Wearable fitness devices are widely used to monitor physiological signals such as heart rate (BPM) and speed during physical activity. However, these signals often sufer from noise, technical inaccuracies, and context-dependent variability. In this study, we investigate unsupervised anomaly detection methods to identify abnormal segments in real-world data collected from runners using wearable sensors. The dataset includes over 180,000 measurements from 43 running sessions, with speed and BPM values aligned and preprocessed to build a multivariate time series. We compare four approaches representative of diferent anomaly detection paradigms: distance-based (k-Nearest Neighbors), classification-based (One-Class SVM), probabilistic (Kernel Density Estimation), and sequence-based deep learning (TadGAN). Classical methods operate on pointwise values and capture global anomalies with high precision, but they fail to detect contextual or collective anomalies. TadGAN, in contrast, is trained on overlapping sequences and demonstrates the ability to identify local patterns of abnormality across time. Our results highlight the complementarity of these methods and the importance of modeling temporal structure when anomalies are subtle or context-dependent. Although TadGAN fails to capture extreme point anomalies, its performance on sequence-level detection suggests promising directions for future research in health-aware fitness monitoring. All analyses were conducted without labels, under purely unsupervised conditions.
CEUR ceur-ws.org
Wearable fitness technologies have become a cornerstone of modern personal health monitoring, ofering non-invasive and continuous access to physiological and kinematic data during exercise. Among the most frequently collected variables are heart rate (BPM) and locomotor speed, which are commonly used as indicators of training intensity, cardiovascular load, and overall physical condition. These metrics are especially critical in endurance disciplines such as running, where performance, safety, and adaptation must be balanced in real time.
However, despite their utility, such measurements are
inherently subject to various sources of uncertainty [
Detecting these anomalies is of dual importance. On
Attribution 4.0 International (CC BY 4.0). © 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Furthermore, since the data are unlabeled — no ground truth exists for “normal” versus “anomalous” episodes — only unsupervised learning techniques are applicable.
These include density-based, distance-based, probabilistic, classification-driven, and generative models.
From a physiological standpoint, both BPM and speed
are computed indirectly. Heart rate is commonly
estimated through photoplethysmography (PPG) via optical
sensors on the wrist. This technique relies on green LED
light absorption by pulsating blood flow and is known
to be sensitive to placement, motion, and skin
characteristics [
Additionally, anomalies may not only stem from the ity, and applicability to temporal or multivariate data.
devices but also from the athlete. For example, during In a broad review by Nassif et al. [
To explore this problem, we test and compare a range ensemble techniques, and regression models. Often,
modof unsupervised techniques: ern systems leverage hybrid combinations of these
categories to enhance robustness.
• Three classical models on pointwise data: Ruf et al. [
TadGAN, a GAN-based approach designed to re- Probabilistic models attempt to fit the probability disconstruct temporal patterns. tribution of normal data and classify low-probability regions as anomalous. Classic approaches use Mahalanobis
Each model captures a diferent perspective on distance, Gaussian mixture models [
Distance-based methods, such as k-Nearest Neigh- anomalies with robustness to noise. Given the nature of
bors (kNN), Local Outlier Factor (LOF), and related al- our dataset—low-dimensional, sequential, and partially
gorithms, operate by measuring the relative distance or corrupted—we selected kNN, KDE, OC-SVM for
pointdensity deviation of samples with respect to their neigh- wise analysis, and TadGAN for sequential modeling. This
bors. Goldstein et al. [
In temporal anomaly detection, time-series models
become essential. Unlike tabular data, time-series re- Our dataset comprises 43 running sessions, for a total of
tains sequential information, and anomalies may occur 180,876 measurements, split into 137,515 training
samin patterns rather than isolated points. The concept of ples and 43,263 for testing. The data includes heart rate
contextual and collective anomalies, discussed in Chan- (beats per minute, BPM) and speed values, recorded
asyndola et al. [
GAN-based models like TadGAN [
It is important to note that for many applications, sampling (e.g., 99.2% of speed delta_T values are 2 or 3 ground-truth labels are missing, so unsupervised learn- seconds, but some anomalous values exist), a matching ing becomes necessary. Many real-world datasets require step was required. To fuse the datasets, we took the BPM domain expert labeling, which is expensive or infeasible timestamps as reference and aggregated the speed values at scale. Therefore, unsupervised models dominate the over those intervals using a mean operation. This reifeld, often evaluated on synthetic datasets or via indirect sulted in the loss of approximately 21% of BPM rows due proxy metrics. to misalignment or missing speed data at corresponding
Our specific application poses several data quality chal- timestamps. The final aligned dataset consisted of 29,147 lenges. Temporal gaps, disjoint session timestamps, and samples. asynchronous recording of heart rate and speed introduce An exploratory data analysis was conducted on both artifacts that could mislead anomaly detection models. signals. The BPM values appeared within plausible physFor example, some training sessions show BPM values iological ranges (60–200 bpm), whereas speed values while no speed is recorded, indicating diferent sensor exhibited substantial variability. Outliers were initially systems. In other sessions, high BPM spikes do not corre- visualized using boxplots (Figure 4), but not removed spond to any physical acceleration, but rather to recovery to preserve diversity and noise for anomaly detection. states or sensor noise. Speed inconsistencies and their dependence on delta_T
One example is the session on 10/02, where a runner re- were analyzed (Figure 3). sumed activity after a one-month break. While the speed Principal Component Analysis (PCA) was employed to returned to prior levels, BPM values were significantly inspect the structure of the dataset and observe potential elevated. Although physiologically plausible, a model separability among patterns or noise (Figure 5). It also not incorporating temporal and contextual information allowed us to monitor the efects of outlier cleaning, as might flag this as an error. shown in Figure 8.
Thus, several methods were examined prior to imple- Finally, to prepare the dataset for modeling using mentation: - Statistical rules, such as the 3-sigma Gaus- GANs, we applied a sliding window approach with a sian rule. - Exploratory data techniques, such as box window length of 100 and a step size of 5. This generated plots and rank analysis. - Density-based clustering, e.g., overlapping multivariate sequences from the final dataset, DBSCAN. - Ensemble frameworks like PyOD. each containing normalized BPM and speed. These se
signals. The BPM values appeared within plausible physiological ranges (60–200 bpm), whereas speed values exhibited substantial variability. Outliers were initially visualized using boxplots (Figure 4), but not removed to preserve diversity and noise for anomaly detection. Speed inconsistencies and their dependence on delta_T were analyzed (Figure 3).
Principal Component Analysis (PCA) was employed to inspect the structure of the dataset and observe potential separability among patterns or noise (Figure 5). It also allowed us to monitor the efects of outlier cleaning, as shown in Figure 8). # runs # points # inf speed points # inf bpm points # NaN speed points # NaN bpm points
Train
Since no ground-truth labels are available in the dataset,
the evaluation of the anomaly detection methods was
carried out through qualitative visual inspection of the
results. In Table 2, we provide a structured comparison
of anomaly scores and detection outputs from all tested
methods. Each method produces scores on diferent
numerical scales; therefore, a MinMaxScaler was applied to
normalize scores in the [
In the first row of Table 2, we observe the distribution of the normalized anomaly scores for each method, with the red horizontal line indicating the 99.9th percentile.
This threshold was then employed to classify outliers.
The second row shows the test dataset over time, where speed and BPM are plotted with marked anomalies based on the threshold. The third row presents a scatter plot of BPM vs speed, where colors reflect the anomaly score, helping to distinguish point and group anomalies. Finally, the fourth row illustrates one representative test session in time-series format, comparing raw signals and detected anomalies.
K-Nearest Neighbors (KNN) and One-Class SVM (OCSVM) showed good capability to highlight global point anomalies but failed to capture more subtle temporal patterns or contextual outliers. Their scores remained mostly stable across time, and rapid local variations in BPM or speed were not reflected in the outputs. As a Figure 13: Number of outliers according to a tolerance level. result, the methods could only detect extreme values that deviate globally from the norm.
The Kernel Density Estimator (KDE), instead, demon- integration, it missed several subtle dips and transient strated higher sensitivity to group anomalies and moder- events that would qualify as contextual anomalies. ately abnormal sequences. It successfully identified both TadGAN, a generative model trained over temporal global point outliers and sustained deviations in BPM windows, exhibited a distinct behavior. The anomaly or speed. However, due to its lack of temporal context score was highly sensitive to the shape and structure of the sequence, favoring low variation (plateau) segments.
As a result, sudden changes or transitions received lower scores, which contrasts with expectations. This behavior likely stems from the reconstruction-based scoring mechanism and the use of overlapping sliding windows.
Post-processing of overlapping scores into a single perpoint score may have diluted signal peaks. Further investigation is needed to refine this post-processing, possibly by reducing the sliding window step size or aggregating scores using weighted schemes.
In conclusion, point-wise methods (KNN, OCSVM, KDE) are appropriate for detecting global anomalies and extreme values. Temporal methods like TadGAN are more suitable for discovering contextually anomalous patterns, although they require more careful calibration. The integration of hybrid methods or ensembles could ofer a more balanced performance across diferent anomaly types.
rate (BPM) and speed, with the objective of identifying anomalous patterns that could indicate sensor faults, physiological irregularities, or performance deviations. We evaluated four unsupervised methods: KNN, OCSVM, KDE, and TadGAN. The first three operate on individual datapoints, while TadGAN uses sliding temporal windows.
The results showed that point-wise methods excel at detecting isolated outliers but are limited in capturing sequential patterns. In contrast, TadGAN was more sensitive to contextual anomalies and group deviations over time but struggled with extreme values. Its limitations are likely tied to the aggregation of reconstruction errors across windows.
Future work should investigate improvements in GANbased models by tuning window size, overlap, and scoring methods. Moreover, the integration of contextual metadata or ensemble learning approaches may enhance robustness. Overall, temporal models show promising potential for enhancing anomaly detection in physiological monitoring during exercise.
ChatGPT, Grammarly in order to: Grammar and spelling check, Paraphrase and reword. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. KNN