size and the amount of user history provided as input. In (a), AUC improves steadily as training size increases, with the fine-tuned and PCA-reduced models outperforming both the baseline and frozen representation models, particularly in low-data regimes. In (b), AUC improves as the model is given more prior user cycles to inform predictions. Learned representations (PCA-reduced and fine-tuned) benefit most from deeper input histories, while the baseline model remains the weakest across all settings. These results reinforce the value of pretraining and temporal context when forecasting symptom occurrence from menstrual tracking data. All plotted lines include 95% confidence intervals.

Privacy Evaluations

We evaluate privacy implications using three complementary analyses: membership inference, attribute inference, and memorization. Together, these analyses test whether real user trajectories can be identified as having been included in the training set (membership inference), whether synthetic trajectories enable inference of withheld sensitive attributes (attribute inference), or whether the model reproduces individual training sequences verbatim (memorization). Across all three evaluations, we find no empirical signal of reliable identification of training users, improved inference of sensitive attributes, or systematic reproduction of individual training cycles beyond natural duplication observed in the real data. Taken together, these results support using the generated data to facilitate privacy-sensitive sharing under minimized observed risk in our setting, while recognizing that privacy is attack specific and should be re-evaluated for new data distributions or release conditions.

Membership Inference Attack.

The membership inference attack performed at chance level, indicating that training and non-training records are indistinguishable based on proximity to the training dataset. Across 100 independent runs with 10,000 training members and 10,000 non-members per run, the attack achieved a binary classification accuracy of 50.0%±0.2% and a ROC-AUC of 0.500±0.004, both consistent with random guessing. Under the fixed decision threshold, precision was 50.1%±1.2%, while recall was low (6.7%±0.2%), yielding an F1 score of 0.118±0.004. This pattern reflects a conservative classifier that rarely predicts membership and fails to recover true members. The chance-level ROC-AUC confirms that nearest-neighbor distance does not provide discriminative power even when used as a continuous score. Nearest-neighbor distance distributions for members and non-members were nearly identical. Median distances were 15.13 for members and 15.11 for non-members, with strongly overlapping upper tails (95th percentiles: 138.6 vs. 138.4). Taken together, these results provide no empirical evidence that individual training records can be reliably identified under this membership inference attack, suggesting low observed membership leakage risk. In this setting, attack performance at chance level indicates that synthetic trajectories do not expose distinguishing signals that would allow an adversary to determine whether a given user was part of the training set. Results can be seen in Supplementary Table 9.

Attribute Inference Attack.

The attack yielded limited inference capability, with synthetic data performing worse than a real-vs-real baseline. The attribute inference attack tests whether an adversary with partial knowledge of a user’s trajectory can infer additional sensitive attributes by matching to synthetic data generated by the model. Across 50,000 target users, matching to synthetic continuations achieved a macro-averaged F1 score of 0.246, compared to 0.341 when matching targets to other real users. At the individual user level (macro-user aggregation), synthetic-vs-real inference achieved precision of 39.76% and recall of 24.48%. The real-vs-real baseline showed higher performance with precision of 41.38% and recall of 41.42%. Importantly, lower performance in the synthetic-vs-real setting relative to the real-vs-real baseline indicates reduced attribute inference risk. In other words, an adversary gains less information from synthetic trajectories than from naturally similar real users alone. Results can be seen in Supplementary Table 10.

Memorization Analysis.

The memorization analysis revealed minimal evidence of training data reproduction in model-generated sequences. We assess memorization by testing whether the model reproduces a user’s unseen future cycle verbatim when prompted with that user’s observed history, compared against natural similarity between different real users. Across 50,000 target users with five synthetic continuations each, only 0.82% of users had any generated continuation that exactly matched their true fourth cycle, compared to a real-vs-real exact match rate of 27.3% among naturally similar users within the cohort. Near-duplicate rates (normalized edit distance ≤ 0.05) were similarly low: 0.93% for real-to-synthetic versus 27.5% for real-to-real. Mean normalized edit distances were 0.462 for real-to-synthetic compared to 0.332 for real-to-real, indicating generated sequences were substantially less similar to users’ true continuations than those users’ nearest neighbors in the real data. Lower exact-match and near-duplicate rates in real-to-synthetic comparisons relative to real-to-real baselines indicate that the model generalizes beyond individual training trajectories rather than memorizing them. Results can be seen in Supplementary Table 11.

Discussion

In this work, we present the first large-scale generative foundation model trained on self-tracked menstrual health data. The model bridges gaps in menstrual health research by supporting two key capabilities: generating realistic synthetic data, and learning user-level representations that transfer effectively to a wide range of downstream forecasting tasks. Our results demonstrate that the model can both generate plausible future cycles and improve performance across discrete and continuous forecasting tasks, especially in low-data regimes. We introduce a novel framework for representing menstrual cycle data as token sequences, enabling the application of generative language models to this complex, temporally irregular domain. Additionally, to support broader adoption and benchmarking, we contribute a reusable evaluation framework that includes fidelity metrics for generative outputs and a task-based setup for assessing predictive performance which together establish standardized tools for future menstrual foundation models. This work directly addresses long-standing data and tooling limitations in menstrual health research, offering a reusable modeling backbone that supports broader data accessibility and enables new avenues for scientific inquiry in this underexplored area.

Our work builds on three existing types of foundation models: tabular, time series, and Electronic Health Record (EHR) based models. Prior tabular and time series foundation models are often constrained by independent and identically distributed (i.i.d.) assumptions, regular sampling requirements, and limited support for sparse or irregular inputs, making them poorly suited for self-tracked health data. By contrast, our approach draws inspiration from recent advances in EHR-based foundation models, which have shown promise in capturing the temporal, multi-modal nature of longitudinal health data^{40,41,50–52}. However, unlike EHR data, menstrual tracking datasets are predominantly self-tracked, behaviorally mediated, and subject to unique patterns of structured missingness. Prior menstrual modeling efforts have largely focused on prediction within narrow scopes (e.g., cycle length or ovulation forecasting) using task-specific models^30,53–56. By contrast, we unify generative modeling and representation learning in a single model trained across users and cycles, facilitating broader utility across forecasting tasks. Our model explicitly addresses limitations in earlier work by handling irregular, sparse input and learning representations that generalize across diverse forecasting tasks, a crucial step toward bridging the methodological advances in foundation modeling with the specific needs of menstrual health research.

Consistent with prior foundation models for biological and clinical sequence data, we adopt an autoregressive (AR) pretraining strategy as the modeling backbone in this work. AR language model pretraining is a widely used approach for learning transferable representations from heterogeneous, temporally ordered biomedical data, including biological sequences and longitudinal electronic health records^{40,41,50,51,57}. Even with relatively generic objectives, such models have been shown to capture meaningful structure and support flexible adaptation to diverse downstream tasks via prompting or fine-tuning, which aligns with our goal of learning reusable representations from self-tracked menstrual data rather than optimizing for a single prediction task. At the same time, our results highlight clear limitations of this choice. Menstrual physiology is inherently cyclical, while AR objectives emphasize local temporal dependencies and do not explicitly encode cycle-level or time-to-event structure, which may limit performance on longer-horizon or phase-specific outcomes. Recent work in EHR foundation modeling has shown that alternative pretraining objectives and architectures, such as time-to-event or survival-based models, can better capture irregular timing and periodic phenomena⁵⁸. Exploring such objectives and architectures is an important direction for future work.

Within this setting, realistic synthetic data generation plays a central role in our approach, offering a pathway to broaden access to high-fidelity menstrual health data while accounting for the sensitivity of user-tracked information^41,42. Our generative model captures both physiological patterns (e.g., cycle length distributions, bleeding dynamics, and symptom prevalence) and behavioral characteristics (e.g., engagement patterns) observed in real user cohorts, enabling flexible downstream use. In practice, the model supports two complementary modes. First, real user histories can be used as prompts to simulate future cycles, enabling cohort simulation, subgroup-specific analysis, and the development of privacy-sensitive tools. Second, fully synthetic user trajectories can be generated to explore population-level questions such as cycle variability, symptom co-occurrence, or engagement dynamics without direct reliance on identifiable records. More broadly, generative models of this form may support stress-testing predictive systems under targeted cycle conditions⁵⁹, bootstrapping data for underrepresented subgroups^60–62, generating causal hypotheses, or studying causal counterfactuals (e.g., how symptom trajectories might change under hypothetical interventions)^63–65. Given the sensitivity of menstrual health information and the increasing politicization of reproductive data, privacy-preserving data synthesis is not just a technical challenge but also an ethical imperative³⁶. Consistent with this need, our privacy evaluation finds no evidence of elevated individual-level data leakage under membership inference attacks, attribute inference attacks, or memorization analyses, supporting the use of the generated synthetic data for privacy-sensitive research and development.

Learning meaningful representations is especially critical in the context of menstrual health, where cycles vary widely between individuals and symptoms reflect complex hormonal and behavioral interactions^1,18. Forecasting tasks provide a natural and practical framework for evaluating representation quality in this domain, as they require the model to anticipate future physiological and behavioral outcomes based on prior user data. Our model’s ability to distill these dynamics into user-level embeddings enables researchers to rapidly adapt to new prediction tasks without retraining from scratch. This is particularly impactful in low-data settings, where traditional models often struggle due to limited labeled data and high inter-user variability. We show that even with constrained supervision, the foundation model supports accurate and robust outcome forecasting. Additionally, the model’s architectural flexibility, supporting both autoregressive generation and representation-based forecasting, positions it as a versatile tool that can support a wide range of research workflows, from exploratory data analysis to applied predictive modeling in digital health contexts.

Despite these contributions, several limitations remain. First, while we conduct extensive validation of menstrual cycle structure, variability, temporal consistency, and age-conditional patterns, we do not separately evaluate cohorts defined by hormonal interventions or condition symptom reporting on finely defined cycle phases. Such analyses are scientifically important but require careful cohort construction and phase definitions that introduce additional modeling and labeling considerations beyond the scope of the present study. Second, while our evaluation spans both generative and predictive performance, there are few, if any, task-specific benchmark models established for menstrual health prediction, limiting our opportunity for direct comparison. This highlights the need for future work to develop and validate task-specific baselines that can serve as reference points for this emerging area. Third, our predictive evaluation focuses on outcomes within a single upcoming cycle and does not assess performance across longer time horizons. In addition, we observe that end-to-end fine-tuning can be sensitive to extreme label skew and behaviorally mediated outcomes, a phenomenon that has been noted in prior work on fine-tuning instability in pretrained models⁶⁶. Under such conditions, frozen or low-dimensional representations with stronger implicit regularization may yield more stable performance. Future work should explore multi-step and longitudinal prediction settings, as well as alternative adaptation strategies better suited to highly skewed outcome distributions. Finally, although we contribute a task-based evaluation framework, standardized benchmarks and broader stakeholder input will be essential for assessing downstream utility in real-world contexts.

Future research may explore several directions to address these limitations and build on this foundation. First, this modeling approach could be extended to other user-generated menstrual datasets to evaluate generalizability and broaden our understanding of different menstruating populations. For example, model transferability could be assessed in cohorts with chronic conditions such as endometriosis, where symptom patterns and tracking behaviors may differ substantially. Second, the suite of downstream prediction tasks could be expanded to include more clinically grounded or behaviorally relevant outcomes, and compared against task-specific models to benchmark performance. In a similar vein, future work could extend the predictive task framework to evaluate zero-shot inference performance, providing insight into how it compares with fine-tuned models and simple classifiers trained with learned representations. Third, future work could explore how a flexible menstrual foundation model might be integrated into real-world menstrual tracking apps, for example, to personalize symptom forecasts, simulate next-cycle scenarios, or recommend tailored tracking prompts. Finally, consistent with broader challenges in EHR foundation modeling, ongoing work is needed to improve longitudinal validation by assessing model robustness across multiple future time horizons.

Menstrual health represents a critical frontier for foundation modeling. It is a domain rich with temporal, dynamic signals and deeply entangled with social and biological factors, yet historically marginalized in health data science. Our work demonstrates that generative foundation models can effectively learn from menstrual data, generating synthetic users and supporting robust downstream predictions. By advancing data-driven tools that center menstrual health, we not only elevate an understudied area of science but also help close longstanding data gaps in women’s health and expand the scope of questions that can be asked, and answered, in menstrual health research.

Methods

Dataset

This study utilizes a snapshot of a de-identified dataset from the Clue data warehouse, comprising longitudinal self-tracked menstrual cycle data from over 4 million users. From this population, a random sample of approximately 1.2 million users was selected to form the study cohort used for model development and evaluation. The only inclusion criterion was that users must have recorded at least three complete menstrual cycles. Due to the structure of the dataset provided by Clue, users may register before age 18, but no self-tracked menstrual data are included until users reach age 18. The resulting study cohort was subsequently divided into a pretraining cohort (862,675 users) and an evaluation cohort (344,244 users), as summarized in Table 1. The full study cohort includes more than 23 million unique menstrual cycles and over 405 million individual tracked events. For evaluating the utility of the model’s learned representations through predictive performance on downstream tasks, the full held-out evaluation cohort was further partitioned into training and testing sets to use for task specific forecasting models, as shown in Supplementary Note 1. To assess the generative capabilities of the menstrual foundation model, specifically, the realism of its synthetic data output, a subcohort of users aged 21–33 with no tracked usage of birth control was taken from the full held-out evaluation cohort and used as seen in Supplementary Table 1. After excluding users with fewer than three tracked cycles and removing anomalous cycles per Li et al.¹⁸, the final comparison cohorts included 98,524 real users and 98,566 synthetic users.

This study was conducted in accordance with relevant ethical guidelines. The dataset was provided to the researchers in a fully de-identified form by Clue by BioWink GmbH. All Clue App users included in this study consented to the use of their data for research purposes. The study protocol was reviewed by the Columbia University Institutional Review Board and determined to be Not Human Subjects Research under 45 CFR 46.

Foundation Model Building

Existing approaches for modeling structured data fall short in the setting of self-tracked menstrual health data. Tabular foundation models typically assume that each row in the dataset is i.i.d., making them ill-suited for capturing temporal dependencies^44,67,68. They also struggle to handle common features of self-tracked health data, such as missing entries, irregular sampling, and behavioral sparsity^43,69. Time series models often require regular sampling and numerical features, limiting their ability to capture irregular, multivariate symptom logs^45,70,71. In contrast, transformer-based EHR models have demonstrated strong performance on similarly irregular and heterogeneous health data^40,72,73. Our model adapts this methodology to the menstrual health domain, enabling personalized modeling and synthetic data generation across user timelines.

In our work we develop a generative, autoregressive transformer model for self-tracked menstrual health data. The raw Clue dataset consists of event logs, where each row corresponds to a user-reported symptom recorded on a given date. To capture temporal and individual-level structure, we group events by user, sort them chronologically, and encode them into structured token sequences that reflect event type, timing, and categorical attributes (Figure 5). These user-level sequences are used to train the model to predict the next token in the sequence, enabling both generative modeling and representation learning.

Figure 5. User representation design.

The user representation preserves demographic information, tracked menstrual cycle symptoms, and temporal intervals between events. Each sequence begins with a demographics prompt that includes the user’s age at sign-up, the year of first tracking, the categorical BMI if provided, and the type of birth control used at sign-up (if any). It also includes a flag indicating whether the user has ever documented any form of birth control use while using the app. If symptom-tracking events occur prior to the first recorded period, a pre-cycle block is inserted between the demographics prompt and the first cycle block. Each cycle block retains all tracked symptoms in chronological order. Artificial time tokens (ATTs) are included to represent the number of days between events, preserving temporal structure. Each user representation encodes a complete longitudinal timeline for that individual.

User Representations.

To model each user’s menstrual tracking history, we encode their event data into a single, long-form text sequence that preserves the temporal structure of their logged events. This representation strategy transforms chronologically ordered symptom and cycle data into a sentence-like format that is well-suited for generative language models. Our approach extends patient encoding techniques used in clinical foundation models such as CEHR-GPT⁴⁰. Each user sequence begins with a demographics prompt, followed by the user’s cycle and symptom tracking data in chronological order, as illustrated in Figure 5. The demographics prompt includes the user’s age, categorical BMI, and the type of birth control (if any) recorded at sign up. It also includes the year of their first tracked event and a binary flag indicating whether the user has ever been on any form of documented birth control during their time using the Clue app.

Following the demographics prompt, the representation includes one or more menstrual cycle blocks. Each cycle block begins with an artificial cycle start token (CS) and ends with a cycle end token (CE). A menstrual cycle is defined as the span of days beginning with the onset of menstrual bleeding greater than spotting and ending the day before the start of the next such bleeding event, consistent with prior menstrual cycle research using Clue app data^18,48. To define a period episode, we follow Li et al.¹⁸ in identifying bleeding events (heavier than spotting) that occur within 10 days of a previous bleeding event as belonging to the same episode. A new period episode is defined by the absence of qualifying bleeding for at least 10 days. Within each cycle, tracked events—such as bleeding and symptoms—are tokenized using a structured format in which each token contains the event category followed by the specific event. For example, a tracked period with light flow would be represented as period:light, and an emotional state of happiness would be recorded as emotion:happy. In some cases, users begin symptom tracking before their first recorded period. To preserve these early observations, we include a pre-cycle block that follows the demographics prompt and precedes the first menstrual cycle block. This block, if present, is enclosed by the pre-cycle start (PCS) and pre-cycle end (PCE) tokens and contains all symptom-tracking events that occur prior to the user’s first documented period event.

Events within each cycle block are separated by artificial time tokens (ATTs), which encode the time elapsed between successive events. When multiple events occur on the same day, they are ordered alphabetically by token to ensure consistent sequence construction. We quantify the prevalence and impact of within-day ordering choices in Supplementary Note 4, including statistics on multi-event days and an ordering-sensitivity ablation on held-out validation data. If the gap between events is 60 days or fewer, a day-based token is used (e.g., D₁ for a one-day interval). For intervals between 61 and 120 days, the gap is expressed using week-based tokens, rounded down to the nearest full week (e.g., W₇ for a 7-week gap). For gaps longer than 120 days, the interval is approximated using month-based tokens, rounded down assuming 30 days per month (e.g., M₃ for a 3-month interval). This discretization captures temporal spacing while reducing the sparsity introduced by long-tailed time gaps.

Generative Model.

We used a standard auto-regressive transformer-based architecture for generative pretraining. Training was performed on 862,675 unique user representations comprising 16.9 million menstrual cycles and over 290 million tracked symptoms. Each user’s input sequence was constructed as described above, starting with a structured demographics prompt followed by a temporally ordered series of tracked symptoms and cycle events. These sequences were then tokenized into discrete concept tokens and embedded using a learned embedding table, supplemented with trainable position embeddings.

While generative modeling can employ a variety of training strategies we chose the Next Token Prediction objective due to its alignment with the chronologically ordered, event-level nature of menstrual tracking data. This approach involves predicting the next token in a user’s sequence given all previous tokens, thereby encouraging the model to learn fine-grained temporal dependencies and event transitions. This objective enables the model to capture the evolving patterns of symptoms and cycles over time, reinforcing the temporal structure embedded within the user representations. We trained the model for five epochs, a choice that balanced training efficiency with modeling capacity. Inspection of training loss revealed that convergence occurred within this window, with loss plateauing by the fifth epoch and no additional gains observed on the validation set. Additional training pipeline details can be found in Supplementary Note 3.

Evaluating Foundation Model Through Synthetic Data Generation

We assess the generative quality of the foundation model by comparing synthetic user data generated by the model to real-world self-tracked data, a common approach for assessing medical generative models^40,42. By evaluating physiological patterns and behavioral engagement in matched real and synthetic cohorts, we test whether the model captures realistic distributions and user behaviors. This provides insight into the model’s ability to encode and reproduce the underlying structure of menstrual health data.

Synthetic Data Generation.

To generate synthetic data, we prompted the pretrained generative model with each user’s menstrual tracking history and then extended their logged data with synthetic generated menstrual cycles. For each user in the evaluation cohort, we initialized generation with their demographic block and first three tracked cycles (Figure 2b). The model then generated additional cycle data using top-k sampling with k = 100. Generation continued until the model either stopped producing new tokens or reached a manually defined maximum context length of 1024 tokens. This procedure was repeated for all users in the cohort, resulting in a synthetic Clue app user dataset. These synthetic user timelines were then directly compared to real user data in downstream analyses.

Menstrual Cycle Comparison.

To evaluate the similarity between synthetic and real user data, we performed cycle-level comparisons across matched users and metrics. Following the definition established by Li et al.¹⁸, we focused on users aged 21–33 with no documented birth control use and at least three logged cycles, excluding any anomalous cycles per the authors’ criteria. Consistent with prior definitions¹⁸, we excluded cycles shorter than 10 days from both real and synthetic cohorts, as such intervals do not constitute complete menstrual cycles and may reflect truncated final cycles rather than true cycle durations. Synthetic cycles were matched to real cycles by unique user-cycle identifiers to ensure one-to-one comparisons. We also adopted the cycle length difference (CLD) metric from Li et al.¹⁸, defined as the absolute difference between consecutive cycle lengths. Users were labeled as highly variable if their median CLD was greater than or equal to 9 days, and not highly variable otherwise. These rules were applied identically to the real and synthetic cohorts to ensure comparability.

We assessed cohort-level alignment using a range of statistical tests. Aggregate user metrics (including number of cycles, average cycle length, period length, and CLD) were compared using the Mann–Whitney U test (Supplementary Table 1). To evaluate the similarity of distributional shapes for cycle and period lengths, we computed Kullback–Leibler (KL) and Jensen–Shannon (JS) divergence scores. Symptom tracking frequencies across 19 categories (Table 2) were compared using a chi-squared test. Together, these analyses enabled a comprehensive comparison of physiological, behavioral, and variability-related patterns in menstrual cycles between the real and synthetic user cohorts.

User Engagement Pattern Comparison.

To evaluate temporal engagement patterns, our approach builds on prior work by Elhadad et al.⁴⁹, which emphasizes the role of patterns of engagement in understanding self-tracking behavior in menstrual health contexts. Here we transformed daily tracking logs into fixed-length binary arrays, where each position indicates whether a user logged any activity on a given day. These arrays were generated for 84-day windows beginning after the third cycle. PCA was then applied to these binary arrays and the reduced projections were clustered by k-means clustering (k = 3) to group users based on engagement patterns⁴⁹. For visualization, we sampled up to 200 users per cluster and sorted each cluster by total engagement before displaying the binary engagement matrix (Supplementary Figure 4). This analysis was performed independently on both the real and synthetic cohorts to assess whether the generative model reproduced typical patterns of user engagement.

Evaluating Foundation Model Through Downstream Forecasting Tasks

We evaluate the representational quality of the foundation model by testing how well its learned embeddings support predictive modeling of future symptoms and cycle-related outcomes⁴⁶. Performance across both discrete (classification) and continuous (regression) forecasting tasks reflects the model’s ability to generalize and extract informative structure from user histories. This evaluation probes whether the model’s internal representations encode features relevant for personalized menstrual health forecasting.

Forecasting Task Definitions.

To evaluate the utility of the pretrained foundation model for personalized menstrual health forecasting, we defined six downstream tasks: four discrete (classification) forecasting tasks and two continuous (regression) forecasting tasks. Each task targets a symptomatic or temporal outcome occurring in the user’s next menstrual cycle. The tasks are as follows:

1. Period Flow Intensity Forecasting: Predicting the intensity of period flow, categorized as light, medium, or heavy, on the first day of the menstrual phase of the next menstrual cycle.

2. Cramps During Menstrual Phase: A binary prediction of whether the user will track cramps at any time during the menstrual phase of the next menstrual cycle.

3. Spotting After Menstrual Phase: A binary prediction of whether the user will track spotting at any point after the menstrual phase during the next menstrual cycle.

4. Cramps After Menstrual Phase: A binary prediction of whether the user will track cramps at any point after the menstrual phase during the next menstrual cycle.

5. Next Cycle Length Forecasting: Predicting the total length, in days, of the user’s next menstrual cycle.

6. Days Until First Emotion Symptom Logged: Predicting the number of days from the start of the next menstrual cycle until the user logs their first emotion-related symptom. For example, a label of 0 indicates that an emotion was logged on the first day of the cycle. This task includes only users who track at least one emotion symptom during the cycle and forecasts when the first such entry occurs, not whether one will occur.

All evaluations were conducted using the held-out evaluation cohorts, which was randomly split into 80% training and 20% testing subsets. For Tasks 1–5, this full split was used for training and evaluation. For Task 6, which requires that users log at least one emotion symptom during the cycle, the same 80/20 split was applied to the full cohort, after which the training and test sets were filtered to include only users with at least one emotion-tracking entry in their next cycle. Supplementary Note 1 summarizes the demographic characteristics of the train/test cohorts used in all forecasting tasks.

Model Types and Comparisons.

We compared four modeling strategies that differ in how they represent user histories and whether they adapt the pretrained foundation model to the forecasting task. This comparison allows us to evaluate the representational quality of the foundation model and assess how it supports downstream menstrual health forecasting across tasks and data regimes. The modeling strategies include:

• Unlearned baseline representation (count-based): User histories were represented using sparse count vectors of token frequencies derived from each user’s encoded input sequence, forming a simple bag-of-concepts feature set.

• Frozen learned representation (768-dim): For each user, we extracted the final hidden state of the last token in their input sequence from the pretrained foundation model, without any further task-specific training.

• PCA-reduced frozen representation (100-dim): To support low-data settings, we applied principal component analysis (PCA) to reduce the dimensionality of the frozen learned representations to 100 dimensions.

• Task-specific fine-tuned model: The pretrained foundation model was fine-tuned separately for each forecasting task, allowing all parameters to adapt to the task objective.

For the first three approaches, the resulting user representations were used as input features to lightweight, task-specific classifiers or regressors. In the fine-tuned setting, the full pretrained model and task-specific head were trained jointly.

Experimental Setup.

We evaluated all modeling strategies using task-specific prediction models. For the unlearned baseline and embedding-based approaches (frozen and PCA-reduced), user representations were used as input features to lightweight supervised models: L1-penalized logistic regression for discrete forecasting classification tasks and ridge regression for continuous forecasting regression tasks.

For the fine-tuned models, we appended a special «PRED» token to each input sequence and trained the model end-to-end using supervised learning. In discrete forecasting classification tasks, the model was trained to generate the correct label as the next token, with outputs restricted to a task-specific vocabulary. In continuous forecasting regression tasks, the hidden state of the «PRED» token was passed through a linear layer to predict a log-transformed continuous target.

Performance was evaluated using task-appropriate metrics, with multiple runs at each training set size to enable confidence interval estimation and assess model stability. For classification tasks, we reported area under the ROC curve (AUC), macro F1, and micro F1 scores. For regression tasks, we reported mean absolute error (MAE) and root mean squared error (RMSE). Regression targets were log-transformed during training, capped at the 95th percentile to reduce the influence of outliers, and exponentiated at evaluation to return to the original time scale. All results reported in the main text reflect comparisons across the unlearned baseline, the frozen and PCA-reduced embeddings, and the fine-tuned foundation model. For each task, models were trained using a range of labeled user counts (from 100 to the full 275,395-user training set), with repeated runs at each size using randomly sampled subsets to assess performance variability. Final metrics were computed on the held-out test set.

Privacy Evaluations

We adapted the privacy evaluation framework from Yan et al.⁴⁷ to assess privacy risks in our model through three complementary approaches: a membership inference attack, an attribute inference attack, and a memorization analysis. Consistent with our model’s design for generating realistic synthetic user data, all evaluations use each user’s demographics and first three menstrual cycles as the conditioning context, with subsequent cycles serving as the target of inference or reproduction. For each target user, we generated five synthetic continuations by prompting the model with this prefix. We represent users as fixed-length vectors over the token vocabulary and employ nearest-neighbor approaches to assess similarity and inference capability.

Membership Inference Attack

We evaluated membership inference risk by testing whether an adversary could determine if a real user trajectory was included in the training set based on its similarity to other training records, following prior work^47,74. For each run, we constructed an attack set A containing 20,000 users: 10,000 sampled from the training set (members) and 10,000 sampled from a held-out evaluation set (non-members). We also constructed a separate neighbor set T^′, consisting of 100,000 training users disjoint from the training members in A, to prevent trivial self-matches. Each user trajectory was represented as x, a token-frequency count vector. For each user a ∈ A, we computed the Euclidean distance to its nearest neighbor in T^′:

$$d\left(a,T\text{'}\right)=\underset{t\in T\text{'}}{\text{min}}\Vert {\text{x}}_a-{\text{x}}_t{\Vert }_2$$ (1)

We classified a user as a training member if d(a,T^′) < τ, with a fixed threshold τ = 5, based on the hypothesis that training members would be closer to other training records than non-members. We evaluated performance using accuracy, precision, recall, F1 score, and ROC-AUC. The full procedure was repeated for 100 runs with independently resampled attack sets, and we report mean ± standard deviation.

Attribute Inference Attack

We evaluated attribute inference risk by testing whether an adversary with partial knowledge of a user’s trajectory could infer additional sensitive symptoms by matching to synthetic data and copying remaining attributes, following prior work^47,74. We define the adversary’s known attribute set K to include (i) demographics and (ii) the most prevalent symptom tokens in the first three cycles, operationalized as the top 5% prevalence tier ranked by user-level prevalence in the training data. All remaining symptom tokens are treated as sensitive attributes S. We sampled 50,000 target users from the held-out evaluation set. Each user is represented as a binary vector over the known-attribute vocabulary. For each target user a, we identify the nearest neighbor in known-attribute space using Euclidean distance. In the primary attack (Synthetic→Real), we match each target to the nearest synthetic continuation; in the baseline (Real→Real), we match each target to the nearest other real user from a disjoint evaluation search pool. The adversary predicts the target’s sensitive attribute set Ŝ(a) by copying sensitive symptom tokens from the matched record. We compare Ŝ(a) against the target’s ground-truth sensitive set S(a) and report macro-averaged precision, recall, and F1 across users. We compare Synthetic→Real performance against the Real→Real baseline to contextualize inference difficulty.

Memorization Analysis

We assessed memorization by testing whether the model reproduces training users’ trajectories verbatim when conditioned on their observed prefix, following established prefix-suffix evaluation methods^75,76. We randomly sampled 50,000 users from the training set and partitioned each trajectory into a prefix and a suffix (fourth menstrual cycle). We extracted the fourth cycle from each synthetic continuation and compared it to the user’s true fourth cycle using exact match rate (fraction of users with at least one generated continuation identical to their true fourth cycle), near-duplicate rate (fraction with normalized edit distance ≤ 0.05), and mean normalized edit distance (token-level Levenshtein distance, averaged across all users taking the minimum across five continuations per user). To control for natural similarity among users, we established a real-to-real baseline by identifying each user’s nearest neighbor within the same training cohort using cosine similarity, then computing the same metrics between each user’s true fourth cycle and their nearest neighbor’s fourth cycle.

Data Availability

The data supporting the findings of this study were provided by Clue by BioWink GmbH (Berlin, Germany) and consist of de-identified user-generated menstrual health records. Due to the sensitive nature of reproductive health data and associated contractual and ethical obligations, the dataset is not publicly available. Researchers interested in accessing the minimum dataset necessary to interpret, reproduce, or extend the analyses may submit a request to Clue by BioWink GmbH, subject to review and approval and the execution of a data use agreement outlining permitted uses, privacy protections, and restrictions on re-distribution. Requests may be initiated by contacting the authors or by emailing Clue’s research collaboration contact at research@helloclue.com. Access requests are evaluated on a case-by-case basis, and Clue generally responds to data access requests within 4–6 weeks, depending on the scope of the request and legal review. Data use is restricted to approved research purposes and governed by the terms of the data use agreement.

Code Availability

The code will be made publicly available upon publication.