Livestock Health Monitoring: Alerts Are Imperfect and Health Gains Are Unproven

The use case

In June 2021, microphones were installed in 16 pens holding 256 pigs on a commercial farm in Oliveira, Minas Gerais, Brazil, and a neural network (software that learns patterns from labelled examples) sorted the recordings into coughs and other sounds¹. That is the clearest example found of the use case: a sensor turns an animal signal into a count or an alert that a farmer or vet can act on. Other examples are collars and activity monitors on dairy cows that estimate eating, rumination and heat from movement²³.

Scope note. Here AI means machine learning. The cough detector qualifies, because it is a hybrid neural network¹. For collars, the sources read do not say whether the software that classifies movement is machine learning; the 2021 collar test refers only to a "classification algorithm"². Results for such systems are therefore reported as results of sensor systems, not credited to AI. The older alternatives are human: visual observation and laboratory hormone tests (progesterone) are the reference methods against which heat alerts are judged³. Milking robots, feeders and virtual fencing appear in the farm surveys but are out of scope⁴.

Where it is running

Farm surveys show sensors in use, but small samples and loose definitions of "technology" limit what they say about AI.

The Italian sample was an online questionnaire answered by 53 farms, mostly free-stall Holstein herds in the Po Valley, so it cannot stand for Italian dairy farming⁵. The Irish sample was larger: 250 farms in the 2024 national survey, estimated to represent 11,880 Irish dairy farms⁴. The Irish collar figure is more than three years old and no newer one was found in the text read, so the Italian and Irish figures cannot be compared. No named commercial deployment with audited user numbers was identified, and no vendor user counts were used.

What the evidence shows

Heat alerts: moderate accuracy. A 2026 narrative review (not a systematic one) reports that six studies using progesterone as the reference had pooled sensitivity of about 73% (range 59 to 94%), specificity of about 92% (57 to 100%) and positive predictive value of about 70% (40 to 94%)³. Positive predictive value is the share of alerts that are right. These are measured results against a reference test; the dates of the underlying studies are not stated in the text read. The wide ranges reflect different farms, sensors and herds.

Alerts and what people see. In one study of 109 partially synchronised cows, four of six systems detected 15 to 35% more cows in heat than visual observation, with sensitivity against progesterone of 42 to 81%, and only 47% of cyclic cows showed standing heat³. On four farms, only 43% of cows with a heat alert stood to be mounted³. The likely reason is that visual checks rely on a behaviour fewer than half of cows show, while sensors read general activity, so the two methods disagree.

Outcomes. In studies cited in the review, cows with an early heat alert were inseminated on the alert in 83 to 86% of cases against 45 to 48% without one, and pregnancy at first insemination was 43 versus 29%, 51 versus 43% and 45 versus 40%³. These compare cows that did and did not show an alert, so cow health and herd factors can distort them. They concern fertility, not illness.

Cough detection. A study published on 6 May 2025 reported recall of 98.6%, specificity of 99.7%, precision of 98.8% and accuracy of 99.6% on 10,048 labelled sounds, of which 1,110 were coughs, judged against cough events annotated live by an observer (measured without an independent control group; one farm; two of the five authors are affiliated with the company PecSmart)¹. The research team did not conduct clinical assessment of the pigs¹. The cough detector's 98.6% recall measures how well it finds coughs, not how well it finds sick pigs.

Collar behaviour estimates. A 2021 test with 20 Holstein-Friesian cows compared a commercial collar with visual scan sampling: correlation was 0.85 for feeding and 0.83 for rumination, and concordance was 0.86 for both; limits of agreement were -19.60 to 17.46 minutes per hour for feeding and -15.80 to 15.00 for rumination with the updated software (measured against direct observation, no control group)². The authors describe precision as "relatively limited"². Because an hour has 60 minutes, differences of 15 to 20 minutes are large for a single hourly reading. This figure is older than three years and no newer one for the device was found.

What it takes: cost, integration, governance

Cost and integration. In the Italian survey, installation costs and poor system integration were "consistently identified as major constraints" (abstract)⁵. In Ireland, adoption was associated with herd size, farm size and family farm income, and milking robots, the technology most often cited as not adopted, had financial factors as the highest barrier⁴. Robots are a different technology from health sensors, but they show that cost sorts who adopts. No per-unit price from an official source was read, so none is quoted.

People and data. A cough count or a heat alert is a prompt, not a diagnosis. The cough study involved no clinical assessment, so someone must decide what a rising count means¹. The heat-alert outcomes cited depend on staff inseminating on the alert³.

Governance. No law, regulator guidance or standard specific to this use was read for this paper. Rules on animal health data, device approval and AI regulation have not been assessed.

Where it falls short

• Accuracy is judged against imperfect references. Observers and progesterone tests are themselves imperfect, and the best figures come from the narrowest tests: one farm and six days¹, or 20 cows²³.
• Detection is not health. No study read compared herds with and without alerts on disease, deaths, treatment or cost. The outcome evidence read concerns insemination and pregnancy³.
• Effects are uneven. Cows with retained placenta or metritis had about a 20-percentage-point lower probability of early heat detection, and cows with hyperketonaemia about 10 points lower³.
• Self-reports overstate. The introduction to a 2026 study cites a 97-farm Canadian study of automatic milking systems in which only 39% of farmers accurately perceived the change and the rest reported improvements the data did not support (date not stated; milking systems, not health sensors)⁶. The farm surveys read record use and opinion, not outcomes⁴⁵⁶.
• Sustained use is unknown. A dedicated search for failures, pauses, withdrawals and critical evaluations found nothing specific, and the survey text read reports no abandonment data. Several journal pages, including PubMed Central and ScienceDirect, could not be opened, so critical evaluations may exist that were not seen.

Phase: Pilot with results — why

The hypothesis was "In early use". The phase chosen is "Pilot with results", the lower of the two plausible phases. Only the cough detector is confirmed to use machine learning, and the evidence read for it is a single-farm test over six days¹. Collars and activity monitors are in use on some farms⁴⁵, but the sources do not say whether their software is machine learning, so they cannot count towards the phase. The share and scale of genuine AI use cannot be established from this evidence.

What to watch next

• A later adoption survey with a national sample that reports collar use, to replace the 2018 Irish figure⁴.
• A repeat of the cough-detector test on more than one farm with clinical assessment of the pigs; the 2025 study used one farm and none¹.
• A controlled comparison of herds with and without sensor alerts on disease or treatment outcomes; none was found.
• Published statements of which collar and ear-tag systems use machine learning, which would let surveys separate AI from other sensors.

Questions to ask before adopting

• Does the supplier say whether alerts come from machine learning or fixed rules, and publish accuracy measured on farms like yours?
• What share of alerts proved right in an independent test, and against what reference?
• What change in disease, treatment or losses has been measured in herds that act on alerts versus those that do not?
• Who will respond to alerts, how many a day can they handle, and what happens when staff are absent?
• What do installation, integration with existing farm software and running costs come to, and who owns the data if the supplier changes?