Source-linked AI summary
The Metrology Atlas: describing recorded physiological data as a measurement system, with VitalDB as a first application
Andrew J. Goodwin
TL;DR
Recorded physiological data are often treated as exact, although clocks, measurement windows, quantisation, and rounding make their true times and values uncertain. The Metrology Atlas reconstructs a dataset’s measurement system and applies that approach to VitalDB, reducing median timing errors in two signal pairings while documenting data-quality issues.
Problem
Recorded physiological data are often treated as exact even though clocks, measurement windows, rounding, and storage affect the recorded time and value, and time-related analyses inherit these uncertainties.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
A shared schema allows different datasets’ measurement properties to be read, compared, and reused, with VitalDB serving as the first proof of principle.
Source: §1 · p. 2, §11 · p. 11The atlas is reconstructed from records without access to devices, firmware, or recorder PCs, and its chain to UTC is broken, so absolute times are not traceable.
Source: §12 · p. 11
Abstract
from arXiv · showhide
Every value in a recorded physiological dataset is a pair: a recorded time and a recorded value. Neither is the truth. The time comes from a clock that drifts and is corrected in jumps, and often describes a window that ended earlier. The value was quantised, rounded, sometimes heaped by hand, and measured by an instrument with its own error. We present the Metrology Atlas, a dataset-independent way of describing a recorded dataset as a measurement system, and its first application, to VitalDB, an open dataset of 6,388 surgical patients, as a proof of principle. For every variable, the atlas states what was measured, the measurement model linking it to the recorded pair, and what is known about the true time and true value given a record. Each part of the model carries its evidence: 69 studies, 49 parameters and 54 observations, each marked as measured, derived, estimated or assumed. The atlas maps 11 clocks and 14 devices into one timing network, and closes loops in that network to test its own model. Synchronisation reduces the median timing error between the EEG and ECG from 794 ms to 42 ms, and between ventilator and monitor waveforms from 378 ms to 55 ms. The atlas also documents data-quality issues, several of which correct the dataset's own documentation. It is published as a website and a single machine-readable file, with open-source code and a scheme of time-aware labels. Everything is open under CC BY 4.0 and MIT licences. The engine, schema and timing tools carry over unchanged, and the aim is to describe many other datasets the same way.
1 Motivation
Recorded physiological data are often treated as exact, although device clocks, measurement windows, rounding, and storage can make both recorded time and value uncertain. The Metrology Atlas applies a documented measurement-system approach, first to VitalDB, while keeping its models and results specific to that dataset and release.
1 Motivation
- Clock drift and resets, delayed monitor windows, rounded chart events, and reduced storage precision make records inexact, affecting analyses that relate signals in time.Examples include pulse arrival time, EEG–haemodynamic coupling, and matching an alarm to its cause.Source: §1 · p. 1
- A complete, checkable description of one dataset’s measurement system is missing, despite prior reviews of timing uncertainty and solutions to specific alignment problems.The atlas applies metrology’s language of measurands, measurement models, and stated uncertainty with evidence, reconstructing the model from records.Source: §1 · p. 1
- The Metrology Atlas is dataset-independent: its schema, engine, and timing tools are shared, while each dataset is described by its own folder.VitalDB is the proof of principle because it is open, large, and records many devices on separate clocks.Source: §1 · pp. 1–2
- The goal is a family of atlases with one shared schema so different datasets’ measurement properties can be read, compared, and reused in the same way.VitalDB’s numbers illustrate the approach, not values to assume for another dataset, which needs its own measurements.Source: §1 · p. 2
- VitalDB results in the paper match an archived release snapshot, while the living atlas may revise models and numbers as evidence is added.Superseded results remain marked on the site so quoted values can still be traced.Source: §1 · p. 2
2 Intended audience
The atlas is intended for users who need to interpret VitalDB records, build models and benchmarks, document datasets, or study data quality.
2 Intended audience
- VitalDB researchers can use the atlas to assess clock separation, numeric delays, and whether charted times were rounded.Source: §2 · p. 2
- Model and benchmark builders need time-aware labels that respect when values could first have been known and information about distributions that may be artifacts.Source: §2 · p. 2
- Dataset curators and device makers can use the VitalDB atlas as a worked documentation example, including specific corrections, and use the engine to describe their own datasets.Source: §2 · p. 3
- Metrologists and data-quality researchers may find the framing useful for other datasets because the engine is dataset-agnostic.Source: §2 · p. 3
3 The VitalDB dataset
VitalDB contains high-resolution physiological and clinical data from 6,388 noncardiac surgical patients, recorded across devices with separate clocks. The atlas organizes dataset-specific measurement information using a reusable schema and engine, with timing models, evidence, and linked documentation.
3 The VitalDB dataset
- VitalDB contains high-resolution waveforms, numerics, and clinical information from 6,388 noncardiac surgical patients, recorded by devices and hospital systems on separate clocks.The release shifts each case to a common start and shuffles case ids, so absolute dates and times of day are unavailable.Source: §3 · p. 3
4 The Metrology Atlas
- The atlas’s YAML files, one per item, generate a website and a machine-readable file, with each item page linked both ways to related items.For another dataset, the engine and tools carry over while its own variables, clocks, devices, and studies are described afresh.Source: §4 · p. 3, Figure 2 · p. 4
- The atlas separates variables, clocks, links, devices, parameters, studies, observations, methods, techniques, software, and references, with evidence status and relationships recorded for each.Parameter statuses include measured, derived, estimated, assumed, withdrawn, or superseded; observations are classed as established, likely, open, or withdrawn.Source: §4 · p. 3
- For each variable, the atlas shows where true time and value may lie given a record and traces the time stamp from UTC; its planner estimates misalignment between variables.The current release includes 136 variables, 11 clocks, 14 links, 14 devices, 49 hand-written parameters, 69 studies, and 54 observations.Source: §4 · p. 3
5 Themes of investigation
The atlas treats recorded time and value as two measurands, each assessed for accuracy and resolution. It inverts a forward measurement model, producing a distribution over true time and value when steps are many-to-one.
5 Themes of investigation
- Each axis is assessed for accuracy, the systematic distance from truth, and resolution, how finely it can be read.The studies are grouped into nine themes.Source: §5 · p. 4, Table 1 · p. 5
- Inverting a record yields a joint distribution over true time and value when many-to-one steps map the truth to an interval rather than a point.Measured parameters can undo steps such as drift; rounding and a coarse stamp grid cannot be uniquely reversed.Source: §5 · p. 4, Figure 3 · p. 5
- Figure 4 places VitalDB time stamps on accuracy and resolution axes, with arrows linking raw stamps to synchronised stamps or per-case labels.Waveform samples sit at millisecond resolution; numerics are shown in orange.Source: Figure 4 · p. 6
6 Signal synchronisation
Signal synchronisation reconstructs device timelines against a working reference, SNUADC stamp-model time. Across VitalDB, it substantially reduces median timing errors, while clock-network loop checks expose under-stated uncertainties.
6 Signal synchronisation
- Device rates are relative to other clocks that also drift, and the recorder PC’s link to UTC is assumed, so VitalDB times are not traceable to UTC.SNUADC stamp-model time is the working reference against which delays are stated.Source: §6 · p. 7
- 794 ms to 42 ms for BIS EEG against ECG, and 378 ms to 55 ms for Primus airway pressure against the monitor, are the median timing-error changes after synchronisation.The Primus comparison is leave-one-out accuracy over 1,067 recordings; its 90th percentile is 187 ms.Source: §6 · p. 7
- Weighted least-squares adjustment uses clock-network loops to estimate adjusted rates and flag bad links; median misclosures are +1.3 ppm for BIS loops and −2.9 ppm for Primus loops.Variance factors of 2 to 5 indicate that stated standard errors remain too small.Source: §6 · p. 7
7 Data-quality issues found
The atlas records both established findings and unresolved or qualified observations, including corrections to VitalDB documentation and issues in the released files.
7 Data-quality issues found
- Of 54 observations, 39 are established, 7 likely, 7 open, and 1 withdrawn; the atlas import already corrects scale factors that differ from VitalDB documentation.The VitalDB team will review the findings after the atlas is public.Source: §7 · p. 8
- File issues include recorder pauses in 29% of files, duplicated records in 8 recordings, 12 recordings stored at 100 Hz, and three arterial tracks with plethysmogram scale factors.Loaders placing samples by stamp lose about 0.48 s at each SNUADC reset.Source: §7 · p. 8
8 Software and labels
The atlas is delivered through general-purpose software and dataset-specific components, alongside time-aware labels that record when their information could have been known.
8 Software and labels
- The metrology-atlas engine validates YAML folders against a schema and compiles them into a website and machine-readable file; physiosync provides dataset-independent timing tools.physiosync models stamp clocks, returns corrected times, and adjusts clock networks to close loops.Source: §8 · pp. 8–9
- Figure 7 compares timing error per 10-minute window across the release before and after synchronisation, with medians marked by dashed lines.Its checking signals are not used in the corrections they test.Source: Figure 7 · p. 9
- vitaldb-resolution converts a recorded VitalDB value or time into a distribution over its true value using versioned model files.Its engine is general and is to be released separately.Source: §8 · p. 9
- Time-aware labels carry a span, parameters, and an available_at value for when their information could first have been known.The availability rule prevents a model built on the labels from using values before they could have been known.Source: §8 · p. 9
9 Access, licensing and use
The atlas is openly licensed and can be read, loaded as data, or applied through corrected readers. It is a research resource, not for clinical use.
9 Access, licensing and use
- The atlas text, figures, machine-readable file, and VitalDB-derived results use CC BY 4.0; code uses MIT, and recordings and per-case tables are not redistributed.The atlas, its code, and an archived snapshot are listed in Table 4.Source: §9 · p. 10
- Read variable pages for measurands, delays, resolution, accuracy, clock chains, and the synchronisation planner’s expected misalignment between variables.The planner estimates expected misalignment between any two variables.Source: §9 · p. 10
- Load the JSON to look up parameters and propagate stated distributions instead of hard-coding delays or drift rates.Source: §9 · p. 10
- Apply corrected readers in vitaldb-bench to return times on the working reference with time-aware labels attached.Source: §9 · p. 10
10 How to contribute
Contributions can update the atlas’s YAML pages, models, studies, or code while preserving evidence and a visible revision history. The atlas also identifies open questions where further evidence would be valuable.
10 How to contribute
- Contributions edit or add YAML pages through pull requests, which the checker tests for broken references, unknown fields, missing figures, and forbidden patterns.The checker is run with python -m metrology_atlas check.Source: §10 · pp. 10–11
- Value models are versioned rather than edited in place, and changed results supersede earlier studies or parameters so their history stays visible.Source: §10 · p. 11
- The atlas classifies contributions as unexplained phenomena, weakly supported explanations, untestable assumptions, or conflicts between studies.Outside evidence would help especially with the recorder PCs’ link to UTC and the hospital records’ clock.Source: §10 · p. 11
11 Beyond VitalDB
The atlas is intended to extend to other open physiological datasets using shared tools and schema, while each dataset supplies its own registry and fitted models. A shared schema can also support comparisons across datasets.
11 Beyond VitalDB
- The schema, engine, timing tools, rounding-model engine, true-value distributions, and model versioning transfer, while variable registries, recording rules, and fitted models are dataset-specific.A separate dataset-independent value-model package is planned after a second dataset tests which assumptions generalise.Source: §11 · p. 11
- Findings such as early-window numerics, rounded charted times, and converters with fewer bits than documented are likely to recur in other datasets.Source: §11 · p. 11
- A shared schema makes it possible to compare datasets on the same terms, including how well two signals can be aligned in each.Curators and researchers are invited to use the engine and VitalDB folder as a template.Source: §11 · p. 11
12 Limitations and related work
The atlas reconstructs measurement models from records, with limits on access to device information and on absolute-time traceability. It builds on metrology, coarse-data, and synchronisation work while describing a dataset as a complete measurement system.
12 Limitations and related work
- Models are inferred from records without access to devices, firmware, or recorder PCs, and have not been reviewed by device makers.Source: §12 · p. 11
- Absolute times are not traceable because the chain to UTC is broken at its first link.Population results describe the release rather than an individual recording; per-setup or per-recording estimates are given where supported, and some findings are preliminary.Source: §12 · p. 11
- The work builds on the GUM and VIM uncertainty framework, coarse and heaped data methods, and network time synchronisation.It extends prior work on clinical database timing, EEG synchronisation, physiological-data quality, pulse arrival time, and high-frequency signal storage; the authors know of no prior open dataset described as a complete measurement system.Source: §12 · p. 11