Freyja Cohort Design Lab

ARTIFACT 03 · DATASET & PROTOCOL DESIGN

Build the ground-truth infrastructure for molecularly aware health models.

Freyja is not another cycle-tracking app. Reproduction is the first proving ground for a broader platform: combine continuous wearable physiology, independent molecular observations, personal context, and longitudinal outcomes to create training-ready datasets.

01Wearablescontinuous physiology

02Molecular anchorspecific event or analyte

03Contextbaseline + confounders

Longitudinal
ground-truth
dataset

THE LONG-TERM VISION

Molecular measurements should not stand alone.

A sensor can provide biochemical specificity. Wearables provide the continuous physiological context needed to interpret what that measurement may mean for this person, at this moment.

FREYJA'S PLATFORM THESIS

Turn sparse molecular observations into context-aware physiological intelligence.

A molecular value is often difficult to collect, temporally sparse, and ambiguous in isolation. Freyja’s long-term objective is to combine it with sleep, temperature, heart rate, HRV, respiration, activity, circadian timing, individual baselines, and known confounders—then train models that estimate state, uncertainty, and when another measurement would be most informative.

NOT A cycle calendar with better predictions → BUILDING A ground-truth, dataset, and model-training platform

01

MOLECULAR SIGNAL

Specific, but difficult and sparse

LH today; additional analytes only after the measurement, sampling, and validation requirements are understood.

02

WEARABLE PHYSIOLOGY

Continuous, but often indirect

Temperature, RHR, HRV, respiration, sleep, and activity reveal trajectories that a single biochemical reading cannot.

03

PERSONAL CONTEXT

The difference between signal and confounder

Illness, exercise, travel, alcohol, medication, circadian phase, and personal baseline can materially change interpretation.

04

MULTIMODAL MODEL

Personalized interpretation with uncertainty

Models learn temporal relationships, identify missing information, abstain when evidence is weak, and improve through versioned datasets.

WHY SENSOR FUSION MATTERS · FUTURE EXAMPLE

Cortisol is hard to measure. And even harder to interpret in isolation.

The meaning of one cortisol measurement can depend on sampling time, circadian rhythm, sleep loss, recent exercise, illness, medication, acute stress, and the person’s own baseline. A future molecular sensor could become more useful when a model evaluates that reading alongside continuous wearable trajectories and context.

Improve interpretationEstimate confidenceChoose informative sampling windowsDefine sensor requirements

ONE COHORT, A REUSABLE ARCHITECTURE

The first study starts a data flywheel.

The near-term output is not merely an ovulation feature. It is the infrastructure for collecting, labeling, aligning, and learning from multimodal human physiology.

01Collectwearables + independent molecular observations

→

02Aligntimestamps, provenance, baseline, and context

→

03Labelground truth, quality, confounders, and outcomes

→

04Train & validateprobabilistic, personalized, uncertainty-aware models

→

05Inform sensorswhich analytes, precision, and sampling cadence add value

ONE VALIDATION STORY

Three artifacts, three questions

The Cohort Design Lab closes the loop between measurement qualification, responsible inference, and prospective data generation.

01

VALIDATION LAB

Can the devices produce usable, provenance-aware data?

Completeness, environmental sensitivity, placement, metric semantics, and inter-device agreement.

Qualify the measurement layer

02

GROUND TRUTH LAB

Can the inference expose uncertainty and fail responsibly?

Posterior probabilities, abstention, confounders, missingness, stress tests, and independent anchors.

Qualify model behavior

03

COHORT DESIGN LAB

How do we generate the real prospective evidence?

Recruitment, repeated molecular observations, participant burden, data yield, endpoints, and decision gates.

Design the evidence engine

WHY REPRODUCTION FIRST

Use an accessible molecular event to prove the full learning loop.

LH can be measured repeatedly at home and aligned in time with wearable physiology. That makes reproduction an unusually practical first domain for building the collection, labeling, temporal-alignment, provenance, quality-control, and confounder-handling infrastructure required for harder physiological axes.

1Accessible anchorLH-observed transition

→

2Ground-truth datasetwithin-person trajectories

→

3Reusable platformmodels + future sensors

WHAT THE FOUNDER COHORT CREATES

✓

A versioned multimodal datasetWearable streams, LH observations, context, provenance, quality, and outcomes.

✓

A ground-truth methodologyRules for aligning sparse molecular events with continuous physiology.

✓

Models and sensor requirementsEvidence about which signals, sampling windows, precision, and confounders matter.

×

Not yet clinical validationNo diagnostic, contraceptive, stress, or treatment claim is supported.

FOUNDER COHORT · DRAFT 0.1

A small cohort, deeply characterized.

Prioritize repeated within-person observations and high-quality event labels over a large but weakly labeled sample.

25–50planned participants

≥3cycles per participant

Dailywearable collection

LHminimum molecular anchor

STUDY FLOW

From screening to expansion decision

Select a phase to inspect its purpose, procedures, and exit criteria.

COLLECTION MATRIX

Every stream has a distinct role

Signals are not collapsed into a single number; provenance and measurement semantics remain explicit.

Layer	Examples	Cadence	Purpose	Required?
Wearable physiology	Temperature, RHR, HRV, respiration, sleep, activity	Nightly / daily	Continuous indirect physiology	Core
Molecular anchor	At-home urinary LH result + timestamp	Targeted cycle window	Independent event labeling	Core
Reproductive context	Bleeding, cycle day, symptoms	Event-driven / daily	Interpretation and segmentation	Core
Confounders	Illness, alcohol, travel, exercise, sleep disruption	Daily quick check	Quality weighting and exclusions	Core
Provenance	Device, model, placement, metric identity, algorithm version	Automatic + change event	Reproducibility and harmonization	Core
Confirmatory subset	Optional PdG or clinician-selected reference procedure	Subset / escalation	Stronger endpoint characterization	Optional

DRAFT ELIGIBILITY LOGIC

Recruit for protocol feasibility, not broad representativeness yet.

✓

Include adults able to consent, use a compatible wearable, perform at-home testing, and complete repeated follow-up.

✓

Document cycle variability, medication, hormonal contraception history, known conditions, shift work, and device use.

!

Do not silently exclude complexity. Predefine whether each factor is exclusionary, stratified, or recorded as a confounder.

PARTICIPANT BURDEN

Automate the passive layer; concentrate effort around the event.

Passive wearable synclow daily effort

Daily context checkbrief structured log

LH testing windowtime-sensitive active task

Cycle closeoutreview and deviation check

SYNTHETIC PARTICIPANT

One cycle is a sequence of measurements, context, and decisions.

Switch scenarios to see how missingness, illness, and device changes alter data quality without changing the underlying protocol.

CYCLE VIEW

Complete, evaluable cycle

Evaluable

Temperature deviation Resting heart rate LH observation Context event

Normalized synthetic signals · illustrative protocol behavior only

↯

DESIGN PRINCIPLE

A missing value is not a normal value.

The cohort stores absence, reason, provenance, and quality impact explicitly. A cycle can remain useful for some analyses while being excluded from others.

UseDown-weightExclude endpointExtend follow-up

INTERACTIVE PLANNING MODEL

Enrollment is not the same as analyzable evidence.

Adjust operational assumptions to estimate participant retention, evaluable cycles, wearable nights, and molecular observations.

EXPECTED PRIMARY OUTPUT

80 evaluable cycles

from 32 retained participants across 3 planned cycles.

74%cycle yield

◎

32retained participants

◌

2,451valid wearable nights

⌁

800planned LH observations

≥2

31participants expected with ≥2 evaluable cycles

EVIDENCE FUNNEL

Where planned observations are lost

Expected values

MOST LEVERAGED ASSUMPTION

Cycle evaluability

A 5-point improvement produces the largest gain in evaluable cycles under the current settings.

EVIDENCE BEFORE SCALE

The cohort succeeds by resolving decisions, we are not lookint for an attractive dashboard.

Each endpoint should map to an operational, analytical, product, or study-design decision.

PRIMARY FEASIBILITY QUESTION

Can Freyja obtain repeated, independently anchored, provenance-aware cycles with acceptable participant burden and analyzable wearable coverage?

→

OUTPUTExpand, revise, narrow, or stop

GATE 01

Operational feasibility

Can participants complete repeated cycles and time-sensitive LH testing?

Illustrative threshold≥80% complete ≥2 cycles

If missedSimplify protocol or increase support

GATE 02

Data yield

Do wearable coverage and molecular observations produce evaluable cycles?

Illustrative threshold≥75% cycles evaluable

If missedFix collection rules or cohort assumptions

GATE 03

Signal identifiability

Can wearable patterns be evaluated against an independent event without leakage?

ReportDetection, timing, calibration, abstention

If missedReframe model or add stronger anchors

GATE 04

Expansion readiness

Which devices, variables, subgroups, and procedures should enter the next cohort?

Required outputVersioned protocol + analysis plan

If uncertainRun a focused bridge study

ENDPOINT ARCHITECTURE

Separate feasibility from model performance

A clean endpoint hierarchy prevents operational failures from being misread as algorithm failures.

01

PRIMARY · FEASIBILITY

Proportion of planned cycles meeting prespecified wearable and LH data requirements.

Cohort execution

02

SECONDARY · ANALYTICAL

Event detection, signed timing error, absolute timing error, calibration, and abstention coverage.

Model behavior

03

EXPLORATORY · SYSTEM

Device effects, missingness patterns, confounder sensitivity, subgroup heterogeneity, and participant burden.

Next design

RISK REGISTER

Failure modes become protocol variables.

GOVERNANCE BOUNDARY

What must exist before real recruitment

01

Qualified clinical and research reviewEligibility, safety, consent, and escalation procedures.

02

Privacy and data-governance planConsent, minimization, access control, retention, and deletion.

03

Prespecified protocol and analysis planVersioned endpoints, deviations, exclusions, and model lock.

04

Participant-facing supportClear instructions, contact path, burden monitoring, and withdrawal.

THE NEXT MILESTONE

Not another cycle app.
A prospective ground-truth and model-building platform.
Laying the foundation for the next generation of molecular sensors.

The Founder Cohort is the first evidence engine for Freyja’s broader vision: build versioned multimodal datasets, train context-aware physiological models, and learn how future molecular sensors can be made more interpretable alongside continuous wearable data.