The widespread adoption of wearable devices has significantly transformed health monitoring and sports analytics by enabling continuous and non-invasive collection of physiological data. Sensors embedded in smartwatches, fitness trackers, and medical wearables can record Heart Rate (HR), movement patterns, blood oxygen saturation (SpO₂), and other vital signals in real time. These data streams provide valuable insights into an individual’s physical condition, training intensity, and overall health status [1], [2].

In the context of sports and exercise science, understanding exercise intensity and duration is essential for optimizing performance, preventing injury, and designing personalized training programs. Traditional approaches often rely on predefined activity labels or subjective measures such as self-reported exertion levels. However, these methods suffer from several limitations, including labeling inaccuracies, inter-individual variability, and the inability to capture physiological responses that differ across athletes performing the same activity [3].

Recent advances in machine learning have enabled data-driven analysis of wearable sensor data, particularly through supervised classification techniques that attempt to map physiological signals to predefined activity types. While effective in controlled settings, supervised methods require large, accurately labeled datasets and tend to focus on recognizing activity names (e.g., walking or running) rather than underlying physiological workload. In real-world scenarios, the same activity can induce vastly different physiological responses depending on intensity, duration, fitness level, and environmental conditions [4].

Unsupervised learning provides a promising alternative by discovering latent patterns directly from physiological measurements without relying on activity labels. Clustering techniques, in particular, allow the identification of natural groupings in data that may correspond to meaningful exercise patterns such as low-intensity endurance activity or short-duration high-intensity exertion. Such approaches are especially valuable in wearable-based health monitoring, where labeled data are scarce or unreliable [5], [6].

Another challenge in wearable data analysis is model interpretability. Many studies employ high-dimensional feature sets or dimensionality reduction techniques such as Principal Component Analysis (PCA) to improve performance. While effective in certain contexts, these approaches may reduce physiological interpretability and complicate clinical or sports-related insights. Selecting a compact set of meaningful physiological features can reduce model complexity while preserving explanatory power [7].

Motivated by these challenges, this study proposes an interpretable unsupervised framework for categorizing exercise patterns using wearable sensor data. Athletes are grouped into three exercise categories (short-duration high-intensity activity, long-duration low-intensity activity, and moderate activity) based solely on physiological responses. A limited set of key physiological features is selected to maintain interpretability and reduce complexity. Clustering results are validated using activity status information that is not included during model training. Additionally, PCA is applied only for comparative visualization to assess whether dimensionality reduction enhances cluster separation.

The remainder of this paper is organized as follows. Section 2 reviews related work on wearable-based exercise analysis and unsupervised learning approaches. Section 3 describes the dataset, feature selection process, and clustering methodology. Section 4 presents experimental results and discussions. Section 5 outlines limitations and future research directions, and Section 6 concludes the paper.

The increasing availability of wearable sensor data has motivated extensive research in the fields of activity recognition, exercise intensity estimation, and health monitoring. Early studies primarily focused on supervised activity recognition, where physiological and motion-based features were used to classify predefined activity types such as walking, running, or cycling. These approaches demonstrated high accuracy in controlled environments but relied heavily on large, well-labeled datasets and often failed to generalize to real-world conditions [8], [9].

To address labeling challenges, several researchers explored machine learning techniques for estimating exercise intensity rather than activity type. HR–based models and Metabolic Equivalents (METs) have been widely used to quantify physical workload during exercise. While these methods provide useful approximations, they often assume population-level relationships and may not accurately reflect individual physiological responses, especially across different fitness levels [10], [11].

Unsupervised learning approaches have gained increasing attention as a means of discovering latent exercise patterns without relying on explicit labels. Clustering algorithms such as k-means, hierarchical clustering, and density-based methods have been applied to wearable sensor data to identify groups of similar physiological responses. These studies demonstrated that unsupervised methods could reveal meaningful exercise states and transitions that are not captured by traditional activity labels [12], [13] .

Recent advances in unsupervised representation learning for time-series sensor data have increasingly adopted contrastive learning frameworks, which leverage carefully designed temporal augmentations and instance discrimination objectives to learn robust and invariant latent representations without requiring labeled supervision. However, these methods primarily optimize for representation quality and downstream predictive performance, while offering limited insight into interpretable and semantically meaningful activity pattern discovery [14].

Dimensionality reduction techniques, particularly PCA, are commonly employed in wearable data analysis to address high dimensionality and redundancy among features. Several studies have reported improved classification or clustering performance after applying PCA or similar transformations [15], [16]. However, the resulting components are often difficult to interpret physiologically, limiting their usefulness in clinical and sports science applications where explainability is essential.

Recent studies have further investigated weakly and self-supervised learning strategies for wearable sensor-based human activity recognition, focusing on reducing annotation requirements while maintaining competitive performance across standard recognition benchmarks. However, these approaches remain primarily centered on improving recognition accuracy under limited supervision, without explicitly addressing fully unsupervised and interpretable discovery of latent activity patterns [17].

More recent research has emphasized the importance of model interpretability and feature selection in health-related machine learning. Studies have shown that using a compact set of physiologically meaningful features can achieve comparable performance to high-dimensional models while improving transparency and trustworthiness [18], [19]. This shift aligns with the growing demand for Explainable Artificial Intelligence (XAI) in medical and wearable-based analytics.

Recent works have further advanced unsupervised representation learning for time-series sensor data through contrastive learning frameworks, which exploit temporal augmentations and instance-level discrimination to learn robust and invariant feature representations without labeled supervision. However, these approaches are primarily designed to enhance representation quality for downstream tasks such as classification or clustering, rather than enabling interpretable and semantically grounded pattern discovery [20].

Recent advances in wearable-based human activity recognition have increasingly leveraged large-scale self-supervised learning to overcome the limitations imposed by scarce labeled data. Notably, recent work has demonstrated that training on massive unlabeled datasets, spanning hundreds of thousands of person-days, can significantly enhance the generalizability and robustness of learned representations across diverse populations, devices, and activity domains. Despite these promising developments, such approaches primarily focus on improving downstream recognition performance, with limited attention to interpretable and fully unsupervised pattern discovery [21].

Recent advances in wearable sensor-based human activity recognition have increasingly shifted toward self-supervised and unsupervised representation learning to mitigate the reliance on large-scale labeled datasets. In particular, learning meaningful latent structures from raw time-series data has emerged as a promising direction for capturing intrinsic activity patterns. Within this line of research, recent studies have explored the use of self-supervised frameworks to learn compact and discrete latent representations, enabling more structured and semantically meaningful modeling of human activities without manual annotations. Despite these advances, existing studies often focus on either activity recognition or intensity estimation in isolation, and few works explicitly validate unsupervised exercise clusters using independent activity information. Moreover, the necessity of dimensionality reduction in low-dimensional, interpretable physiological feature spaces remains insufficiently explored [22].

In parallel, recent studies have leveraged hybrid deep architectures combining convolutional and transformer-based models within self-supervised frameworks to learn rich temporal representations from wearable sensor data, achieving strong recognition performance with limited labeled data. However, these approaches predominantly focus on optimizing predictive accuracy through complex model architectures, with limited attention to interpretable and structurally meaningful pattern discovery [23].

In addition, recent studies have explored semi-supervised activity detection frameworks that operate directly on wearable devices, emphasizing personalized modeling and privacy preservation while leveraging sparsely labeled data. However, such approaches typically remain focused on classification-oriented objectives and require partial supervision, limiting their applicability to fully unsupervised and interpretable pattern discovery settings [24].

More closely related to unsupervised pattern discovery, recent studies have investigated hybrid deep clustering frameworks that combine representation learning with clustering mechanisms to uncover latent activity structures from wearable sensor data, demonstrating improved capability in identifying complex and transitional activity patterns without labeled supervision. Nevertheless, these approaches often prioritize clustering performance and latent separability, with limited emphasis on interpretability and validation using independent physiological or activity-based evidence [25].

Recent research has systematically explored the trade-offs across supervised, unsupervised, and weakly/self-supervised learning paradigms in wearable-based human activity recognition, emphasizing that reducing reliance on labeled data remains a fundamental challenge, as unsupervised approaches often lack discriminative power while supervised methods demand costly annotations. This tension has motivated the development of hybrid and weakly supervised frameworks; however, the problem of fully unsupervised and interpretable pattern discovery remains largely unresolved [26].

In contrast to prior work, the present study adopts an unsupervised clustering framework based on a small set of key physiological features to categorize exercise patterns by intensity and duration. The proposed approach emphasizes interpretability, avoids reliance on activity labels during training, and includes an explicit validation step using withheld activity status information. Additionally, PCA is evaluated only as a comparative tool, providing insight into its practical necessity in this context.

Across all analyses, the proposed unsupervised framework successfully identifies meaningful exercise patterns that align with established physiological principles. The results demonstrate that clustering based on a small set of interpretable features can effectively differentiate exercise intensity and duration without relying on activity labels. Validation using independent activity status information and comparative PCA analysis further confirms the robustness and practicality of the proposed approach.

This study presented an interpretable unsupervised learning framework for analyzing exercise patterns using wearable sensor data. By focusing on a compact set of physiologically meaningful features, the proposed approach successfully categorized athletes into three exercise groups based on intensity and duration without relying on predefined activity labels.

The clustering results demonstrated strong alignment with established physiological principles and were further validated using independent activity status information. Comparative analysis using PCA confirmed that dimensionality reduction did not enhance cluster separability, supporting the decision to preserve the original feature space for improved interpretability.

Overall, the findings highlight the effectiveness of unsupervised clustering in capturing latent exercise behaviors from wearable data while maintaining transparency and practical relevance. The proposed framework offers a scalable and data-driven foundation for personalized exercise analysis and wearable-based health monitoring applications.

Finally, integrating the clustering results into real-time wearable systems could enable personalized feedback and adaptive training recommendations, highlighting the practical potential of unsupervised learning in sports and health monitoring.

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

All data are included in the text.

The authors, being non-native English speakers, used AI tools to refine the English language of this manuscript.