Use of Wearable Devices to Detect Sedentary Behaviour in an Office Environment

Sedentary Behaviour (SB) is a significant public health problem, which is associated with chronic illness, poorer quality of life, and increased healthcare costs. Wearable devices like Fitbit provide the possibility of 24/7 objective activity monitoring and could be useful in the identification of sedentary behaviour in the general population. This study, however, analyzed a publicly available Kaggle Fitbit dataset, and it focused on classifying the activity levels and identifying office-based sedentary patterns using Machine Learning (ML) and Deep Learning (DL) approaches.

Five supervised ML models, such as Random Forest, Support Vector Machine, XGBoost, K-Nearest Neighbour, and Logistic Regression, were compared with two DL models, such as a feedforward neural network and Long Short-Term Memory (LSTM). Features obtained from daily activity, heart rate, and sleep were used as inputs, whereas the SHAP values were used to explain model predictions. Performance was measured in terms of accuracy, precision, recall, F1 score, and AUC. Stratified train-test splits and cross-validation ensured robustness.

The XGBoost achieved the best classification performance (accuracy 0.95, macro-F1 0.84), followed by the Random Forest. Logistic Regression (Accuracy 0.94, macro-F1 0.79). The feedforward neural network achieved moderate results (accuracy 0.93; macro-F1 0.70), and the LSTM underperformed (accuracy 0.62; macro-F1 0.47), which is typical of the difficulty in using sequence models on small, imbalanced data.

A Streamlit application was created to enable users to upload Fitbit CSV files, perform automated preprocessing, model selection, and predictions with confidence scores. It also produces daily summaries of sedentary instead of active time and offers SHAP-based explanations.

Keywords: Sedentary behaviour detection, wearable devices, physical activity monitoring, machine learning, deep learning, health informatics, human activity recognition.

1. INTRODUCTION

Sedentary behaviour (SB) is any waking activity that involves sitting, reclining, or lying that involves little energy expenditure (≤1.5 METs) [1], [26]. It is becoming an increasingly recognised serious cause of concern in public health, with prolonged SB associated with cardiovascular disease, obesity, musculoskeletal disorders, type 2 diabetes, some forms of cancer, and premature mortality. [2]-[5], [23]-[25], [27].

Modern office environments exacerbate this problem, as workers spend most of their working life in a seated position, and are contributing significantly to daily sedentary time [5], [7], [8], [62].

Early detection and regular monitoring of SB is the key to designing interventions and encouraging active lifestyles. Traditional methods, such as self-reports and observational studies, tend to have problems with recall bias, limited resolution, and scalability. In contrast, wearable devices such as Fitbit allow us to have continuous, objective, and fine-grained measures of activity, heart rate, and sleep in free-living conditions [6], [8], [9], [10], [31], [32].

Open datasets such as the Fitbit collection on Kaggle used for this paper enable the development and validation of ML and DL models for SB detection [9], [11]-[14], [67]. Prior research shows that multimodal features—combining activity, physiological, and sleep data—improve classification accuracy compared with step counts alone [11], [15], [16].

This paper introduces a comprehensive SB detection framework using the Kaggle Fitbit dataset. Five ML models are systematically compared with a TensorFlow-based DL model [9], [10], [17], [18], [70]. SHAP (Shapley Additive Explanation) values are employed to explain ML predictions, while the DL model captures more complex patterns. Performance is assessed with multiple evaluation metrics, and the models are deployed in a Streamlit application for real-time, user-friendly analysis of Fitbit CSV uploads.

The remainder of this paper is organised as follows: Section 2 reviews the literature related to the areas of the use of wearable devices and DL in Activity Recognition in the context of SB; Section 3 outlines the data used and the preprocessing; Section 4 details the proposed methodology; Section 5 presents the results obtained; Section 6 describes the application developed; and Section 7 concludes with recommendations for future research.

2. Literature Review

• • Sedentary Behaviour: Definition, Risks, and Measurement Challenges

Sedentary Behaviour (SB) generally refers to waking activities that involve sitting, reclining, or lying down, with energy expenditure no greater than 1.5 METs [1], [26]. Spending long periods inactive has been linked to a range of health issues, such as heart disease, obesity, type 2 diabetes, musculoskeletal disorders, [2]-[5],[23]-[25], [27], [28], and a heightened risk of early death. In today’s office settings, extended sitting is widespread and plays a significant role in overall daily inactivity.

Conventional approaches, such as self-report surveys, suffer from recall bias and are subjective, whereas direct observation is time-consuming and impractical for use at scale [7], [8], [29], [30]. Objective device-based techniques, including accelerometers and inclinometers, make ongoing and dependable monitoring of posture and movement possible. However, these tools, which are often research-grade, are often costly and require specialised knowledge to process data. By comparison, consumer wearables such as Fitbit aim to provide a more accessible solution, one that is affordable, scalable, and reasonably accurate for monitoring in wide populations [31], [32].

• Wearable Devices and Fitbit in Sedentary Behaviour Detection

Devices such as Fitbit constantly collect all kinds of data, [31]-[33], [67]-[69], such as steps, heart rate, activity levels, and sleep patterns. Several studies have confirmed the accuracy of Fitbit in step detection [13], [14], [31], [33], and activity classification showing similar performance as research-grade accelerometers, hence confirming its suitability for large-scale research applications [8], [61], [62].

Boudet et al. [7] emphasised the importance of integrating various behavioural measures to obtain a more comprehensive picture of SB in workplace health studies. Similarly, Bongers et al. [8] determined that Fitbit was a reasonable and practical tool [12], [15], [67] for measuring SB over prolonged periods in the workplace. More recently, in their study, Wang et al. [9] also showed that step count data from Fitbit can be used effectively in Machine Learning models for classifying SB, and the results were similar and comparable with high-end accelerometers.

• Machine Learning Approaches to SB Classification

A variety of Machine Learning algorithms such as Random Forest, Support Vector Machines (SVMs), XGBoost, K-Nearest Neighbour (KNN), and Logistic Regression have been strongly employed in the analysis of data from wearable devices for the detection of sedentary behaviour (SB) [9]-[11], [40]-[42], [47]. These models do tend to work well when applied to structured data such as daily step counts, minutes of activity, heart rate readings, and sleep statistics.

Kantoch [10], who also reported on high accuracy in recognising SB using wearable sensors during Activities of Daily Living (ADLs), whose performance could further be improved by using Feature Engineering. Similarly, Papathomas et al. [11] have used gradient boosting on daily step count data, revealing the efficiency of lightweight models for the prediction of SB. In addition to accuracy in predictions, the interpretability of ML outputs is critical to their adoption in healthcare settings. Tools like SHAP facilitates to understand how individual features impact model predictions, which improves trust [50] and clinical applicability [65].

• Deep Learning in Activity Recognition and SB Detection

Deep Learning techniques can capture some meaningful patterns directly from raw or weakly processed wearable data, so they can capture complex relationships that might not be detected by traditional ML models [17], [18], [36]- [39], [46], [70], [72]. Models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTMs) networks have given a powerful performance in recognising human activity, including the recognition of Sedentary Behaviour.

Plotz et al. [17], [36], [37] carried out an extensive comparison of deep, convolutional, and recurrent models using wearable sensor datasets, coupled with their superior performance with respect to conventional ML approaches. Similarly, Zhao et al. applied a deep residual bidirectional LSTM for wearable sensor data [18], [70] which exhibited the enhanced ability in capturing temporal patterns. In particular, these results highlight the usefulness of DL models for Fitbit data analysis in SB detection [37], [72], particularly for time-series and high-dimensional inputs.

• Research Gaps

Despite progress in detecting Sedentary Behaviour, several key challenges persist:

(i) Multimodal integration: Most of the research has been conducted by focusing on step count data, and many other useful data, such as heart rate and sleep pattern [9], [11], [15], [16], are ignored.

(ii) Direct ML–DL comparisons: There is a lack of direct comparisons between traditional ML models and DL techniques on the same Fitbit dataset and its advantages [9], [11], [17].

(iii) Practical deployment: Most existing research is still contained within controlled environments, with very few attempts at creating practical tools for everyday use. To overcome this gap, this research presents a Streamlit-based framework that is designed to be deployed in the real world [24], [66], [73].

(iv) Office-centric applications: Very few studies have focused specifically on sedentary behaviours in the office environment, even though this is of great relevance to workplace health [5], [7], [62].

3. Methodology

• • Dataset Description

The study was based on a publicly available dataset of Fitbit data from Kaggle [9], [12], [13], which includes activity tracker data of 30 participants over two months. The dataset is made up of the following components:

(i) dailyActivity_merged.csv – contains daily summaries, including steps taken, distances travelled, active minutes, sedentary minutes, and calories burned.

(ii) sleepDay_merged.csv – contains sleep-related metrics, such as the number of sleep records, total minutes asleep, and total time spent in bed.

(iii) heartrate_seconds_merged.csv – includes second-by-second heart rate measurements, and this will allow for detailed analysis of the cardiovascular activity.

(iv) minuteStepsNarrow_merged.csv – this logs step counts every minute, allowing the average steps per minute to be calculated.

• Data Preprocessing and Integration

The Fitbit data set was cleaned and integrated in a step-by-step manner. Date formats were standardised, and files were merged using the date field ActivityDate to provide consistency between the activity, heart rate, sleep, and step records.

Feature Engineering was then used to help with a more meaningful analysis. For heart rate data, the following statistical measures were computed as a measure of general cardiovascular activity: daily average (AvgHeartRate). Step data were aggregated to obtain daily average steps per minute (AvgStepsPerMinute), which is thought to be an intensity measure of activity. Sleep-related aspects were also extracted, such as the total number of sleep data (TotalSleepRecords), total minutes asleep (TotalMinutesAsleep), and total time in bed (TotalTimeInBed), to offer information on rest and recovery.

Missing activity entries were replaced with zeros, in line with common practices in wearable data analysis. The final set of features included: TotalSteps, TotalDistance, TrackerDistance, LoggedActivitiesDistance, VeryActiveDistance, ModeratelyActiveDistance, LightActiveDistance, SedentaryActiveDistance, VeryActiveMinutes, FairlyActiveMinutes, LightlyActiveMinutes, Calories, TotalSleepRecords, TotalMinutesAsleep, TotalTimeInBed, AvgHeartRate, and AvgStepsPerMinute.

Although the original dataset contained the SedentaryMinutes column, it was used solely to create class labels for sedentary and active days (see Section 3.3). To avoid label leakage, it was excluded from the predictor set during model training.

This research involved an iterative, Agile-inspired development lifecycle [73]. Each cycle included a literature study, data cleaning, model training, evaluation and reflection. A Kanban board was used to plan and track progress, increasing flexibility and accountability throughout the project.

• Sedentary Behaviour Labelling

Labels for training the models were derived from the SedentaryMinutes field in the Fitbit daily activity file. Based on guidance from occupational health and epidemiology, [19]-[23], [24], [60] a threshold of 600 minutes (10 hours) of daily sedentary time was used. Days with SedentaryMinutes ≥ 600 were labelled as sedentary (class = 1), whereas days with values below this threshold were labelled as active (class = 0).

This cut-off is in line with evidence from other studies. Koster et al. [19] reported that estimates of sedentary time using activPAL and ActiGraph devices usually converged at about 8 to 10 hours per day. Similarly, Rowlands et al. [20] found good agreement with activPAL reference standards using similar thresholds in free-living conditions. Kerr et al. [21] also showed that sedentary postures classified using hip-worn accelerometers were reliably identified at a rate of around 10 hours per day. In addition, public health guidance, including that of the UK Chief Medical Officers [60], identifies prolonged sitting (beyond 8 to 10 hours per day) as a critical risk factor for adverse health outcomes.

Using the 600-minute threshold helped align the study’s labelling approach with both research findings and public health advice. To avoid label leakage, the SedentaryMinutes column was not included in the feature set during training, helping to maintain the integrity of both ML and DL models.

• Machine Learning Models

This study implemented 5 supervised Machine Learning algorithms with a selection of algorithms based on their success in prior research of Human Activity Recognition (HAR) and Sedentary Behaviour detection. [40]- [42], [47] The Models included Random Forest (RF), Support Vector Machine (SVM), XGBoost, K-Nearest Neighbour (KNN), and Logistic Regression (LR). Collectively, these models represent several classes of methods such as ensemble learning, kernel methods, instance-based methods, and linear classifiers.

The hyperparameters of each model were optimised with grid search and cross-validation to ensure that you can select settings based on real performance instead of guesswork. This systematic process of tuning helped in mitigating overfitting and improving generalisability to unseen data. In order to improve the model’s interpretability, SHAP (Shapley Additive Explanations) was used to obtain feature-level insight into model predictions [50]-[53], [65]. Combining predictive modelling with explainability. The goal of the study was not only to evaluate how well the model was performing, but to build transparency and trust, which are two critical factors when applying ML in healthcare.

• Deep Learning Model

In addition to the machine learning models, two Deep Learning (DL) architectures were implemented using the TensorFlow/Keras framework. The first was a fully connected feed-forward neural network, [43]-[46], [48], [49], to classify sedentary and active days according to daily Fitbit features. The network had an input layer with 18 features, two hidden layers using ReLU activation functions to capture non-linear relationships. To reduce overfitting, dropout regularisation was used, while early stopping was used to stop training when validation performance started to plateau. The output layer made use of a sigmoid activation function for a binary classification, and a model optimiser (Adam optimiser) and binary cross-entropy loss were used to optimise the model.

To capture temporal patterns, the Long Short-Term Memory (LSTM) network is used to model the sequential nature of the Fitbit data at the minute level. The architecture consisted of an input layer using masking (to take care of padded time steps), a 64-unit LSTM layer using dropout and recurrent dropout. A dense layer using ReLU activation was used to feed the final output sigmoid layer. Training included class weighting to cope with the class imbalance, [57]-[59], as well as early stopping and learning rate reduction callbacks for improved convergence stability.

Evaluation of the LSTM showed that there was weaker performance compared to both the model using ML and the neural network with a dense layer. Although it obtained a relatively strong precision for active days (0.93), the accuracy for sedentary days was low (0.50), and the overall results were modest (Accuracy 62.2%, Macro F1 0.47, ROC-AUC 0.35). These results are a reflection of the challenges of working with temporal DL models using small, imbalanced Fitbit datasets, pointing out the need for larger datasets or more sophisticated designs, such as transfer learning or concatenated CNN-LSTM designs, to better capture patterns in real-world activities [37], [39], [70], [72].

• Tools and Environments

All experiments were implemented in Python 3.10 [47]-[49], using the following libraries: pandas, NumPy, scikit-learn, XGBoost, TensorFlow, Keras, and SHAP. The Matplotlib and seaborn libraries were used for visualisation. GitHub was employed for version control, and Streamlit served as the deployment framework for the final application.

• Evaluation Metrics

Model performance was evaluated using a variety of metrics that indicate both overall accuracy and the accuracy for each class. Accuracy was a general indicator of predictive performance, and the Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) was used to provide a threshold-independent measure of class discrimination between sedentary and active days [54], [55], [56]. To account for class imbalance, precision, recall, and F1-score were also reported. Precision measured the number of correct instances of sedentary or active days out of all instances predicted, while recall measured the number of correct instances out of all instances making up the actual class. The F1-score, which is the harmonic mean of the precision and recall, was especially useful to characterize the tradeoff between the two scores.

To promote robustness, hyperparameter tuning for the ML models used k-fold cross-validation [54], [56], and minimised the risk of overfitting to any individual partition of the data. For the last evaluation, the data set was split into training (80%) and testing (20%) data sets stratified by class to maintain the same distribution of sedentary and active days in the original data. This two-step approach, which used cross-validation for tuning and stratified hold-out testing for final evaluation, helped to produce results that were both dependable and applicable beyond the training data.

• Application Deployment

To demonstrate the practical use of the framework, a web-based application was created with Streamlit, which offers a simple interface for non-technical users. Users can upload raw Fitbit CSV files, which are automatically cleaned and formatted by the system to fit the workflow for the prediction. This makes the process simple for the user, as no technical steps are needed for simplistic preprocessing. The greatest advantage is that it eliminates common formatting errors, and the data is uniform before entering it into the model. However, if someone submits highly customised or incomplete Fitbit exports, they may require some lesser adjustments to be done by hand, which is a practical limitation of using user-provided files. They can then select between Machine Learning and Deep Learning models for generating the predictions, with each prediction accompanied by a confidence score that indicates the certainty of the model.

For more transparency, the application includes SHAP-based explanations, which can reveal how individual features impact each prediction. In addition, the system produces visual charts to display daily activity patterns as well as classify behaviour into sedentary or active categories. A very unique property of the tool is that it gives per day summaries, reporting the percentage of sedentary versus active behaviour and giving a good interpretation, such as “70% sedentary, 30% active — overall sedentary day” or “80% active, 20% sedentary — overall active day.” These outputs both support personal health tracking and workplace wellness efforts with insights that are easy to interpret and apply.

Crucially, this application helps cover a major gap found in the literature and, in particular, the lack of practical tools to monitor Sedentary Behaviour in real-world settings. By creating a bridge between academic knowledge and practical tools, the system has demonstrated the possibility of using the data received from wearable technology at the consumer level for large-scale health monitoring. The combination of accurate classification, interpretability, and user-centred design is representative of the possibilities associated with the integration of ML techniques and wearable technology in the Sedentary Behaviour management in office-based environments in a manner that is evidence-based.

• Ethical and Professional Considerations

The dataset used for this research from Fitbit was fully anonymised and publicly accessible in accordance with the GDPR and guidelines and ethics for research. No personal information was collected or shared. To prevent the leaking of the labels, the SedentaryMinutes feature was excluded from the predictor variables. From the social perspective, the work encourages awareness of well-being without being intrusive, and the use of consumer-grade devices guarantees a sustainable and low-cost solution.

3. Results and Discussion
   • Model Performance

Table 1 shows the comparative performance of all the models implemented. We have five Machine Learning models, a fully connected Deep Neural Network (DNN), and a Long Short-Term Memory (LSTM) model. The evaluation metrics show trade-offs between overall accuracy, macro-averaged F1, weighted F1, and class-specific scores.

Table 1. Model comparison highlights.

• Visual Analysis

Fig. (2) sorts the overall accuracy, macro-F1, and weighted-F1 score for each machine learning and deep learning model. XGBoost, Random Forest, and KNN (scaled) seem fairly consistently good and balanced in all three of these metrics, indicating good and balanced classification. By contrast, Logistic Regression and LSTM have very low values for the macro-f1, which supposedly suggests an issue with class imbalance. The strong correlation between the accuracy and the weighted-F1 values indicates that the high performance is largely compensated for by the performance of the majority class.

Fig. (3) focuses on the Class-0 (sedentary=minority) metrics; we display the precision, recall, and F1 for each model. It emphasises the challenge models have to identify the minority class. SVM (scaled) has a perfect recall (1.00) for Class-0, but has a very low precision, meaning that a lot of false positives appear. XGBoost and Random Forest are better overall balance with increased F1 for Class-0. The deep learning models (Dense NN and LSTM) have a much weaker performance for the minority class, which proves they are more sensitive to the class imbalance.

(Fig. 4). Weighted-F1 scores provide a more realistic view of the performance of each model for each class, considering the distribution of classes. XGBoost, Random Forest, and KNN (scaled) come on top once again with values greater than 0.95. These results have proven their strength and the ability to preserve the performance even with the uneven class frequency. Logistic Regression and LSTM are further down the list, with the LSTM model being the lowest performing out of all the algorithms reviewed in this metric.

(Fig. 5) compares ROC-AUC scores for all models. The Random Forest and KNN models achieve the highest ROC-AUC (1.00), indicating near-perfect separation between the two classes. SVM also performs strongly, followed by XGBoost and Logistic Regression. Deep learning models show comparatively weaker performance, with Dense NN achieving a moderate ROC-AUC (0.65) and LSTM the lowest (0.35). These results confirm that tree-based models remain the most effective for this dataset and task.

(Fig. 6) Cross-validation results showing the mean macro F1-score (± standard deviation) across machine learning models. XGBoost achieved the highest mean performance (0.828) [40], [41], [9], [11], [16], followed by Random Forest (0.772) and SVM (0.775). KNN performed moderately (0.703), while Logistic Regression showed the lowest macro F1 (0.616). Error bars indicate variability across folds.