Integrating supervised and unsupervised machine learning for behavior segmentation reveals latent frailty signatures and improves aging clocks in isogenic and outbred mice

This study demonstrates that combining supervised expert-defined features with unsupervised behavioral signatures discovered by Keypoint-MoSeq significantly improves the prediction of age and frailty in mice, while revealing that these behavioral aging markers are highly strain-specific and do not generalize across different mouse populations.

The Gist

Imagine you are trying to tell how "old" and "tired" a mouse is. In the past, scientists had to play the role of a strict, tired teacher. They would watch a mouse walk around a box, squint at its back, count its steps, and give it a report card based on a checklist of 27 different things (like "is the fur messy?" or "is the spine bent?").

This method works, but it's slow, boring, and depends entirely on the teacher's mood. If Teacher A is tired, they might give the mouse a bad grade. If Teacher B is happy, they might give the same mouse a good grade. This "teacher bias" makes it hard to compare results between different labs.

To fix this, the researchers at The Jackson Laboratory built a robot teacher (a computer program) that watches the mouse videos. But here is the twist: they didn't just build one robot; they built two different kinds of robots and asked them to work together.

The Two Robots

1. The "Expert" Robot (Supervised Learning)
This robot was trained by human experts. The humans said, "Hey robot, look for these specific things: Is the mouse walking slowly? Is it hunching over? Is it turning sharply?"

• The Analogy: Think of this like a driving instructor. They know exactly what to look for: "Are you holding the wheel correctly? Did you check your blind spot?" They are great at spotting the specific mistakes we already know are bad.

2. The "Curious" Robot (Unsupervised Learning)
This robot was given a video and told, "Just watch everything. Don't look for anything specific. Just find patterns in the movement." It didn't know what "old" looked like. It just watched the mouse move and started grouping similar movements into "chapters" or "words" (called syllables).

• The Analogy: Think of this like a baby learning to speak. The baby doesn't know what a "car" or a "dog" is yet. They just hear sounds and start noticing patterns: "Oh, every time the dog barks, it wags its tail." It might discover a tiny, weird twitch in the mouse's tail that no human ever thought to write down on a checklist, but which actually means the mouse is very tired.

The Big Experiment

The researchers tested these robots on two types of mice:

• The "Identical Twins" (B6J mice): All genetically the same.
• The "Mixed Family" (DO mice): A genetically diverse group, like a human family with different heights, eye colors, and personalities.

They asked the robots to predict three things:

1. How old the mouse is (Chronological Age).
2. How "frail" or sick the mouse is (Biological Frailty).
3. How much of its life is left (Proportion of Life Lived).

The Results: Why "Teamwork" Wins

Here is what they found, translated into plain English:

1. The "Expert" Robot was usually better at telling time.
If you just want to know how many weeks a mouse has lived, the robot trained by humans (looking for slow walking and hunching) was the most accurate. It knew the classic signs of aging.

2. The "Curious" Robot was surprisingly good at spotting sickness.
When it came to figuring out how frail or sick the mouse was, the Curious Robot did just as well as the Expert Robot. It found hidden patterns—tiny changes in how the mouse moved—that the humans hadn't thought to check.

3. The "Super-Team" (Combining both) was the champion.
When they let the Expert Robot and the Curious Robot share their notes, the results were the best of all.

• The Analogy: Imagine you are trying to guess the weather.
  • The Expert says, "It's cloudy, so it might rain." (Based on known facts).
  • The Curious says, "I noticed the birds are flying low and the grass smells different." (Based on hidden patterns).
  • Together, they can predict a storm with much higher accuracy than either one alone.

The Surprising Catch: One Size Does Not Fit All

There was one big problem. The robots were great at predicting the age of the "Identical Twins," but if you took the robot trained on Twins and tried to use it on the "Mixed Family," it failed miserably.

• The Analogy: Imagine you learn to drive a Toyota perfectly. You know exactly how the brakes feel and how the steering responds. If you then get into a Ford that handles completely differently, your muscle memory fails. You crash.
• The Science: Aging looks different depending on the mouse's genes. A "frail" movement in a genetically identical mouse looks different than a "frail" movement in a diverse mouse. The robots learned the specific "dialect" of aging for one group but couldn't speak the dialect of the other.

Why This Matters

This paper is a huge step forward for aging research because:

1. It removes the human bias: We don't need tired teachers anymore; computers can do the grading objectively.
2. It finds the invisible: By using the "Curious" robot, we can find subtle signs of aging that humans miss.
3. It builds a better clock: Combining human knowledge with computer discovery gives us the most accurate "aging clock" we've ever had.

The Bottom Line: To understand how animals (and eventually humans) age, we need to listen to what the experts tell us and what the data is whispering in the background. When we combine human wisdom with machine curiosity, we get a much clearer picture of the aging process. However, we still need to build specific "clocks" for different genetic groups, just like you need a different map for every different city.

Technical Summary

1. Problem Statement

Manual Frailty Index (FI) assessment in mice is the gold standard for predicting morbidity and mortality in geroscience but suffers from critical limitations:

• Labor Intensity & Scalability: Requires expert technicians to score specific health deficits, making large-scale studies difficult.
• Subjectivity & Variability: High inter- and intra-rater variability (up to 42% of variance in some datasets), which can exceed biological variation caused by interventions like caloric restriction.
• Bias in Automated Approaches: Previous automated solutions (Visual Frailty Index or vFI) relied on supervised learning using expert-defined features (e.g., gait, posture). While effective, this "top-down" approach is anthropocentric; it only detects behaviors humans already know are associated with aging, potentially missing subtle, latent, or non-canonical signatures of decline.
• Data Waste: Supervised classifiers annotate only a small fraction of video data, discarding the vast majority of high-dimensional behavioral information.

2. Methodology

The study utilized a large, combined dataset of C57BL/6J (B6J) isogenic mice and Diversity Outbred (DO) genetically diverse mice under various dietary interventions (Ad Libitum, Fasting, Caloric Restriction).

A. Data Processing Pipeline

1. Supervised Features (S): Extracted using the JAX Animal Behavior System (JABS). These included 59 expert-defined metrics covering open-field activity, gait parameters, and morphometrics.
2. Unsupervised Features (U): Extracted using Keypoint-MoSeq (KPMS), an unsupervised behavioral segmentation algorithm.
   • Training: A KPMS model was trained on a stratified subsample of 150 videos (balanced by strain, sex, age, and diet).
   • Inference: The model was applied to the full dataset (1,138 videos) to assign every video frame to a discrete behavioral "syllable" and generate a continuous low-dimensional latent embedding.
   • Feature Derivation: Three classes of unsupervised features were generated:
     • Syllable Usage: Bout counts, total duration, and average length per syllable.
     • Syllable Transitions: Transition probabilities between syllables, aggregated into 5 "meta-syllables" via hierarchical clustering to reduce sparsity.
     • Latent Embeddings: Statistical summaries (mean, median, SD) of the continuous pose trajectories.
3. Predictive Modeling:
   • Targets: Chronological Age, Biological Frailty (Cumulative Frailty Index, CFI), and Proportion of Life Lived (PLL).
   • Algorithms: Penalized Linear Regression (Elastic Net), Random Forest, and Gradient Boosting (XGBoost).
   • Validation: Repeated nested cross-validation (10 folds, 5 repeats) to prevent overfitting.
   • Feature Sets Compared: Supervised-only (v), Unsupervised-only (uv), and Combined (bv).

3. Key Contributions

1. Hybrid Framework: Demonstrated that fusing expert-defined (supervised) features with data-driven (unsupervised) features yields superior predictive accuracy compared to either method alone.
2. Unsupervised Discovery: Validated that unsupervised segmentation (KPMS) can uncover latent behavioral signatures of aging that are missed by human-defined metrics, particularly regarding the temporal structure of behavior.
3. Community Resources: Released pre-trained KPMS models, population-specific aging clocks (for B6J, DO, and combined), and the full dataset of 1,138 open-field videos with manual FI scores to the scientific community.
4. Interpretability Bridge: Mapped high-importance unsupervised "syllables" back to canonical behaviors (e.g., turning, darting, stationary), proving that unsupervised methods capture fine-grained kinematic variations within known behaviors.

4. Key Results

A. Predictive Performance

• Combined Models (bv) are Superior: Models combining supervised and unsupervised features consistently achieved the lowest Median Absolute Error (MAE) and highest $R^2$ across all targets and strains.
  • Frailty (bvFI): MAE of 1.11 (B6J), 1.47 (DO), and 1.31 (Combined). This error margin is comparable to or better than inter-rater variability in manual scoring.
  • Chronological Age (bvFRIGHT): MAE of 10.6 weeks (B6J), 12.4 weeks (DO).
  • Proportion of Life Lived (bvPLL): Improved $R^2$ from 0.33 (supervised) to 0.37 (combined).
• Feature Set Dynamics:
  • Chronological Age: Supervised features generally outperformed unsupervised ones, likely because age tracks monotonic declines in locomotion captured well by gait metrics.
  • Frailty: Unsupervised features performed comparably to supervised features (and slightly better in some cases), suggesting frailty involves complex, multisystem reorganizations of behavior better captured by holistic segmentation.
  • Feature Importance: In combined models, supervised features remained the primary drivers of accuracy, but unsupervised "syllable usage" features provided the most significant complementary gain.

B. Cross-Strain Generalization

• Failure to Generalize: Models trained on B6J mice failed to predict outcomes in DO mice (and vice versa), regardless of whether features were supervised or unsupervised.
  • Example: B6J-trained models applied to DO resulted in MAE > 46 weeks for age prediction.
• Implication: Behavioral manifestations of aging are highly population-specific. The genetic diversity of DO mice creates a broader behavioral landscape, leading KPMS to assign distinct "syllables" to DO-specific motifs that do not exist in B6J mice.

C. Interpretability of Unsupervised Features

• Top Syllables: Permutation importance analysis revealed that the most predictive unsupervised features corresponded to recognizable motor patterns:
  • Syllables 3, 6, 25, 28: Turning left/right at various speeds.
  • Syllable 29: Darting/locomotion.
  • Syllable 35: Stationary/resting.
• Mechanism: The value of unsupervised learning lies not in discovering entirely new behaviors, but in quantifying finer-grained kinematic variations (e.g., distinguishing slow vs. fast turns) and temporal allocation (time spent in specific states) that supervised metrics might average out or miss.

5. Significance

This study establishes a highly sensitive, scalable, and non-invasive framework for objective geroscience in rodents.

• Overcoming Human Bias: By integrating unsupervised discovery, the framework captures the "unknown unknowns" of aging behavior, moving beyond the limitations of human-defined ethograms.
• Precision vs. Generality: The findings highlight a critical trade-off: while combined models improve precision within a specific strain, the lack of cross-strain generalization necessitates population-specific calibration for aging clocks.
• Future Impact: The release of tools and data lowers the barrier for automated phenotyping, enabling continuous, longitudinal monitoring of biological aging in home-cage environments, which is more ecologically valid than open-field assays.

In conclusion, the paper argues that the future of automated aging clocks lies in a synergistic approach: using supervised features to anchor biological interpretability and unsupervised features to capture the full, latent complexity of the aging behavioral repertoire.