Research note 01

A pen-and-paper mortality log beats a bird-mounted sensoron warning time.

August 2026Detection lead time12 min read

We assumed no credible published figure for detection lead time from grower-recorded data existed. We were wrong. The figure exists, it comes from eight commercial farms, and it is better than we expected.

Until recently our own website said that a credible published figure on detection lead time from manually recorded farm data did not exist. That statement was false, and we have removed it. This note explains what is actually published, what it says, and why the finding is more interesting than the claim we replaced.

The figure, and where it comes from

Gonzales and colleagues analyzed daily mortality, feed intake, and water intake taken from the production charts of eight commercial Dutch broiler farms during highly pathogenic avian influenza episodes. Using a daily mortality threshold of 0.17 percent, they recovered per-farm detection lead times of one, one, two, two, two, two, five, and five days ahead of the point at which the outbreak was reported.

The false alarm rate at that threshold was 2.0 per 100 production days. The work was first posted as a preprint and subsequently restated in the peer-reviewed literature.

What the data source actually was

This matters more than the headline number. The input was not a sensor network. It was the production chart — the daily record a grower keeps. A companion study on layer farms describes its data explicitly as “production calendars manually recorded by farmers.” The lead times above were extracted from handwriting.

Detection lead time per farm at a 0.17 percent daily mortality threshold Horizontal bar chart of eight commercial Dutch broiler farms. Lead time ahead of the farm's own reported detection was 5, 2, 2, 5, 1, 2, 1 and 2 days. All eight farms signalled. Farm 1 5.0 d Farm 2 2.0 d Farm 3 2.0 d Farm 4 5.0 d Farm 5 1.0 d Farm 6 2.0 d Farm 7 1.0 d Farm 8 2.0 d All 8 of 8 farms signalled. False alarms: 2.0 per 100 production days.
Lead time ahead of each farm's own reported detection, from daily mortality counts already written on the production chart. Eight commercial Dutch broiler farms. Preprint, not yet peer reviewed. Gonzales et al. 2023

The comparison that changed our thinking

Once we had that number we placed it next to the published lead times for instrumented approaches, expecting the sensors to win by a wide margin. They do not win at all.

Published detection lead times by data source
ApproachLead timeSpecies / settingSource
Manual daily mortality record1–5 daysBroilers, 8 commercial farms, NetherlandsGonzales et al. 2023; restated in Pathogens 2024
Manual production calendar2 d (HPAI), 5–7 d (LPAI)Layers, commercialGonzales & Elbers 2018, Scientific Reports
Manual daily mortality recordmedian 1 dayMule ducksLambert et al. 2025, Veterinary Research
Electronic feeder plus machine learning1–3 daysInstrumentedAlves et al. 2024
Per-bird wearable sensor14.9 hoursIndividual birds, before deathOkada et al. 2014

Hardware buys resolution. It does not appear to buy horizon.

A bird-mounted sensor gives fourteen point nine hours before that individual bird dies. A grower’s handwritten daily mortality column, read properly, gives one to five days at flock level. Electronic feeders with a machine learning model on top give one to three days — the same range as the paper record.

This is not an argument that sensors are useless. It is an argument about what they are for. Continuous instrumentation gives you granularity, per-house attribution, automation, and signals that mortality simply does not contain. What the published evidence does not show is that it gives you more warning time for a condition that expresses itself in mortality. If that holds, then the constraint on early warning is not sensing. It is that nobody reads the record that already exists.

Where instrumentation clearly does win

The honest counterweight. Dawkins and colleagues studied optical flow analysis of flock movement and found it detected footpad dermatitis and hockburn — conditions that produce no useful mortality signal at all. In that work optical flow outperformed water consumption, bodyweight, and mortality as a predictor.

That is the correct division of labor, and it is why our roadmap treats camera and sensor work as a later layer rather than a first product. For welfare and lameness conditions, the daily log is close to blind and a camera is not. For anything that kills birds, the daily log is already carrying the signal.

Four limitations that matter

  • All of it is avian influenza. Every lead-time figure above concerns a notifiable epidemic disease with a violent mortality curve. Endemic production problems — coccidiosis, enteritis, ascites, litter and air quality conditions — have no equivalent published lead-time literature that we could find.
  • All of it is retrospective. These are analyses of records from outbreaks that had already been confirmed by other means. The threshold was chosen with the answer visible. A prospective study, where the model runs live and its alerts are recorded before anyone knows the outcome, has not been done.
  • None of it is a US broiler house. The broiler work is Dutch, the duck work is French. US houses differ in size, stocking, ventilation design, flock programs, and record-keeping practice.
  • Eight farms is eight farms. Per-farm lead times ranged fivefold, from one day to five. That spread is as important as the mean, and with n=8 it is not characterized.

What this means for what we are building

It sharpens the claim rather than weakening it. We are not proposing that a statistical model on daily records can see things a sensor cannot. We are proposing something narrower and better supported: the signal is already in the record the grower keeps, published work has recovered days of warning from exactly that kind of record, and no product currently reads it for them on a US broiler farm.

It also tells us precisely what our own validation study has to produce to be worth anything: a prospective lead-time figure, on endemic conditions, in US houses, with the false alarm rate reported alongside it. A lead time without a false alarm rate is not a result. It is half of one.

The claim we will not make

We will not tell a grower that PoultraIQ Daily will give them one to five days of warning. That figure belongs to eight Dutch farms during avian influenza outbreaks, analyzed after the fact. Borrowing it would be exactly the kind of thing this note exists to stop.

Sources

References.

  1. Gonzales, J.L. et al. Detection of highly pathogenic avian influenza in broiler flocks using production data. Preprint, 2023. Eight commercial Dutch broiler farms; 0.17% daily mortality threshold; 2.0 false alarms per 100 production days. biorxiv.org
  2. Peer-reviewed restatement of the above. Pathogens, 2024. pmc.ncbi.nlm.nih.gov
  3. Gonzales, J.L. & Elbers, A.R.W. Effective thresholds for reporting suspicions of avian influenza in layer farms using production calendars manually recorded by farmers. Scientific Reports, 2018. pmc.ncbi.nlm.nih.gov
  4. Lambert, S. et al. Early detection of highly pathogenic avian influenza in mule duck flocks. Veterinary Research, 2025. Median 1 day. pmc.ncbi.nlm.nih.gov
  5. Alves, A. et al. Machine learning on electronic feeder data for early disease detection. Computers and Electronics in Agriculture, 2024. sciencedirect.com
  6. Okada, H. et al. Wireless sensor system for detection of avian influenza in poultry. Journal of Sensor Technology, 2014. 14.9 hours before death. content.scirp.org
  7. Dawkins, M.S. et al. Optical flow, flock behaviour and broiler chicken welfare. 2017. Optical flow outperformed water, bodyweight and mortality for footpad dermatitis and hockburn. pubmed.ncbi.nlm.nih.gov

This is the study we want to run.

A prospective lead-time measurement on endemic conditions in US broiler houses, with the false alarm rate published alongside it. If you are at an extension program or a land-grant poultry department, we would like to talk.

Request a pilot All research notes