Study: AI-Powered Camera System Predicts Nearly 95 Percent Of Thoroughbred Foalings
An estimated 70 to 80 percent of foalings occur at night, requiring breeding farms to maintain continuous surveillance during foaling season — a practice researchers say carries substantial labor, economic, and psychological costs.
Researchers in Japan say a combination of thermal imaging and artificial intelligence could eventually lower that burden.
A study published July 17 in the journal Animals tested a completely non-contact foaling prediction system on 115 Thoroughbred mares at 13 commercial breeding farms in Hokkaido, Japan. The system detected 94.8 percent of foalings and, in detected cases, provided an average lead time of 89 minutes.
The system, called Umamori, uses a thermal infrared camera paired with a conventional visible-light camera mounted above the mare's stall. Rather than requiring a wearable sensor or an invasive device, the cameras continuously monitor the mare while software analyzes several physiological and behavioral indicators.
Researchers found those indicators tended to emerge in two distinct phases.
Locomotor activity and body surface temperature began increasing roughly 70 to 90 minutes before foaling. More obvious behavioral changes occurred later: Tail-raising increased significantly beginning about 45 minutes before foaling, while changes between standing and lying down increased significantly beginning about 25 minutes before delivery.
The researchers tested two versions of the prediction model. Using only locomotor activity and surface temperature, the system detected 92 of 115 foalings, or 80 percent, with an average of 186 minutes between the first alert and foaling.
When posture changes and tail-raising behavior were added, the system detected 109 of 115 foalings, or 94.8 percent, and the average interval between detection and foaling narrowed to 89 minutes. Two of those mares, however, were not detected until the time of foaling, while four mares were never detected before delivery.
The model had previously been trained using an independent dataset of 53 foalings and was applied to the 115 mares in the new study without additional training or parameter tuning.
The system wasn't immune to premature alerts. About 28 percent of mares generated at least one notification two days before foaling and about 41 percent generated one the day before. Researchers said those alerts were generally isolated, while alerts became more frequent and sustained as delivery approached.
The authors cautioned that the study was conducted in one region of Japan and monitoring occurred only while mares were stalled. Further validation would be needed under different climates and management systems.
Two study authors are employees of Noritsu Precision Co., which developed the Umamori system. The researchers reported that the study received no external funding.
This story was originally published by Paulick Report on Aug 14, 2026, where it first appeared in the Horse Care section. Add Paulick Report as a Preferred Source by clicking here.
