The new approach accurately detects squinting in response to CGRP and amylin, and proves superior to a manual strategy.
Accurately assessing pain, including migraine pain, is crucial for developing new therapies. But current pain assessment relies primarily on a person’s verbal report and response to pain questionnaires. That makes it challenging to assess pain in people who can’t effectively communicate it and in animals, too, who obviously can’t tell us how they feel.
Now, new research led by Brandon Rea, University of Iowa, US, reveals that automated facial detection software can pick up eye squinting in mice and be used as an objective and real-time measure of pain – in this case, pain in response to calcitonin gene-related peptide (CGRP), a migraine trigger, and to amylin, a related peptide. Importantly, the automated assay performed better than manual scoring of facial features.
“This study represents a number of developments I’m very happy about,” said Jeffrey Mogil, a pain researcher at McGill University, Canada. “I’m happy people are trying to get measures of spontaneous pain – pain that emerges from within the animal – rather than responses to experimenters poking and prodding the animal from the outside. I’m happy people are automating the assessment of pain, because that’s the problem with spontaneous measures – there is a lot of manual labor that goes into it. Being able to get data more quickly will help research go faster, which puts us in a better position to make progress with new therapies,” said Mogil, who was not involved in the new research.
The study appeared in the August 2022 edition of the journal PAIN.
By machine or by hand?
While researchers are in hot pursuit of objective ways to measure pain in people, they want to do the same in animals, too. That will enable successful translation of knowledge gained from studying creatures like mice and rats into new therapies for humans.
Mogil has pioneered an approach towards a more objective assessment of pain, based on a fact that those outside the world of animal pain research may be unaware of: Rodents grimace in response to pain, just like people do. His and his colleagues’ work has shown that measuring changes in facial features, such as squinting, when a mouse grimaces is a viable way to more objectively assess pain in animals. But, so far, this strategy has relied on manual scoring of what are often subtle alterations, which is extremely time consuming.
The new study offers a streamlined way to detect squinting, but that wasn’t the initial intention of the authors. Working in the lab of Andrew Russo, a migraine researcher and study co-author, first author Rea and senior author Levi Sowers were studying migraine using a range of behavioral assays in mice. Rea was using electromyography, which records electrical activity produced by skeletal muscles, in mice.
Specifically, he was recording activity from the muscles of the eyelid, called the orbicularis muscles, to measure squinting in response to bright light. Photophobia – sensitivity to light – is a common symptom among people with migraine, making it an important area of study.
However, an unexpected finding led Rea and Sowers to reconsider their approach.
“We noticed that mice were not squinting or closing their eyes like you would expect them to in bright light,” Rea explained. “But they were squinting when they were given CGRP, a known migraine trigger – they were squinting in both the light and the dark after CGRP administration. This led us to think the squinting response wasn’t a photophobic response, but more of a pain response.”
But Rea and Sowers soon ran into problems when using the mouse grimace scale that was originally developed by Mogil as a way to explore pain behaviors: It took a very long time to score the video recordings of the grimaces by hand.
After previously showing that, in response to CGRP, orbital tightening – a narrowing of the area around the eyes, or squinting – was the largest contributor to the mouse grimace scale, Rea and Sowers decided to focus on just the squint response, working with a partner to detect squinting in a more time-efficient, automated way.
“We collaborated with a company called FaceX, which uses deep learning and shape regression models,” said Sowers. Deep learning is a subset of machine learning, while shape regression models locate the eye and present an objective area measurement of squinting through detection, alignment, and labeling of facial landmarks.
Setting up an experiment to detect squint
To test the automated squint assay, Rea and Sowers used male and female mice of two different strains (white mice known as CD1 mice, and black mice called C57BL/6J mice). To keep the mice in one spot so that they were always the same distance from video cameras, the team put the animals in a customized collar restraint, while eight synchronized cameras, each recording at 10 frames per second, would record facial expressions from different angles. Keeping the mice restrained was important.
“The collar is not that different from other methods of restraining a mouse. The animal can move its limbs inside the restraint and look around; the collar just localizes the mouse in one spot,” Rea said. “The mice act very calm, unlike in a lot of other studies, where researchers make an incision in the head and fix the skull to a bar. That’s obviously problematic when you’re looking at pain expression or responses.”
After the mice acclimatized to the restraint, the researchers made a five-minute baseline recording of squint behavior in response to normal room lighting. Those recordings were used to train the automated facial detection software to pick up squinting.
The mice then received an injection of formalin, CGRP, amylin, or phosphate-buffered saline (as a control) and returned to their cages to rest. Formalin is an inflammatory agent commonly used to study pain responses in rodents.
Following a short recovery period, the investigators again restrained the mice for an additional five minutes and made more video recordings. The recordings were then fed into the trained facial detection software to determine changes in squinting behavior. The scientists also manually scored squinting using the mouse grimace scale, as a comparison.
Better than human eyes
A greater number of training images were required for the black mice, compared to the white mice, because of the difficulty the software had differentiating between the eyelid border and the animals’ fur. Nonetheless, the software accurately tracked facial features in both strains of mice, including in response to formalin injection.
Being able to quantify pain in mice of different colors is a step in the right direction, according to Mogil.
“The system in the current study appears to work equally well in black- and white-colored mice, which is a plus. In our previous machine learning model, we only worked with white mice,” Mogil said. Mogil and colleagues have actually now developed a software platform that simplifies and standardizes grimace analyses in black mice; the paper describing this approach is now available as a preprint in bioRxiv).
After this validation of their automated approach, Rea and Sowers tested how well it detected squint in response to CGRP. Here, the automated assay did better than manual scoring did. In fact, unlike with manual scoring, it even detected squint in response to the lowest dose of CGRP that the researchers used. Interestingly, unlike the squinting responses to higher doses that all the animals showed, only female mice displayed an automated squint response at the lowest dose.
“The automation goes much further than the human eye can; the sensitivity just isn’t there on the traditional, hand-scored assays. Showing that a previously sub-threshold dose of CGRP induces a sex-specific pain phenotype highlights that the machine is doing better than what we wanted to be able to do with our eyes.”
Amylin also caused squinting that the automated assay could detect, but only in female mice.
A question of restraint
Although the findings are promising, one limitation is the need for animal restraint.
“At the moment it is very difficult for artificial intelligence to accurately measure orbital tightening while the mouse is moving around. To simplify this process, they restrained the mice,” Mogil said. “However, restraining mice can cause problems, because it produces stress, which in turn produces analgesia. And the analgesia gets rid of the very thing you’re trying to measure – the squinting.”
Mogil continued, “The way around that is habituation to the restraint. But were the mice habituated enough? It’s hard to know. From an earlier paper, the authors seem to have defined habituation as not struggling. So then the question is, Have they stopped struggling because there’s no stress, or is it learned helplessness?” The latter refers to when animals (or people) stop trying to change their behavior in response to an aversive stimulus because the stimulus seems unchangeable or inescapable.
Rea and Sowers hope they can develop the automated software for measurement of squint in mice that are free to move around. The pair also want the software to measure additional behaviors, like paw movements. But this takes time and money.
“I don’t think we ever planned on doing this when we began doing facial grimace scoring,” Sowers said. “It was much more work than we ever anticipated,” he said with a laugh. “But it was good to learn that exploring new ideas can be quite the undertaking, even when you have a company on board.”
“It also shows that even when the results aren’t what you expect, it always leads to the next finding,” said Rea. “The path to figuring things out is never straightforward.”
Lincoln Tracy is a research fellow and freelance writer from Melbourne, Australia. You can follow him on Twitter @lincolntracy.
Correction: Due to an editing error, in its original version, this news story incorrectly referred to Dr. Mogil’s previous machine learning model as one used in black mice, when this model was in fact used in white mice. We have corrected the error. We have also added mention of Mogil’s new study, available as a preprint in bioRxiv, which uses black mice.
Automated detection of squint as a sensitive assay of sex-dependent calcitonin gene-related peptide and amylin-induced pain in mice.
Rea et al.
PAIN. 2022 Aug 1;163(8):1511-19.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Sign Up For The MSC Newsletter
Share this article:
Messe Wien Exhibition & Congress Center, Hybrid, Vienna, Austria and online
UCLA Meyer and Renee Luskin Conference Center, Los Angeles, California