Data-Driven Reduction of External Load Variables in Indoor Team Sports Using Local Positioning System

Area of Science:

• Sports Science
• Biomechanics
• Data Science

Background:

• Local positioning systems (LPSs) generate extensive external load data in team sports.
• Variable redundancy in LPS data complicates performance analysis and load management.
• Dimensionality reduction is needed to identify key external load metrics.

Purpose of the Study:

• To reduce the dimensionality of external load variables from LPS data.
• To identify data-driven external load profiles using PCA and clustering.
• To assess the effectiveness of PCA in preserving clustering structure.

Main Methods:

• Principal Component Analysis (PCA) for dimensionality reduction.
• K-means clustering applied to full and PCA-reduced datasets.
• Silhouette analysis and elbow method for optimal cluster determination.
• Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) for solution agreement.

Main Results:

• The first two principal components captured 53.7% of the variance, representing high-intensity and neuromuscular load.
• Clustering consistently identified three distinct profiles in both full and PCA-reduced spaces.
• PCA-based clustering showed improved separation (silhouette = 0.362) compared to the full space (silhouette = 0.319).
• High agreement (ARI = 0.981, NMI = 0.971) between methods confirmed structure preservation.

Conclusions:

• PCA effectively reduces complex external load variables into interpretable dimensions.
• PCA combined with clustering offers a framework for summarizing LPS data and identifying load profiles.
• Identified profiles require prospective validation for practical training-load management applications.