Machine-Learning-Guided Video Analysis Identifies Sound-Evoked Pain Behaviors from Facial Grimace and Body Cues in Mice

This study develops and validates a machine-learning-guided video analysis framework that quantifies sound-evoked pain in mice by detecting specific facial grimaces and body posture changes, thereby providing an objective tool to investigate the mechanisms of auditory pain.

The Gist

Imagine you are trying to figure out if your pet mouse is in pain. You can't ask them, "Does this loud noise hurt your ears?" and they can't tell you. Usually, scientists have to guess based on how the mouse moves or if it flinches, which can be tricky and sometimes biased.

This paper introduces a high-tech "pain translator" for mice. It's like giving the mouse a voice without them saying a word. Here is how they did it, broken down into simple stories:

1. The "Smart Camera" Detective

The researchers built a special room with a camera that never blinks. They trained a super-smart computer program (called Deep Learning, think of it as a digital detective) to watch the mice 24/7.

This detective doesn't just watch; it draws a skeleton on the mouse's face and body in real-time. It tracks 13 specific points, like the tip of the nose, the corners of the eyes, and the tips of the ears.

• The Analogy: Imagine a video game character where you can see all the joints moving. The computer is watching these joints to see if the mouse is "grimacing" (making a pain face) or "hunching over" (curling up because it hurts).

2. The "Migraine Test" (Calibrating the Machine)

Before they could trust the computer to detect pain from sound, they had to teach it what pain looks like. They used a known painful situation: giving mice a shot that causes a migraine-like headache (using a molecule called CGRP).

• The Result: The computer watched the mice and noticed a pattern. About 30 minutes after the shot, the mice started squinting their eyes, flattening their ears against their heads, and hunching their bodies.
• The "Pain Score": The computer combined all these tiny changes into a single "Pain Score." It learned that if the score drops below a certain line, the mouse is definitely in pain. They even found two levels: a "ouch" level and a "really bad" level.

3. The "Loud Noise" Experiment

Now that the computer knew what pain looked like, they tested the big question: Does loud noise hurt mice?

They put mice in the room and played sounds ranging from a quiet whisper (70 decibels) to a jet engine taking off (120 decibels).

• The Discovery: When the noise got above 100 decibels (which is like a jackhammer or a rock concert), the mice's "Pain Scores" dropped. They started grimacing and hunching just like they did during the migraine test.
• The Conclusion: Yes, loud noise actually causes pain in mice, not just a startle reflex. It's not just "that's too loud"; it hurts.

4. The "Deaf Mouse" Proof (The Smoking Gun)

To prove that the pain was coming from their ears and not just the noise vibrating their skin, the researchers used a special group of mice. These mice have a broken part in their inner ear (they lack a protein called TMIE), which means they are effectively deaf to sound. They can't "hear" the noise at all.

• The Test: They played the same loud noises for these deaf mice.
• The Result: The deaf mice didn't care. Their "Pain Scores" stayed normal. They didn't grimace or hunch.
• The Lesson: This proved that the pain comes from the sound being processed by the ear. If the ear can't hear the sound, the brain doesn't feel the pain. It's like turning off the microphone; if the microphone is broken, the speaker doesn't get feedback.

Why Does This Matter?

Think of this like a new universal translator for animal welfare.

• For Science: It helps us understand why some people (and animals) have "painful hearing" (hyperacusis), where normal sounds feel like needles in the ears. This could lead to better treatments for migraines and hearing disorders.
• For Ethics: It gives scientists a way to measure pain objectively without guessing. If a noise experiment causes pain, they can stop it sooner, making research more humane.

In a nutshell: The researchers taught a computer to read a mouse's "pain face" and "pain body language." They proved that loud noises really do hurt mice, but only if the mice can actually hear them. It's a giant leap toward understanding and treating sound-induced pain.

Technical Summary

1. Problem Statement

• The Clinical Gap: Humans can experience pain triggered by sound (sound-evoked pain), ranging from extreme loudness (acoustic trauma) to "pain hyperacusis" where typically tolerable sounds become painful. The underlying mechanisms of this phenomenon are poorly understood.
• The Methodological Gap: There is a lack of sensitive, objective, and high-throughput behavioral methods to measure sound-evoked pain in animal models. Existing methods often rely on reflexive withdrawal (which measures sensitivity, not pain) or require multiple camera angles to capture both facial expressions and body posture simultaneously.
• The Objective: To develop and validate an automated, machine learning-based framework using a single-camera setup to quantify persistent pain in mice by analyzing facial grimace and body position changes during sound exposure.

2. Methodology

The study employed a deep learning approach to analyze video recordings of freely moving mice.

• Experimental Setup:

  • Mice were placed in a custom clear chamber with scent holes on one side to encourage a profile view relative to a single camera (GoPro Hero 11).
  • Validation Model: Migraine pain was induced via intraperitoneal injection of Calcitonin Gene-Related Peptide (CGRP), a known painful stimulus with a predictable peak at ~30 minutes.
  • Sound Exposure: Mice were exposed to an interleaved protocol of broadband sound (2–20 kHz) at varying intensities (70–120 dB SPL), alternating with 2-minute "No Sound" periods.
  • Genetic Model: Tmie knockout mice (Tmie-/-), which lack functional cochlear mechano-transduction, were used to determine if inner ear function is required for sound-evoked pain.

• Machine Learning Pipeline:

  • Tool: DeepLabCut (DLC), a deep neural network for pose estimation.
  • Training: The network was trained on 816 annotated frames to identify 13 specific landmarks on the mouse face and neck (e.g., nose tip, eye corners, ear tips, neck).
  • Feature Extraction:
    • Facial Grimace (6 parameters): Ear Ratio, Ear Position, Ear Tip Tilt, Eye Ratio, Snout Position, and Mouth Position.
    • Body Position (4 parameters): Relative Nose Tip Position (exploration), Percentage of Nose Tip in Top 1/3 (rearing), Vertical Line Crosses (turning/locomotion), and Face Inclination (hunching).
  • Scoring: Individual parameters were converted to percentage changes from baseline and summed to create two composite scores: Facial Grimace Score and Body Position Score.

3. Key Contributions

• Single-Camera Multimodal Analysis: The study successfully demonstrated that a single camera angle can simultaneously capture and quantify both facial grimace and complex body posture changes, eliminating the need for multi-camera setups previously required for comprehensive pain assessment.
• Definition of Pain Thresholds: By analyzing the CGRP-induced migraine model, the authors established two distinct pain thresholds (Level 1: Non-Peak; Level 2: Peak) based on composite scores, providing a quantitative scale for persistent pain.
• Automated High-Throughput Framework: The method reduces observer bias and analysis time, allowing for the processing of thousands of video frames to detect subtle behavioral changes associated with pain.

4. Key Results

A. Validation with CGRP-Induced Migraine

• Behavioral Changes: Mice injected with CGRP showed significant changes in both composite scores compared to saline controls.
• Peak Detection: An empirical pain peak was identified between 27–32 minutes post-injection.
  • Facial Grimace: Significant reduction in Eye Ratio and Mouth Position during the peak.
  • Body Position: Significant reduction in rearing ("Percent in Top"), turning ("Vertical Line Crosses"), and an increase in hunching ("Face Inclination").
• Thresholds: Two pain levels were defined:
  • Level 1 (Non-Peak): -19.04% (Grimace) and -132.4% (Position) change from baseline.
  • Level 2 (Peak): -40.62% (Grimace) and -179.5% (Position) change from baseline.

B. Sound-Evoked Pain in Wild-Type Mice

• Intensity Dependence: Sound exposure at 100 dB SPL and above elicited significant changes in both Facial Grimace and Body Position scores.
• Threshold Exceedance: At 100–120 dB SPL, the behavioral scores surpassed the Level 1 pain threshold established by the migraine model, indicating that loud sounds induce a pain state comparable to migraine.
• Reversibility: Behavior returned to baseline during "No Sound" periods, except after 120 dB SPL exposure, where body position changes persisted, suggesting potential cochlear damage or prolonged distress.

C. Role of Cochlear Mechano-Transduction (Tmie Knockout)

• Loss of Response: Tmie-/- mice (lacking functional hair cell transduction) did not exhibit significant changes in Facial Grimace or Body Position scores at any sound intensity, even up to 120 dB SPL.
• Conclusion: Normal cochlear mechano-transduction is strictly required to generate sound-evoked pain behaviors. This rules out the hypothesis that pain is caused solely by mechanical stress on the tympanic membrane or middle ear stretch receptors; the signal must be transduced by the inner ear.

5. Significance

• Mechanistic Insight: The findings confirm that sound-evoked pain is a central phenomenon dependent on functional cochlear input, distinguishing it from simple mechanical trauma to the ear.
• New Animal Model: This study provides the first robust, quantitative animal model for pain hyperacusis (sound-evoked pain), a condition previously difficult to model due to the lack of objective behavioral metrics.
• Clinical Translation: The developed framework offers a sensitive tool for screening analgesics and investigating the peripheral and central neural circuits responsible for sound-evoked pain, potentially leading to better treatments for hyperacusis and migraine.
• Methodological Advancement: The integration of facial and body metrics into a single composite score via deep learning sets a new standard for automated, unbiased pain assessment in preclinical research.