Autoencoders Score Athlete Performance from Wearable Data

New research explores the use of autoencoder architectures and other dimensionality reduction techniques to interpret large, high-dimensional training data from wearable devices worn by everyday runners. The study evaluated five models—three autoencoder variants, PCA, and a Variational Autoencoder—on their ability to condense nine sensor-derived runner profiles into a single scalar performance indicator, referred to as the latent score. Since the setting is entirely unsupervised, model quality was assessed based on both reconstruction error and the interpretability of the latent score, using metrics like Spearman correlation and Mutual Information. The Deep Autoencoder emerged as the top performer, achieving the lowest reconstruction error and the highest composite score. Key physiological metrics such as running pace, aerobic decoupling, and average heart rate consistently appeared as the most influential drivers of the latent performance score across all models.