The research, published in Preventive Veterinary Medicine, analysed records from 1,040,253 cattle inducted into three Australian feedlots between January 2018 and December 2024. Bovine respiratory disease, or BRD, accounted for 5,813 of the 15,446 deaths recorded in the study — 37.6% of mortality. Approximately 69% of BRD deaths occurred during the first 60 days on feed, with the estimated hazard peaking at around 35 days. Most strikingly, a measure based on the historical BRD mortality associated with an animal’s source property showed the strongest association with subsequent BRD mortality.
The immediate implications are about animal health. But the study also illustrates something important about biological data: finding patterns in more than one million animals is only possible because individual records can be connected across animals, places, events and time. That raises another question — what turns a million livestock records into scientific evidence?
Scale can create an illusion of certainty. A database containing a million animals sounds inherently powerful, and it can be — but the scientific value of a dataset is not determined simply by the number of records it contains.
Imagine receiving one million livestock records containing animal identifiers, source properties, arrival dates, breed classifications, vaccination histories, health events, treatments, mortality records and production information. Before analysing anything, we need to know what those fields mean. Were definitions consistent throughout the seven-year period? Were the same coding systems used at all three feedlots? How were missing values handled? Did software systems or diagnostic practices change? Did identifiers remain stable? How were animals connected to their source properties, and how was cause of death determined? Which records were excluded from analysis, and how were derived variables calculated?
Without answers to questions like these, the number of records tells us remarkably little about their evidential quality. Large-scale biological research therefore has two problems: the first is collecting enough data, and the second is preserving enough context to know what those data actually represent.
One particularly interesting finding in the Australian study concerned the source property. The researchers used a standardised vendor Property Identification Code risk score representing historical BRD mortality at the source-property level, and it showed the strongest association with BRD-specific mortality in the study, with a cause-specific hazard ratio of 4.307 and a 95% confidence interval of 4.149–4.471.
That result is important partly because it connects information generated at different points in an animal’s history. The relevant evidence does not begin when the animal arrives at the feedlot — its history already matters. The conceptual record might therefore look more like: source property → animal → pre-entry management → vaccination → sale or direct sourcing → transport → feedlot induction → health observations → disease event → outcome.
The animal connects these events, but so do identifiers, timestamps and relationships between datasets. Lose those relationships and the history fragments. Preserve them, and it becomes possible to investigate how events occurring at one stage of the production system influence outcomes later.
This is where provenance becomes important. Data provenance describes the history of data: where they originated, how they were generated, and what happened to them before they reached their current form. In livestock systems, provenance may involve considerably more than the animal itself — a record may need to connect animal → property → management event → observation → database → transformation → analysis → result.
Suppose a variable in an analysis indicates that an animal was vaccinated before entering a feedlot. Where did that information originate? Was it entered by a producer, imported from another system, recorded during induction, or derived from a lot-level record? Was the date known, the vaccine known, the information complete? The value “vaccinated = yes” may look perfectly clear in an analytical dataset, but its meaning depends upon the chain of information behind it — a distinction that becomes increasingly important as datasets are combined.
Livestock data rarely originate in a single system. Identification may exist in one database, movements in another, veterinary treatments somewhere else. Laboratory results may be held by diagnostic laboratories, genomic information generated by specialist platforms, production records kept within farm or feedlot management software, and environmental information sourced from sensors or weather services.
Each dataset may be internally useful, and combining them can make them dramatically more powerful. But integration also creates opportunities for error: two animals can be incorrectly matched, an identifier can be reused or reformatted, a property can change ownership, a timestamp can use a different timezone, a diagnostic code can change meaning, a missing record can be interpreted as a negative result, a duplicated record can be counted twice, or a historical correction can propagate into one dataset but not another. At small scale, some of these problems can be found manually. At one million animals, they become infrastructure problems.
Absence is particularly dangerous in biological datasets. Consider vaccination: if a database contains no vaccination record, several explanations are possible. The animal was not vaccinated. The animal was vaccinated but the event was not recorded. The event was recorded in another system. The information was lost during data migration. The vaccination occurred before the period covered by the dataset. Or the source record could not be linked to the animal.
These are scientifically different states, and a blank field cannot tell us which one occurred. This is why good data architecture distinguishes, where possible, between absence of an event and absence of information about an event. The distinction sounds small, but across a million records it can materially affect analysis.
Scientific datasets often contain variables that never existed in the original operational systems — they are created during analysis. The source-property risk score used in the BRD research is a good example of the general principle: a derived variable may combine previous observations, apply mathematical transformations, or classify continuous information into categories. Once created, it can become one of the most important variables in an analysis, but it has provenance of its own.
A reproducible system should therefore be able to answer which source records contributed to a given value, which algorithm generated it and which version of that algorithm was used, what time period was included and what exclusions were applied, when it was calculated, and whether it could be calculated again. Otherwise, a derived variable risks becoming detached from the observations that created it.
Longitudinal datasets introduce another challenge: time. The Australian cattle in this study entered feedlots between 1 January 2018 and 5 December 2024. Across almost seven years, operational systems can change substantially — software gets upgraded, diagnostic terminology evolves, management practices change, new vaccines become available, staff change, data fields are added, coding conventions are revised, and sensors and laboratory technologies change. Even if a column retains the same name, its meaning may not remain completely stable.
This creates a need for semantic provenance as well as technical provenance. It is not enough to know that a field came from Database A — we may also need to know what that field meant in 2018 compared with what it meant in 2024. Controlled vocabularies, data dictionaries, schema versioning and documented transformations become part of the scientific record.
Once operational data become research data, another provenance chain begins. The Australian study used several statistical approaches, including Cox proportional hazards models, time-stratified models, Fine–Gray competing-risks regression and sequential attenuation decomposition. Those methods transform millions of individual observations into estimates that can be interpreted scientifically.
The final publication contains the results, but reproducibility requires more than the final table. Ideally we should be able to reconstruct: source records → data extraction → inclusion and exclusion → data cleaning → variable construction → analysis dataset → statistical model → result → figure or table. Every transformation should leave enough information behind for somebody to understand what happened, which is why reproducibility is ultimately an infrastructure problem as much as a statistical one.
The FAIR Guiding Principles describe scientific data as needing to be Findable, Accessible, Interoperable and Reusable. The last two principles are particularly relevant to large livestock datasets: interoperability requires information to work across systems, and reuse requires enough context for somebody other than the original data producer to understand and appropriately use it. The original FAIR principles explicitly identify detailed provenance as part of making data reusable.[4]
That matters because the future user of a livestock dataset may ask a question completely different from the one for which the data were originally collected. Operational records created to manage a feedlot can become epidemiological evidence. Diagnostic records can become surveillance data. Genotypes collected for breeding can contribute to research. Movement records can become part of outbreak reconstruction. Today’s administrative record may become tomorrow’s scientific observation.
There is another important implication. A dataset containing one million animals is not necessarily most valuable on the day it is created — its value can increase. New analytical methods appear. New biological questions emerge. New datasets become available for linkage. Machine-learning techniques improve. Historical patterns become visible only after enough time has passed. An observation collected in 2018 may become important in 2030 because it can suddenly be connected to information that did not exist when the original record was created.
But that future value survives only if the meaning of the record survives with it. A number without provenance gradually becomes archaeology — researchers have to reconstruct what it probably meant. A well-described record remains evidence.
The same issue becomes even more important when large livestock datasets are used to train predictive models. Imagine using a million-animal dataset to predict BRD risk before feedlot entry, drawing on information about source property, breed, vaccination, procurement route, season, transport and historical disease. A model could potentially identify high-risk cattle and support targeted surveillance or preventive intervention.
But now another set of questions appears. Which version of the dataset trained the model? Which animals and farms were represented? How were missing values treated, and which derived variables were used? How were training and validation populations separated, and which model version generated a prediction? If the underlying source data are later corrected, which models depended upon the previous version? The provenance chain therefore expands again: animal → record → integrated dataset → training dataset → model → prediction → decision. At each step, context can either be preserved or lost — the same principle behind <a href="”>how we think about connecting provenance across systems at The BioChain, whatever the species or sector.
The Australian BRD study demonstrates the enormous scientific value contained within operational livestock data. More than a million individual animal histories allowed researchers to examine when BRD mortality occurred and which factors were associated with increased or decreased risk. Its findings reinforce the importance of what happens before animals enter the feedlot: source-property history, procurement pathway, vaccination and other pre-arrival factors can matter to later outcomes.
But there is another lesson beneath the epidemiology. Large-scale livestock science increasingly depends upon our ability to connect information across systems, organisations, animals and time. That means the infrastructure supporting the evidence matters. Can we identify the animal, and the source? Can we reconstruct the event, and distinguish missing information from a negative observation? Can we reproduce a derived variable, identify which version of a dataset was analysed, and reconstruct the analytical workflow? And years later, can another researcher still understand what every important record meant?
A million records can reveal patterns that a thousand never could. But scale does not create trust automatically. Provenance does. The same chain of custody, applied to human and animal pathogen sequencing rather than herd records, is the subject of our piece on building an evidence chain for genomic surveillance.
Analysing over a million cattle across three feedlots between 2018 and 2024, researchers found that a risk score based on an animal’s source property’s historical BRD mortality was the strongest predictor of that animal’s own BRD mortality risk — stronger than anything observed after arrival at the feedlot.
Scale doesn’t guarantee consistency. Across three feedlots and nearly seven years, definitions, coding systems, software and diagnostic practices can all change — without knowing whether they stayed consistent, a million records may tell us far less than the number suggests.
Because a blank field can mean several different things — not vaccinated, vaccinated but unrecorded, recorded elsewhere, or lost in a data migration — and treating all of these as the same “no” can materially distort an analysis run across a million animals.
A derived variable, like the source-property risk score in this study, is calculated from other records rather than observed directly. To trust it, or reproduce it, you need to know which source records, which algorithm and version, and which time period went into calculating it.
FAIR calls for data to be Findable, Accessible, Interoperable and Reusable, and the original FAIR paper specifically extends this to the algorithms and workflows that produce results, calling for detailed provenance to make reused data reliable.
Because the provenance chain has to extend further: not just which animal and which record, but which version of the dataset trained a model, which animals and farms it included, and which model version produced a given prediction — so that if a source record is later corrected, you can identify which models and decisions depended on the earlier version.
If your organisation generates, manages, analyses or governs biological evidence and would be interested in participating in a UK or European provenance demonstrator, The BioChain would welcome the conversation.
Get in touch
Regulatory developments, technical notes and platform news — sent occasionally, straight to your inbox.