Prior findings on this dataset

Every work that has used mcPHASES, as of October 2026, and what each established

Seventeen formal works and about twenty-five repositories touch the dataset. Everything physiological that is published is a pooled group contrast (glucose, wrist temperature, resting heart rate), a classifier score, or a symptom model. For this dataset, no published work reports resting-heart-rate, breathing-rate or HRV phase contrasts in native units, splits a phase contrast into person, cycle and day variation, reports a cross-metric concordance, or anchors an analysis on the 2024 progesterone metabolite (one public code repository has begun to). Within-person luteal changes in these metrics are well established in other wearable studies: sleeping pulse rate (Shilaih et al. 2017), wrist temperature, heart rate, HRV and breathing rate (Goodale et al. 2019), and lower vagal activity in the luteal phase across 37 studies (Schmalenberger et al. 2019). Ribeiro’s 2026 audit has already tested between-person heterogeneity in this dataset, for a different quantity (E3G coupling). This page is the context for the findings here; figures reproduced from other works carry their own licence, stated in the caption.

Which metrics each work used. Rows in date order; colour is the analysis type. Blue rows are the only ones that report an effect in physical units.

Inventory

#	Work	Venue, date	Type	n	Headline number
1	Lin et al. 2023	npj Digit Med, 2023-08	effect estimate	49, 149 cycles	Glucose luteal − late-follicular β = 0.15 [0.09, 0.20] mmol/L
2	Lin et al. 2024	npj Women’s Health, 2024-10	effect estimate	45, 123 cycles	Wrist temperature 34.0 ± 1.0 luteal vs 33.7 ± 0.8 late-follicular; peak 7.6 ± 4.8 d after the LH peak
3	Lin et al. 2024, CHI	CHI, 2024-05	qualitative	30	Five design requirements
4	Lin et al. 2024, IMWUT	IMWUT 8(3), 2024-09	qualitative	50	How participants made sense of multimodal tracking
5	Lin et al. 2025, IMWUT	IMWUT 9(3), 2025-09	qualitative	20	The 2024 return interval as an app study
6	Lin et al. 2025, ACM HEALTH	ACM Trans Comput Healthc, 2025-12	qualitative	50	Barriers to a holistic view
7	Lin 2025, PhD thesis	U Toronto, 2025-10	qualitative	—	Chapters 3–6 are works 4, 3, 5, 6
8	Lin et al. 2026, descriptor · PhysioNet	Sci Data, 2026-02	descriptive	42 (+20)	Repeated-measures ANOVA across phases: temperature F(3, 123) = 11.63, RHR F(3, 123) = 10.26
9	Specht et al. 2026	medRxiv, 2026-04	classifier	41	Symptoms only, four-phase leave-one-subject-out: 67.6 %, macro-F1 0.662; calendar-only 63.0 %
10	Xu et al. 2026	IEEE COMPSAC, 2026-07	predictor	—	Wearable model beats calendar baselines for next-menses timing; no numbers in the abstract
11	Wang & Li 2026, FemWear	arXiv, 2026-08	classifier	42	Nested leave-one-participant-out phase macro-F1 0.2955 vs 0.2988 baseline
12	Wang 2026, Scale-to-Dialogue	arXiv, 2026-08	classifier	42	Six-item vs three-cluster symptom questioning: QWK 0.976 vs 0.913
13	Kogelman et al. 2026	bioRxiv, 2026-08	effect estimate	37, 96 cycles	Between-person ICC: cycle length 0.26, late-follicular 0.51, luteal 0.25, menstrual 0.31
14	Ribeiro 2026	Behav Res Methods, submitted 2026-09	reanalysis audit	41	Heterogeneity of E3G coupling not above a phase-randomised null: RHR p = 0.156, temperature p = 0.874; an earlier p = 0.024 was a join artefact
15	Goyal 2026	Oxford J Student Scholarship, 2026-09	classifier	30	Bleeding-day ROC-AUC from cycle history alone 0.822; symptoms and wearables did not add
16	Nsugbe & Lin 2026	J Intell Med, 2026-09	classifier	42, 126 cycles	Multimodal − best unimodal: median +0.0156 macro-F1
17	Dutta 2026, aadya-m1	Zenodo, 2026-10	forecaster	41	An optional LH-test input narrows the 80 % next-period window from ≈ 12 to ≈ 7 days

Grey literature (repository READMEs, not reproduced): leave-one-subject-out four-phase classifiers in which cycle day carries most of the signal (macro-F1 0.643 from cycle day vs 0.363 from physiology); a 2022 → 2024 distribution shift (KS = 0.322); an LH-aligned clustering of physiology that found no consistent clusters.

What is established, metric by metric

Metric	Published finding	Source	Not published
Glucose	Biphasic over the cycle; luteal − late-follicular β = 0.15 [0.09, 0.20] mmol/L	1	Per-person size; repeatability
Wrist temperature	Group means by phase 33.9 / 33.7 / 33.8 / 34.0 °C; peak 7.6 ± 4.8 d after the LH peak	2, 8	Effect in °C with an interval; per-person shift
Resting heart rate	ANOVA across phases; lowest in follicular, highest in luteal (boxplots only)	8	Any size in bpm
HRV, breathing rate, sleep	Classifier features only	11, 16	Any phase estimate
Phase lengths	Median cycle 31 d; menstrual 6, late-follicular 11, luteal 14; ICCs 0.26 / 0.31 / 0.51 / 0.25	13	—
Symptoms	Luteal vs late-follicular on a probit scale: cramps +0.94, sore breasts +0.97, bloating +0.38; symptom ICCs 0.21–0.66	1, 13, 9	—
Heterogeneity	Between-person spread of E3G-coupling slopes not larger than a phase-randomised null	14	—
2022 → 2024	Feature distributions shift; physiology-only phase accuracy falls	grey	Any confirmatory use of 2024

Published figures worth knowing

Lin et al. 2023, Fig. 2 (CC BY 4.0). LOESS curves of daily median glucose, LH and E3G over the normalised cycle with 95 % bands: glucose falls to about 5.8 mmol/L near 20 % of the cycle and rises after the LH peak.

Lin et al. 2024, Fig. 1 (CC BY-NC-ND 4.0, reproduced without change). Nightly wrist temperature, LH and E3G over the normalised cycle. The whole temperature excursion is about 0.3 °C.

Kogelman et al. 2026, Fig. 2 (CC BY). Mean symptom intensity by phase and symptom ICCs; conditioning on phase changes the ICCs by less than 4 % except for cramps.

Published results beside the estimates here

The papers and this site use the same app-defined late-follicular phase (last day of flow to the first day of the fertile window), but it is shorter in this release (median 6 labelled days, against 7.7–11 days in the papers), the cohorts differ (49 and 45 against 42 participants) and so do the estimators (mixed model or group means against paired participant means). This is a same-direction check, not a replication.

Metric, luteal − follicular	Published	This site	Agreement
Glucose	β 0.15 [0.09, 0.20] mmol/L, 49 participants	+0.16 [0.06, 0.27], 40	Same direction and size
Wrist temperature	34.0 vs 33.7 °C, no interval, 45	+0.23 [0.12, 0.31] °C, 42	Same direction and size
Resting heart rate	F(3, 123) = 10.26, no bpm	+3.0 [2.2, 3.8] bpm, 42	Direction agrees; size unpublished
Breathing rate, HRV	classifier features only	+0.48 /min; −3.9 ms	No published estimate
Heterogeneity	E3G-coupling spread not above null	personal shift vector not detectable over 2–3 cycles	Both say person differences are hard to see with few cycles

Data hazards reported by prior users

Hazard	Who reports it	Handling here
Several Fitbit rows per participant-day; a naive join multiplies observations	Ribeiro audit (p 0.024 → 0.874)	One value per participant-day before any join
Resting heart rate 0 means missing; undocumented	Ribeiro; this project	0 treated as missing
Glucose in mg/dL for two participants	GitHub issue on the dataset repository; this project	Converted on the participant’s median
2022 and 2024 differ: no diary, PdG added, different RHR export, distribution shift	Descriptor Fig. 2; repositories; Kogelman	2024 held out
Published cohort results use 45–49 participants; 42 released	Dataset authors	Same-direction checks only
Symptom diary collected while participants saw their phase predictions	Lin 2024 IMWUT	Symptom contrasts are not blind
Phase lengths depend on the labeller	Works 1, 8, 13	Reference phase stated on every page

Licences of the material reproduced here

Source	Licence	Use on this site
Lin 2023	CC BY 4.0	Figure reproduced with attribution
Lin 2024, Lin 2026 descriptor	CC BY-NC-ND 4.0	One figure reproduced without change, non-commercial
Kogelman 2026	CC BY	Figure reproduced with attribution
Specht 2026	CC BY	Numbers only
arXiv works	CC BY 4.0	Numbers only
Nsugbe & Lin 2026	CC BY 4.0 (full text not accessible)	Numbers from the abstract only
Goyal 2026, Xu 2026	not stated or inaccessible	Numbers from the abstract only