Data governance and annotation

The unglamorous layer that decides whether any of the measurement work survives contact with a clinical study, a regulator, or a reviewer.

Why it is load-bearing here

Three separate requirements converge on it:

  1. Longitudinal comparability. Repeated within-person measurement only works if every session is captured the same way. → Longitudinal within-person measurement
  2. Registered Report discipline. Feature definitions, QC rules, exclusions and the analysis plan must be specified prospectively and frozen before evaluation. → PCI Registered Reports
  3. New label types. The respiratory extension introduces interval categories that do not yet have an agreed annotation manual or measured inter-annotator agreement.

The session-level data contract

What a short speech session must carry, beyond the audio itself:

  • repeatable prompts (the same task, asked the same way)
  • timestamps
  • task and protocol identifiers
  • recording quality indicators
  • respiratory-event handling (how ambiguous intervals were treated)
  • links to clinical reference dates

Together these turn short sessions into a longitudinal research asset rather than a folder of recordings.

Annotation capability in this ecosystem

The DFKI collaboration is the relevant prior experience: LabelMovie (2014) covered expert video annotation, combining and filtering annotations, and bootstrapping machine annotation from manually labelled examples, with quality assurance and crowd-sourcing options.

Transfer value: annotation protocols, reviewer agreement, manual checking of automated outputs, and potential reanalysis of historical recordings.

Safeguard. Historical joint work only. The 2014-era tools are not claimed to be maintained or clinically deployable today; a present-day implementation would require its own capability, licensing, security, and governance review.

Governance constraints already fixed

  • No raw video of children in the public wiki — aggregate findings and dataset descriptions only. Applies to the 2013 educational-games archive.
  • Consent and provenance are named as first-order work-package content in the proposed programme, not an afterthought. → Programme opportunities
  • Danish data access is per-project: access, data type, analysis location, approvals and new-measurement requirements are determined each time. → Michael E. Benros

Keyword cluster from the source material

Longitudinal cohorts · data harmonisation · dataset provenance · human-in-the-loop annotation · annotation quality assurance · inter-annotator agreement · participant-level validation · reproducible analysis · measurement specification.

PCI Registered Reports · Longitudinal within-person measurement · Respiratory signal extension · Supporting collaborations · Programme opportunities