Researchers have developed an AI-based system that automatically and objectively detects the moment when fish experience loss of equilibrium (LOE) due to temperature stress. The system combines DeepLabCut, a deep-learning AI that captures animal posture from video, with ResNet34, a deep-learning AI-based image-classification technology. Their system is expected to help predict the effects of climate change on fish. The research team was led by Specially Appointed Lecturer Tomoya Nakayama and Yoshiya Matsuo (a master's student at the time) of the Institute of Transformative Bio-Molecules (WPI-ITbM) at Nagoya University, in collaboration with Associate Professor Tatsuto Hasegawa and Takuya Kato (a master's student at the time) of the University of Fukui. Their research was published in the journal Scientific Reports on September 10, 2026.
Changes in water temperature associated with climate change threatens the fish and species that live in aquatic environments. Because fish are ectothermic animals that cannot self-regulate their body temperature, changes in the surrounding water temperature directly affect various bodily functions. Therefore, accurately evaluating the temperature tolerance of fish is essential for predicting their response to climate change.
A conventional method for investigating temperature tolerance of fish is to gradually change the water temperature, while recording video, and identify the moment when the fish can no longer maintain its balance as an indicator for LOE. Until now, however, researchers have visually examined videos to determine the moment of LOE. Inherent subjectiveness could lead to variation in results and required considerable time and effort to examine large populations.
The research group first developed a technology that photographs fish individually in separated compartments and automatically identifies the individual fish in each compartment. They then used DeepLabCut to automatically track movements at seven locations on the fish's body: the tip of the nose, the left and right fins, the center of the body, the front and rear portions of the body, and the tail. Combining the information on these movements with video information, an AI model determines the fish's state and automatically detects when a fish reaches LOE.
Across 50 individuals, the accuracy of the AI-based determinations was approximately equal to the variation of determinations made by experienced researchers, verifying its reliability. In addition, the AI produced the same result when subjected to repeated analysis of the same video. As such, the system is well suited for large-scale experiments with many videos.
Next, they investigated the tolerance to low and high temperatures in different of medaka and closely related fish species. They found that temperature tolerance differed among the strains and found a tendency for species inhabiting higher latitudes to be more cold-tolerant, with the Japanese medaka (Oryzias latipes) showing the highest cold tolerance. However, the Taiwanese medaka relative (Oryzias cabaranensis), recently described as a new species in 2025, eluded this trend, showing greater cold tolerance than would be expected based on the latitude of its habitat. This suggests that factors other than latitude may be involved in adaptation to temperature.
This demonstrates a reliable AI-method to automate the evaluation of temperature tolerance in fish with objectivity and reproducibility. They expect this system will enable larger-scale comparisons among strains and species in the future. Through such research in the future, the findings may contribute to predicting how climate change will affect fish and towards their conservation.
