Distributed intelligence, not a database

UK Biobank asked half a million people to trust an institution with their most intimate biological data. That trust has been tested. We’re building something different — where the data never leaves you.

A generation primed by Cambridge Analytica and NHS data-sale rows does not need another citizen-science biobank that quietly ships the asset to a server. Chronobiobank is research infrastructure built for people who want population insight without surrendering personal biology.

Apple made on-device federated learning a competitive advantage, not a compliance checkbox. The model trains locally; only learned weight updates aggregate centrally. Your photos, typing patterns, and health metrics never leave the phone.

Apple’s servers do not process your intimate traces — compute is distributed across a billion devices they do not pay to run. A traditional biobank pulls every sleep log to a central server, cleans it, stores it, and queries it. Participant count scales storage and compute brutally.

• Body-clock state on the edge

  Deepdose computes BTI, proxy DLMO, and dosing windows on the participant’s phone or wearable path — not in a central warehouse.

• Population patterns from model updates

  Chronotype and timing–outcome relationships are learned by aggregating federated weight deltas or privacy-preserving summaries — not raw longitudinal sleep JSON.

• Researchers query the model

  Licensed partners interrogate population-level inference and pseudonymised outcome aggregates — not a queryable lake of individual nights.

• Central infra is coordination

  The Chronobiobank runs consent governance, secure aggregation, and audit — not petabyte storage and batch ETL on intimate biology.

A federated chronotype model must learn phase relationships over weeks and months, not single timepoints. That is a genuinely novel research problem — publishable, partnerable, and potentially patentable: a population chronotype inference model that improves with scale without centralising intimate traces.

Participants can see what lives on their device and delete it. Consent stops being abstract policy and becomes something tangible.

Hybrid Tier B/C paths exist because polypharmacy cohorts skew older and less connected — federated where capable, privacy-preserving upload elsewhere. Designed from day one, not bolted on later.

1. Tier A

   Federated

   Modern phone · Oura / Whoop · reliable Wi‑Fi

   On-device training in the background. Only encrypted weight updates leave the phone — never raw sleep JSON.

   Leaves the edgeGradient contribution per learning round

   01
2. Tier B

   Privacy upload

   Older smartphone · patchy charging · intermittent sync

   When federated training isn’t viable, we queue minimum sufficient statistics — phase features, timing shift, outcome label — with differential-privacy noise before upload.

   Leaves the edgeDP-noised feature bundle when online

   02
3. Tier C

   Assisted

   Care home · carer proxy · questionnaire-only

   Human-mediated entry still improves population models. Pseudonymised prescribing outcomes join the same coordinator as federated rounds — via coarse bands, never names.

   Leaves the edgePseudonymised outcome record only

   03

• Apple HealthKit + Core ML

  On-device health metrics with local inference — the consumer proof that federated health learning works at scale.

• Federated cohort learning

  Google’s FLoC was abandoned for ads, but secure population-level inference without centralised behavioural lakes survived in the research stack.

• OpenMined / PySyft

  Open-source privacy-preserving ML infrastructure built for health research — production-grade federated training and secure aggregation.

• HDRUK Trusted Research Environments

  Not fully federated yet — but the UK’s national move toward governed research access without shipping raw NHS data to every analyst.

• Sleep & phase

  Sleep onset, wake, deep/REM duration, proxy DLMO, social jet lag, and circadian score — from wearables, phone, or TipTraQ validation.

• Medication context

  Medication code, prior vs recommended timing, adherence proxy, and whether the clinician accepted or modified the window.

• Outcomes

  Blood pressure, HbA1c, symptom scores, and adverse events — linked to timing shifts when patients and clinicians record them.

• Patient clinical store

  Raw sleep logs, DLMO estimates, and prescribing data — Supabase RLS; readable only by the patient and consented clinicians.

• Learning coordinator

  Round metadata, aggregated contributions, and global model versions — internal only; no patient_id; never licensed to third parties.

• Chronobiobank licensing store

  Pseudonymised outcome rows and population aggregates — readable only under active data license with full audit trail.

• Patient app

  GetsBTI payload, dosing windows, own clinical data

  NeverGlobal model weights or other participants’ data

• Clinician triage

  GetsPatient BTI, device sync status, recommendation history

  NeverChronobiobank aggregates or training internals

• Learning coordinator

  GetsAggregated contributions per round

  NeverIndividual gradients before secure aggregation

• Enterprise licensee

  GetsFiltered pseudonymised records and cohort statistics

  NeverRaw sleep, re-identification bridge, or model checkpoints

Version one trains a small timing–outcome model across chronotype band, social jet lag, timing shift minutes, medication cluster, and adherence. Capable devices train locally; the coordinator merges; the BTI engine deploys a new model version tag. Patients still see only plain-language window guidance.

1. NowConsent-gated pseudonymised outcome ingest and enterprise licensing
2. NextCapability routing and anonymous BTI telemetry
3. ThenPrivacy upload path for intermittent devices
4. ScaleOn-device federated rounds for timing–outcome v1

• Clinical care

  Share BTI and device status with your linked clinician. Required for prescribing workflows.

• Model improvement

  Help train privacy-preserving timing models. Raw sleep never leaves Tier A devices; withdraw any time.

• Research licensing

  Contribute pseudonymised outcomes to NHS population analytics, pharma R&D, or academic studies — each purpose is explicit.

UK GDPR-shaped workflows · immutable consent audit log · Chronobiobank isolation: UI never exposes model weights.

Deepdose is the working prototype — patient app, wearable ingest, BTI engine, consent-gated Chronobiobank ingest, and enterprise licensing dashboard. The partners below are the credible London stack to scale federated rounds, not claimed relationships.

The prototype is DeepDose

Wearable pull-sync, proxy DLMO fusion, clinician triage, pseudonymised Chronobiobank ingest, and licensed enterprise analytics already ship in the codebase. Federated rounds are the next layer on architecture that is live today.