Feed intake is the obvious signal.It is the one that failed in the house.

Ask a grower what tells them something is wrong and feed comes up quickly. Birds off feed is the oldest observation in the business, it precedes almost everything else, and it fits the way a house is walked. It is also the variable that has performed worst in the one attempt we can find to build a commercial broiler alarm on it. This note sets out both halves of that, because the gap between them is the whole practical question, and because it runs against the version of the story that is easier to sell.

Why feed is the intuitive choice

The physiology is not in dispute. A bird under challenge reduces intake before it shows a clinical sign, and it reduces intake before it dies. Regulators encode this. The EU Annex II early-detection criteria trigger on a feed or water intake drop of more than 20 percent, alongside a mortality rise above 3 percent per week. The Dutch national scheme requires veterinary consultation on a 5 percent or greater reduction in feed or water per day over two days, in parallel with a mortality rule of 0.5 percent per day for two consecutive days in broilers older than ten days.

So the belief that feed carries signal is not folklore. It is written into statutory detection criteria in two jurisdictions. The question a product has to answer is narrower and harder: how often does acting on that signal put a grower in the right house, and how often does it waste the trip.

The 17-day result, and what it was measured on

The strongest positive finding for feed intake is Alves et al. 2024, in Computers and Electronics in Agriculture. Across 95,711 pure-line broilers in 146 trials run between 2017 and 2022 in Oklahoma, daily feed intake diverged 17 days before the mortality event, at p of 0.0028 or better. The behavioural traits moved in a readable sequence: time spent feeding at seven days out, meal length at eight, inter-meal interval and feeding rate at ten, visit-average intake at fifteen, number of visits to the feeder at seventeen.

That is a real, peer-reviewed, well-powered result and it deserves to be taken seriously. It was also produced in experimental pens using RFID individual feeders, one bird at a time. The best classifier in that study reached specificity of 0.952 with sensitivity of 0.643, which is to say it missed roughly a third of events even under those conditions. The authors themselves ask for feasibility and cost-effectiveness research in commercial settings. Individual-bird feeding behaviour and house-level bin draw-down are not the same measurement, and treating a 17-day lead from the first as evidence about the second is the single most common overreach we see quoted from this literature.

The commercial null result

The counterweight is a 2023 bioRxiv preprint, which we label a preprint every time we cite it. Gonzales and colleagues took production charts from eight commercial Dutch broiler farms with nine infected sheds, flock sizes from 15,000 to 63,000 birds, during H5N1 and H5N8 episodes, and tested mortality, feed and water on the same data with the same method. Their sentence on feed and water is short. Reliable thresholds for food or water intake could not be established.

Feed was sensitive. It caught every infected shed. It also raised 37.2 false alarms per 100 production days, which is more than one wrong answer every three days on a single house. Mortality on the same charts caught the same nine sheds at 2.0, with one to five days of lead over actual detection. Sensitivity was never the constraint. Specificity was.

Feed intake caught every infected shed and was still unusable. An alarm that is right about the flock and wrong about the day is not an alarm.

The duck evidence, at its correct weight

There is a peer-reviewed result that points the other way, and it should be reported. In Pathogens in 2021, the same research group examined 53 non-infected and 13 infected commercial duck flocks across the Netherlands, Germany, the United Kingdom and the United States. Where feed and water data were available, and they were available for 38 non-infected and only six of the infected flocks, the alarms based on those production parameters were raised one to two days before the mortality alarm.

One to two days is a useful margin. It is also ducks rather than broilers, six infected flocks with the relevant data, and the thresholds the authors derived, a feed drop of 7 g per duck per day and a water drop of 14 mL per duck per day, were described as sensitive with poor specificity. That is the same trade-off the broiler preprint ran into, stated more gently.

The measurement problem underneath all of it

There is a plainer obstacle that rarely gets discussed. In a January 2024 survey of Tennessee broiler producers, an unpublished MS thesis with 77 respondents from 531 contacted, a 14.5 percent response rate, 95 percent of them contract growers, 88.7 percent reported an automated feed system and 85.2 percent a water flow meter. Automated feed inventory was reported by 11.9 percent.

An automated feed system moves feed. It does not weigh what the birds ate yesterday. Getting a daily feed number requires either bin load cells, whose vendor documentation describes storing 24-hour feed usage for each of the past four days and states no accuracy specification, or in-line feed metering, which one Australian project reported as rated above 98 percent accuracy once calibrated across eight sheds. Both are new hardware on the farm. On a house with bulk deliveries and nothing else, daily feed intake is a derived and lumpy number, and a lumpy input will not fix a specificity problem.

The ratio does not rescue it

The water-to-feed ratio partly normalizes out the growth curve, which is why it gets reached for next. It does not settle the question. Breeder guidance gives 1.8 to 1 for bell drinkers and 1.6 to 1 for nipple drinkers without cups at 21 degrees Celsius. An Australian industry report gives 1.8 to 2.0 to 1 as normal. Measured commercial values in an Arkansas MS thesis ran from 1.93 at day 11 to 1.72 at day 46, with a seasonal spread from 1.724 in winter to 1.838 in summer. The direction of a disease deviation is contested as well: the Australian report treats a widening ratio above 2.0 as a possible health challenge, while a doctoral dissertation at Auburn recorded the ratio decreasing during a disease event because water fell further than feed.

We do not publish a single normal ratio, because the published sources do not agree on one and a fixed band would be wrong for most flocks on most days. What is defensible is deviation from the flock’s own trajectory.

Where this leaves the 17 days

The 17-day figure is real. It is a property of individually monitored birds in research pens, and nobody has shown it survives the move to a bin, a feed line and 30,000 birds sharing them. Treat it as a statement about what the animal does, not about what the farm can see. What does transfer, on the same evidence, is duller and more useful: a daily log, taken at the same time each day, read against where that flock should be. That is the input the commercial dataset actually validated, and mortality is the column in it that worked.