🤖 machine learning

Autoencoder Architectures for Athlete Performance Scoring from Wearable Telemetry

This paper evaluates five unsupervised dimensionality reduction models for converting high-dimensional wearable telemetry into a single scalar athlete performance score, finding that deep autoencoders achieve the best balance between reconstruction accuracy and latent score interpretability, with running pace, aerobic decoupling, and average heart rate identified as the dominant physiological drivers.

Original authors: Mateusz Kubita, Jan Zubalewicz, Krzysztof Siwek

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mateusz Kubita, Jan Zubalewicz, Krzysztof Siwek

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ 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 running coach. Every day, your athletes wear smartwatches and chest straps that act like tiny, over-enthusiastic reporters. They shout out thousands of data points: how fast they ran, their heart rate, how much uphill they climbed, their cadence, their weight, and even how their heart rate drifted compared to their speed over time.

The problem isn't collecting this data; it's making sense of it. A coach can't stare at a spreadsheet of 45,000 rows to figure out who is actually performing well. They need a single, clear number—a "performance score"—that summarizes everything.

This paper is about building a smart machine that can read all those noisy reports and boil them down into that single, meaningful number.

The Challenge: The "Too Many Variables" Puzzle

The researchers started with nine different pieces of information (like pace, heart rate, and elevation). They wanted to compress these nine variables into just one number (the "latent score") that represents an athlete's performance.

Think of it like trying to describe a complex movie in a single sentence. If you just say "it was good," that's too vague. If you list every scene, that's too long. You need a summary that captures the essence without losing the important details.

The Contenders: Five Different "Summarizers"

The team tested five different computer models (architectures) to see which one was the best at this summarizing job:

Simple AE: A very basic, one-step summarizer.
Medium AE: A slightly deeper summarizer with a bit more brainpower.
Deep AE: A complex, multi-layered summarizer (the "heavy lifter").
PCA: A classic, mathematical method that looks for straight-line patterns (like drawing a straight line through a cloud of dots).
VAE: A probabilistic summarizer that tries to group similar athletes together in a "cloud" of possibilities.

The Twist: Accuracy Isn't Everything

Usually, when we test these machines, we just ask: "How well can you rebuild the original data from your summary?" (This is called reconstruction error). If the machine can rebuild the data perfectly, it wins.

But the researchers realized that's not enough. A machine could be great at rebuilding the data but produce a summary score that makes no sense to a human. For example, it might rank an athlete as "great" just because they were heavy, ignoring that they were actually tired.

So, they created a two-part test:

The Rebuild Test: How well did the machine reconstruct the original data?
The "Make Sense" Test: Does the summary score actually reflect what we know about running? (e.g., Does a high score correlate with a faster pace or a healthier heart rate drift?)

They combined these two tests into a single "Composite Score" to pick the winner.

The Results: Who Won?

The Deep Autoencoder (Deep AE) won the competition.

It was the best at rebuilding the data (lowest error).
It was the best at creating a score that made sense.

The Surprise: The researchers found that the "bottleneck" (the single number the data is squeezed into) wasn't the problem. The problem was how much "thinking space" the machine had before it squeezed the data. When they gave the simpler models more hidden layers (more thinking space), they became almost as good as the Deep AE. It wasn't the fact that they had to squeeze nine things into one; it was that the simpler models were too "narrow" to understand the data before squeezing it.

The "Why": What Drives the Score?

The most exciting part is that the machine figured out the rules of running on its own, without being told what to look for.

Using a special "X-ray" technique (called Explainable AI), the researchers looked inside the winning model to see which variables mattered most. The model consistently identified three drivers of performance:

Running Pace: How fast they are going.
Average Heart Rate: How hard their heart is working.
Aerobic Decoupling: This is a fancy term for "heart rate drift." It measures if your heart rate is getting higher while your speed stays the same (a sign of fatigue or dehydration).

This matched what human physiologists already know: if you are running fast with a steady heart rate and no drift, you are in great shape. If your heart rate is spiking while you slow down, you are tired. The machine discovered this pattern purely from the data.

The Bottom Line

The paper shows that we can use advanced AI to turn a messy pile of wearable data into a single, trustworthy number that coaches can use. The best model (Deep AE) didn't just crunch numbers; it learned the "language" of running physiology.

However, the authors are careful to note:

They tested this on data from one specific platform (Golden Cheeta).
They didn't have "ground truth" (like actual race results or lab tests) to prove the score is perfect, but the results matched known science.
The system is currently a "black box" that needs to be interpreted carefully, but they built a framework to make sure the results are stable and not just random noise.

In short: They built a machine that can look at a runner's watch data and say, "Here is your performance score, and here is exactly why," without needing a human to manually calculate it.

Technical Summary: Autoencoder Architectures for Athlete Performance Scoring from Wearable Telemetry

Problem Statement
The proliferation of IoT wearable devices has shifted the primary challenge in sports analytics from data collection to data synthesis. Runners generate high-dimensional, multi-sensor training logs (kinematic, physiological, and environmental) that are difficult for coaches to interpret manually. Existing approaches often rely on hand-crafted summaries with fixed weights, which fail to capture nonlinear interactions between workload, fatigue, and cardiovascular response (e.g., the simultaneous drop in pace and rise in heart rate during uphill running). While dimensionality reduction techniques like Principal Component Analysis (PCA) exist, they are constrained by linearity and typically optimize only for reconstruction accuracy, neglecting the interpretability and stability of the resulting performance scores. There is a lack of a unified framework that compresses multi-sensor telemetry into a single, interpretable scalar score while preserving nonlinear physiological relationships and ensuring stable feature rankings.

Methodology
The study evaluates five dimensionality reduction architectures on a dataset of 45,836 anonymized running sessions derived from the Golden Cheeta platform. The input consists of nine features: pace, average heart rate, elevation gain, total distance, cadence, aerobic decoupling, age, body weight, and sex.

Model Architectures: Five models were compared, all reducing the 9-dimensional input to a 1-dimensional bottleneck (latent score):
- Simple AE: A single-layer mapping.
- Medium AE: Input → 16 → 1 → 16 → Output.
- Deep AE: Input → 32 → 8 → 1 → 8 → 32 → Output.
- PCA: A linear baseline.
- Variational Autoencoder (VAE): A probabilistic encoder with a 32-dimensional intermediate layer.
- Design Note: Hidden layers were kept wide (16 or 32 units) to ensure sufficient representational capacity before the 1D bottleneck, distinguishing capacity limitations from bottleneck limitations.
Evaluation Framework:
- Composite Selection Criterion: Model quality was assessed using a weighted score ( $Score_{sel}$ ) combining reconstruction accuracy and latent score interpretability: $Score_{sel} = \alpha \cdot (1 - MSE_{norm}) + (1 - \alpha) \cdot Q$ . Here, $\alpha=0.5$ , and $Q$ represents Latent Score Quality.
- Interpretability Metrics: To determine $Q$ , feature importance was measured via four metrics: Spearman rank correlation, Kendall rank correlation, Mutual Information, and Permutation Importance (using a Random Forest). These rankings were aggregated using a modified Borda count.
- Stability: Bootstrap validation (resampling with replacement) was used to calculate the probability of features appearing in the top three and to establish 95% confidence intervals for feature ranks.

Key Results

Model Performance: The Deep Autoencoder achieved the lowest reconstruction error (MSE = 0.001781) and the highest composite score (0.972).
Capacity vs. Bottleneck: Widening the hidden layers of the Medium AE significantly improved its performance, bringing it close to the VAE and PCA. This indicates that the limitation in earlier deterministic variants was hidden layer capacity, not the one-dimensional bottleneck itself.
Feature Stability: Across all models and bootstrap resamples, three features consistently emerged as the dominant drivers of the latent score: running pace, aerobic decoupling, and average heart rate. These features achieved a top-3 inclusion probability of 1.000.
Comparison with Linear Baseline: The Deep AE latent score was strongly inversely correlated with a simple linear baseline (Spearman $\rho = -0.850$ ). The nonlinear model prioritized aerobic decoupling, a factor missed by simple linear combinations of pace and heart rate.
Physiological Validity: The latent scores successfully separated athletes into physiologically distinct groups. High-scoring athletes exhibited faster paces (2.2–2.6 min/km), near-zero aerobic decoupling, and heart rates of 125–134 bpm, while lower-scoring athletes showed slower paces and significant cardiovascular drift.

Key Contributions

Unified Evaluation Framework: The paper introduces a method for unsupervised athlete ranking that simultaneously optimizes for reconstruction accuracy and the interpretability of the derived score.
Composite Selection Criterion: It proposes a selection metric that prevents the exclusive selection of models based on reconstruction error, thereby favoring models that yield more interpretable latent representations.
Consensus Feature Ranking: The study utilizes a modified Borda count to aggregate four distinct Explainable AI (XAI) metrics, providing a robust consensus on feature importance.
Stability Assessment: It employs bootstrap validation to confirm that the identified key physiological drivers are stable across data splits and not artifacts of specific training sets.

Significance and Claims
The paper claims that the proposed framework successfully compresses complex, multi-sensor runner telemetry into a single, interpretable performance indicator without requiring labeled outcomes or expert-defined weights. The results demonstrate that deep autoencoders can recover physiological relationships consistent with established sports science (specifically the roles of pace, heart rate, and aerobic decoupling) purely through unsupervised learning.

The authors emphasize that relying solely on reconstruction error (MSE) is insufficient for physiological modeling, as different architectures can produce similar errors but rank athletes in contradictory orders. By integrating interpretability and stability checks, the framework offers a practical tool for coaches to summarize performance and identify fatigue or overreaching. The study concludes that while the current results are promising, future work requires supervised validation against ground-truth metrics (e.g., VO2max, race results) and cross-platform generalization to other device ecosystems.