🤖 machine learning

HealthCAT: An Interpretable Encoder-only Transformer Framework for Health Indicator Prediction and Temporal Interpretation of Wearable Sensor Data

HealthCAT is an interpretable Encoder-only Transformer framework that outperforms existing deep learning baselines in predicting health indicators from wearable sensor data while providing validated, time-step-level insights into the specific behavioral patterns driving those predictions.

Original authors: Xiaotong Yu, Joshua Y. Kim, HaeJin Lee, Kalina Yacef

Published 2026-07-31

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Xiaotong Yu, Joshua Y. Kim, HaeJin Lee, Kalina Yacef

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints, you are looking for clues hidden in a person's daily movements. For years, scientists have used wearable sensors—like the smartwatches on our wrists—to track our health. These devices are like super-observant spies that never sleep, recording every step we take, every beat of our heart, and every moment we sit still. Traditionally, researchers have treated this massive stream of data like a grocery receipt: they just add up the totals. They ask, "How many steps did you take today?" or "How many hours did you sleep?" These "aggregated metrics" are useful, like knowing the total weight of a backpack, but they don't tell you when the heavy books were put in or when the light snacks were eaten.

The problem is that health isn't just about the total; it's about the timing and the rhythm. Just like a song isn't just the total number of notes but the melody and the beat, our health patterns happen at specific moments. Deep learning, a type of computer brain, is great at finding these complex patterns, but it's often a "black box." It can tell you, "This person is at risk," but it can't explain why or when the clues appeared. It's like a detective who points a finger at a suspect but refuses to show the evidence. This paper, titled HealthCAT, steps in to open that black box. It introduces a new way to use computer brains that not only predicts health outcomes but also highlights the exact moments in time that matter most, turning a vague guess into a clear, evidence-based story.

The Paper's Story: HealthCAT

The researchers behind this study, Xiaotong Yu and their team, wanted to build a tool that could look at the continuous stream of data from wearable sensors and answer a very specific question: "Which exact moments in a person's day are actually driving the prediction of their health?" To do this, they created HealthCAT, a flexible framework that acts like a high-tech spotlight.

Think of a wearable sensor's data as a long, continuous movie of a person's day, recorded in tiny, split-second frames. Most computer models watch the whole movie and just give a final grade at the end. HealthCAT, however, uses a special "Encoder-only Transformer" engine. Imagine this engine as a super-smart editor who doesn't just watch the movie but can pause it, rewind it, and point a glowing finger at specific scenes. It uses a clever trick called AttentiveCAT (Attentive Class Activation Token). You can think of this as a "magnifying glass" that the computer uses to scan the movie frame-by-frame. It asks, "Does this specific second of walking help prove this person is healthy? Does this specific minute of sitting down suggest they need to improve?"

The team tested this new framework on two real-world groups of people. The first group was 206 Pacific Island adolescents (aged 10 to 16) wearing wrist devices for 7 days. The goal was to predict if they had a "Healthy Weight Status." The second group was 100 adults (aged 21 to 86) wearing sensors for 6 to 8 hours overnight to predict their sleep health, specifically looking at the Apnea-Hypopnea Index (AHI), which measures breathing interruptions during sleep.

The results were quite promising. When the researchers compared HealthCAT to other popular computer models (like Transformers, GRUs, and LSTMs), HealthCAT came out on top. On the teen activity data, it improved the prediction accuracy by up to 12% and the F1-score (a measure of how well the model balances different types of errors) by up to 17%. On the adult sleep data, it also beat the other models significantly. The paper notes that these improvements were statistically significant, meaning the results were very unlikely to be just a lucky guess.

But the real magic wasn't just in the scores; it was in the "why." To prove that HealthCAT wasn't just guessing, the researchers played a game of "hide and seek" with the data. They took the specific time steps that HealthCAT said were important and "masked" (hid) the rest of the data. They found that even when they removed 75% of the data, leaving only the tiny bits HealthCAT pointed to, the model could still predict the health outcome almost as well as if it had the whole movie. In contrast, when they randomly hid data, the model's performance crashed. This suggests that the time steps HealthCAT identified truly carry the most important clues, rather than just being random noise.

The paper also showed how these findings can be visualized in a way that makes sense to humans. Instead of a confusing wall of numbers, HealthCAT maps the "importance scores" back onto a 24-hour clock. For example, in the study of the adolescents, the system could show that for a healthy teen, the "green light" (positive evidence) for a healthy weight appeared during sustained morning activity, while for a teen needing improvement, the "red light" (negative evidence) was linked to long periods of sitting during the day. It even showed that sitting down isn't always "bad"; it depends entirely on when it happens and what else is going on.

In short, the paper suggests that by combining powerful prediction with this new "spotlight" method, we can move beyond simple daily totals. We can start to understand the specific rhythms of our lives that keep us healthy or put us at risk. The authors conclude that HealthCAT offers a practical way for researchers and doctors to see when and how our behaviors impact our health, paving the way for more personalized and effective health interventions. While the study is limited to these two specific groups and datasets, the framework itself is designed to be flexible, ready to be adapted to other types of wearable data and health questions in the future.

Technical Summary: HealthCAT

Problem Statement

Wearable sensors generate continuous, fine-grained multivariate time-series data that offer significant potential for modeling behavioral patterns associated with health outcomes. However, current deep learning approaches in this domain face a critical limitation: they prioritize predictive accuracy at the expense of interpretability. While existing models can effectively recognize activities (e.g., Human Activity Recognition) or sleep stages, they often fail to provide fine-grained, time-step-level explanations regarding when and how specific behaviors impact health outcomes.

Furthermore, health research traditionally relies on aggregated metrics (e.g., daily total steps or average sleep duration), which obscure temporal patterns and context (e.g., time-of-day routines). The lack of interpretable, temporal evidence prevents the integration of deep learning models into health research, where domain experts require validated insights to support feedback, intervention design, and the translation of data-driven findings into actionable health strategies.

Methodology

The authors propose HealthCAT, a flexible framework designed to predict health indicators from wearable sensor data while providing class-specific, time-step-level interpretability. The framework operates in three stages:

Multivariate Time Series Representation:
- Raw sensor data (e.g., triaxial accelerometers, PPG, skin temperature) are preprocessed into structured multivariate time series.
- For physical activity, Signal Vector Magnitude (SVM) or gravity-subtracted SVM (SVMgs) is computed, and intensity levels (MVPA, LPA, Sedentary) are categorized based on bout lengths.
- For sleep data, multi-sensor signals are resampled and aggregated into minute-level resolutions.
- Data is aggregated into consistent time steps (e.g., hourly for activity, minute-level for sleep) to form the input sequence.
Interpretable Extraction (Encoder-only Transformer + AttentiveCAT):
- Architecture: HealthCAT utilizes an Encoder-only Transformer to capture temporal dependencies in the multivariate time series.
- Interpretability Mechanism: The framework integrates the Attentive Class Activation Token (AttentiveCAT). This mechanism combines:
  - Class Activation Token (CAT): Computed as the element-wise product of the hidden representation and the gradient of the class logit with respect to that representation. This reflects the contribution of a time step to a specific class decision.
  - Self-Attention Weights: The CAT scores are weighted by the self-attention scores from the Transformer layers to incorporate contextual relationships between time steps.
- Output: The system generates a sequence of class-specific importance scores ( $S^c$ ) for each time step, indicating how much that specific moment contributed to the prediction of a particular health class.
Domain-Adapted Visualization:
- To make the raw time-step scores meaningful to domain experts, the framework maps the importance scores back onto domain-relevant behavioral cycles (e.g., 24-hour clock, sleep stages).
- Scores are aggregated by temporal position (e.g., hour of day) and feature buckets (e.g., activity intensity levels like MVPA, LPA, SED).
- This produces visualizations showing when and at what intensity specific patterns support or counteract a health prediction for an individual.

Key Contributions

The paper outlines three primary contributions:

Interpretable Framework: The proposal of HealthCAT, an Encoder-only Transformer framework integrated with AttentiveCAT, capable of generating class-specific, time-step-level interpretations from multivariate wearable data.
Interpretability Evaluation: A quantitative evaluation protocol demonstrating that the time steps identified by HealthCAT carry significantly more predictive value than random selection, validating the interpretability mechanism.
Domain-Adapted Visualization: A flexible method to visualize individual-level interpretations by mapping time-step scores onto domain-relevant cycles (e.g., time-of-day), facilitating the analysis of behavioral patterns.

Experimental Results

The framework was evaluated on two real-world datasets involving 306 participants:

Physical Activity Dataset: 206 Pacific Island adolescents, predicting Healthy Weight Status (HWS).
DREAMT Dataset: 100 adults, predicting Apnea-Hypopnea Index (AHI) status (Healthy vs. Unhealthy).

Predictive Performance (RQ1):

HealthCAT outperformed deep learning baselines (Vanilla Transformer, GRU with attention, LSTM with attention) on both datasets.
Physical Activity: Achieved an F1-score of 0.822 (vs. 0.759 for the best baseline) and accuracy of 0.826 (vs. 0.784).
DREAMT: Achieved an F1-score of 0.808 (vs. 0.749 for the best baseline) and accuracy of 0.843 (vs. 0.795).
Improvements were statistically significant ( $p < 0.05$ ), with gains of up to 17% in F1-score and 12% in accuracy.

Interpretability Evaluation (RQ2):

Masking Experiments: The authors tested whether time steps selected by AttentiveCAT were more predictive than random selections by masking (zeroing out) the bottom 25%, 50%, and 75% of time steps.
Findings: HealthCAT's selected time steps consistently yielded higher F1-scores than random selection across masking conditions. For instance, in the Physical Activity dataset, AttentiveCAT-guided masking significantly outperformed random masking in 10 out of 12 comparisons.
Robustness: Even under aggressive masking (75%), the model retained predictive power when relying solely on the top 25% of AttentiveCAT-selected time steps, whereas random masking degraded performance monotonically.

Visualization (RQ3):

Case studies on two participants demonstrated the ability to surface distinct behavioral patterns. For a participant in the "Healthy Fitness Zone," sustained morning MVPA and consistent LPA were identified as positive contributors. Conversely, for a participant in the "Need Improvement" category, fragmented activity and specific sedentary periods were highlighted as contributing factors. The visualization revealed that sedentary behavior is context-dependent, with some periods contributing positively and others negatively to the prediction.

Significance and Claims

The paper claims that HealthCAT moves wearable sensor analysis beyond aggregated summary metrics toward temporal patterns that support health monitoring and intervention design. Its significance lies in:

Bridging the Gap: It addresses the gap between aggregated features and the need for time-aware behavioral modeling in health research.
Validated Interpretability: Unlike many "black-box" deep learning models, HealthCAT provides time-step-level explanations that are not only class-specific but also quantitatively validated to contain predictive information.
Actionable Insights: By mapping importance scores to domain cycles (e.g., time-of-day), the framework enables researchers and clinicians to inspect when and how physical activity patterns occur. This supports the design of personalized interventions and behavior recommendations based on specific temporal evidence rather than just overall activity volume.

The authors note limitations regarding sample size and the exclusion of contextual metadata (e.g., diet), suggesting future work will focus on larger cohorts and multimodal integration. However, the current work establishes a foundational framework for interpretable, temporal health analytics using wearable data.