📄 health informatics

Sex and menstrual cycle differences in the mood-activity association should inform cycle-aware digital phenotyping

This study demonstrates that the relationship between physical activity and mood varies significantly by sex and menstrual cycle phase, particularly in women, indicating that digital phenotyping models must account for these biological factors to avoid biased predictions and enable personalized mental health insights.

Original authors: Delray, K., Zeitler, J. K., Hayes, J. F., Kandola, A., Keay, N., Evans, R. J.

Published 2026-07-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Delray, K., Zeitler, J. K., Hayes, J. F., Kandola, A., Keay, N., Evans, R. J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your phone is a silent, super-observant detective. It doesn't just know what you type; it watches how you move, where you go, and how you sleep. In the world of digital health, scientists call this "digital phenotyping." It's like building a digital shadow of your daily life to guess how you're feeling inside. For years, researchers have assumed a simple rule: if your shadow moves more (you take more steps), your mood is probably better. They thought this rule was as steady as gravity, working the same way for everyone, every single day. But what if that rule has a glitch? What if the connection between moving your body and feeling good isn't a straight line, but a wavy one that changes depending on who you are and where you are in your body's natural rhythm? This is the question a team of scientists set out to answer, investigating whether the "move more, feel better" rule holds up for women as it does for men, and if it shifts gears during the menstrual cycle.

The researchers, led by Kyra Delray and her team at the University of Oxford, decided to test this assumption using data from the Juli app, a digital health tool where users track their mood and connect their activity data. They looked at a massive dataset: 1,072 people (121 women and 951 men) contributing nearly 14,000 days of records. Think of it as watching a giant, living movie of thousands of people's lives, looking for patterns between how much they walked and how happy they reported feeling.

Their first discovery was that the "move more, feel better" link wasn't the same for everyone. It was significantly stronger for women than for men. If you imagine mood and activity as two dancers, they were dancing much more closely together for the women in the study than for the men. But the real plot twist happened when the scientists looked closer at the women's data. They found that the dance steps changed depending on where the women were in their menstrual cycle.

During the "early luteal phase" (the week or so after ovulation but before the period starts), the connection between activity and mood basically vanished. It was as if the dancers had stopped listening to each other; moving more didn't seem to correlate with feeling better at all. However, as the cycle moved into the "late luteal phase" (the few days right before the period), the connection snapped back into place, and it was actually the strongest it had ever been. During menstruation and the follicular phase (the time after the period ends), the link remained strong and positive.

To make sure this wasn't just a weird fluke of time passing (like "maybe people just feel better on Mondays"), the researchers created a clever control group. They took the men in the study and assigned them fake, made-up "cycles" based on the calendar month. When they ran the same analysis on the men with these fake cycles, the connection between activity and mood stayed flat and unchanging. This proved that the shifting pattern in women wasn't just about the time of month; it was specifically tied to the biological reality of the menstrual cycle.

The team also asked: does knowing about these cycle changes actually help us predict mood better? They ran a test where they tried to guess a woman's mood based on her activity. When they ignored the menstrual cycle, their guesses were okay. But when they built a model that knew which phase of the cycle the woman was in, their predictions got better in 75% of the test cases. It's like trying to predict the weather: knowing it's July helps, but knowing it's July and that a storm front is moving in makes your prediction much sharper.

The study didn't claim to know exactly why this happens—perhaps it's the shifting hormones like progesterone acting on the brain, or perhaps it's something else entirely—but they were very clear that the pattern exists. They also noted that their data came from people who were already interested in tracking their health, so it might not represent every single person on earth. Still, the message is clear: if we want our digital health tools to be fair and accurate for women, we can't just use a one-size-fits-all rule. We need to build systems that understand that for women, the relationship between moving their bodies and feeling good is a dynamic, changing story that depends on the day of their cycle.

Technical Summary: Sex and Menstrual Cycle Differences in the Mood-Activity Association

Problem Statement
Digital phenotyping increasingly relies on passive physical activity data as a proxy for mood, operating under the assumption that the relationship between activity and mood is stable across individuals, time, and physiological states. This stability is a prerequisite for Dynamic Treatment Regimes (DTRs) and Just-in-Time Adaptive Interventions (JITAIs), which use these signals to deliver personalized behavioral support. However, existing literature suggests the activity-mood association may vary by sex, and mood itself fluctuates across the menstrual cycle. The critical gap addressed by this study is whether the within-person association between physical activity and mood varies systematically across the menstrual cycle in women. If this association is not constant but rather cycle-dependent, fixed decision rules in digital health tools will be biased and suboptimal for female users.

Methodology
The study utilized daily mood and physical activity data from the Juli app, a chronic health management application. The analytic sample comprised 1,072 individuals (121 women, 951 men) contributing 13,909 person-days.

Data Processing:
- Mood: Daily self-reported on a 5-point scale (1=very bad to 5=excellent), normalized to each individual's mean as a percentage deviation.
- Physical Activity: Passive step counts and workout minutes derived from Apple Health or Google Fit. Steps were converted to approximate minutes (divided by 100) and combined with workout minutes, then normalized as percentage deviations from the individual's mean.
- Menstrual Cycle: Cycle phases were defined based on user-reported period start dates. The analysis window spanned days -14 to +13 relative to menstruation onset (Day 0). Four phases were defined: Early Luteal (-14 to -9), Late Luteal (-8 to -1), Menstruation (0 to 5), and Follicular (6 to 13). 76% of cycle onsets were algorithmically imputed based on cycle length extrapolation.
Statistical Models:
- Fixed-Effects OLS: The primary analysis used within-person fixed-effects Ordinary Least Squares (OLS) regression to estimate the same-day slope between activity and mood. Models were adjusted for day-of-week and daily weather (temperature, humidity, pressure, pollen).
- Sex Comparison: A pooled model tested for a sex $\times$ activity interaction.
- Cycle Modulation: Within the female subsample, a mood $\times$ cycle-phase interaction term was added to test for phase-specific slope variations.
- Continuous Modeling: A Fourier-basis varying-coefficient Generalized Additive Model (GAM) was employed to characterize the association as a smooth, periodic function of cycle day, avoiding discrete phase binning.
- Negative Control: A "male pseudo-cycle" analysis was conducted where male users were assigned synthetic 28-day cycles anchored to the first Monday of each month to distinguish menstrual effects from general temporal patterns.
- Predictive Validation: Repeated five-fold cross-validation assessed whether cycle-aware modeling improved out-of-sample prediction of mood from activity (measured by held-out $R^2$ ).

Key Results

Sex Differences: The within-person activity-mood association was significantly stronger in women than in men. The slope difference (men minus women) was -0.194 ( $p = 0.019$ ). Women exhibited a slope of +0.289, compared to +0.095 for men.
Menstrual Cycle Modulation: Within women, the association varied significantly across cycle phases (Omnibus interaction $p = 0.022$ $p = 0.022$ ).
- Early Luteal: The association was effectively zero (Slope = 0.00, 95% CI [-0.31, +0.30]).
- Late Luteal: The association was strongest (Slope = +0.46, 95% CI [+0.21, +0.71]).
- Menstruation & Follicular: Significant positive associations were observed (Slopes +0.36 and +0.27, respectively).
- The continuous Fourier model confirmed this pattern: the slope rises from near zero in the early luteal phase, peaks in the late luteal/menstruation window, and remains elevated through the follicular phase.
Negative Control: In the male pseudo-cycle analysis, there was no variation in the activity-mood slope across synthetic phases ( $p = 0.974$ ). This confirms the observed modulation in women is specific to the menstrual cycle rather than a general temporal artifact.
Predictive Utility: Accounting for cycle phase improved out-of-sample mood prediction in women in 75% of cross-validation splits (mean gain +0.054 percentage points in $R^2$ ). In contrast, the male pseudo-cycle model showed no improvement (0% of splits).

Significance and Claims
The paper claims that digital phenotyping systems using physical activity as a mood signal are biased if they fail to account for sex and menstrual cycle phase. The study demonstrates that:

The activity-mood coupling is not a static trait but a dynamic state that fluctuates significantly within women over the course of a cycle.
Treating this association as constant leads to systematically incorrect inferences, particularly during the early luteal phase where the link is absent.
Incorporating cycle phase as a moderator improves the accuracy of mood prediction from activity data for women, thereby enabling more equitable and personalized recommendations in Just-in-Time Adaptive Interventions (JITAIs).

Limitations and Modesty
The authors explicitly note several limitations:

Sample Bias: The cohort is self-selected (Juli users who actively track mood and cycles), limiting generalizability to the broader or clinical population.
Measurement: Cycle phases were estimated from self-reported period start dates without hormonal confirmation or ovulation markers. Mood was a single-item self-report.
Effect Size: While the direction of predictive improvement was consistent, the absolute gain in predictive accuracy was small (under 0.1 percentage points of $R^2$ ), suggesting the practical value requires prospective confirmation.
Mechanism: The study does not establish a causal biological mechanism (e.g., specific hormonal pathways) for the modulation, offering only speculative hypotheses regarding progesterone and estradiol fluxes.
Replication: The findings are based on a relatively small female subsample ( $n=121$ ), and replication in larger, more diverse cohorts is warranted.

The authors conclude that while the pattern is robust within their data, digital health tools should consider sex differences immediately and aim to incorporate menstrual cycle data where available to avoid biased predictions.

Technical Summary: Sex and Menstrual Cycle Differences in the Mood-Activity Association

More like this