Publications

2025

1. Health

   Wake/Sleep Classification via a Label-Efficient Contrastive Learning Approach

   Safa Boudabous, Juliette Millet, and Emmanuel Bacry

   Sep 2025

   working paper or preprint

   Abs PDF

   Data annotation can be highly labor-intensive for sleep staging, where experts need to analyze encephalograms recorded throughout an entire sleep night. To address this challenge, this paper presents CL-WSC, a label-efficient contrastive learning approach for wake/sleep classification. First, we use a contrastive learning model to learn versatile, time-enriched representations at the timestep level. Next, we train a machine learning model on a small labeled dataset for the Wake/Sleep classification to learn to map the representation space to the output space. We evaluate our approach using data from the Multi-Ethnic Study of Atherosclerosis (MESA) dataset, comprising 1186 patient records. We demonstrate that the learned representations effectively capture key characteristics of the input signals, which are also useful for Sleep/Wake classification without requiring any labels during training. Additionally, we demonstrate that our model achieves performances comparable to those of fully supervised models while using only one-tenth of the labeled data.

2024

1. Health

   Combining Signals for EEG-Free Arousal Detection during Home Sleep Testing: A Retrospective Study

   Safa Boudabous, Juliette Millet, and Emmanuel Bacry

   Diagnostics, Sep 2024

   Abs HTML

   ntroduction: Accurately detecting arousal events during sleep is essential for evaluating sleep quality and diagnosing sleep disorders, such as sleep apnea/hypopnea syndrome. While the American Academy of Sleep Medicine guidelines associate arousal events with electroencephalogram (EEG) signal variations, EEGs are often not recorded during home sleep testing (HST) using wearable devices or smartphone applications. Objectives: The primary objective of this study was to explore the potential of alternatively relying on combinations of easily measurable physiological signals during HST for arousal detection where EEGs are not recorded. Methods: We conducted a data-driven retrospective study following an incremental device-agnostic analysis approach, where we simulated a limited-channel setting using polysomnography data and used deep learning to automate the detection task. During the analysis, we tested multiple signal combinations to evaluate their potential effectiveness. We trained and evaluated the model on the Multi-Ethnic Study of Atherosclerosis dataset. Results: The results demonstrated that combining multiple signals significantly improved performance compared with single-input signal models. Notably, combining thoracic effort, heart rate, and a wake/sleep indicator signal achieved competitive performance compared with the state-of-the-art DeepCAD model using electrocardiogram as input with an average precision of 61.59% and an average recall of 56.46% across the test records. Conclusions: This study demonstrated the potential of combining easy-to-record HST signals to characterize the autonomic markers of arousal better. It provides valuable insights to HST device designers on signals that improve EEG-free arousal detection.

2021

1. Mobility

   Dynamic Graph Convolutional LSTM application for traffic flow estimation from error-prone measurements: results and transferability analysis

   Safa Boudabous, Stephan Clémençon, Houda Labiod, and 1 more author

   In 2021 IEEE 8th International Conference on Data Science and Advanced Analytics (DSAA), Sep 2021

   Abs

   The technological advances in the transportation and automotive industry led to the use of new types of sensing systems more cost-effective and adapted to large-scale dense deployment. Those sensing techniques allow continuously gathering traffic measurements times series in different geospatial locations. The accuracy of the obtained raw measurements is often hindered by different factors related to the sensing environment and the sensing process itself and thus fail to capture the short-term traffic variations crucial for real-time traffic monitoring. In this paper, we propose the DGC-LSTM model for area-wide traffic estimation from error-prone measurements time series. The backbone of the DGC-LSTM model is a graph convolutional Long Short Term Memory model with a dynamic adjacency matrix. The adjacency matrix is learned and optimized during the model training. The adjacency matrix values are estimated from the set of contextual features that impact the dynamicity of the dependencies in both the spatial and temporal dimensions. Experiments on a realistic synthetic labelled Bluetooth counts dataset is used for model evaluation. Lastly, we highlight the importance of transfer learning methods to improve the model applicability by ensuring model adaptation to the new deployment site while avoiding the extensive data-labelling effort.

2. Mobility

   Vehicular traffic analysis based on Bluetooth sensors traces

   Safa Boudabous

   Institut polytechnique de Paris, Sep 2021

   HTML

2020

1. Telecommunication

   Simulation model of bluetooth passive scanning for vehicular traffic monitoring

   Safa Boudabous, Julian Garbiso, Shabbir Ali, and 2 more authors

   In 2020 IEEE 92nd Vehicular Technology Conference (VTC2020-Fall), Sep 2020

   Abs

   We propose a simulation model of Bluetooth (BT) passive scanning for vehicular traffic monitoring. The model is designed to simulate packet detection in the physical and Medium Access Control (MAC) layers for fixed BT sensors. At the physical layer, we consider the radio propagation effects of path loss, small- and large-scale fading. At the MAC layer, we model channel hopping and packet collisions. Associated to a vehicular traffic simulator, our proposal provides a cost-efficient way to generate large-scale BT traffic measurement datasets including both sensor and ground-truth data, avoiding deployment costs. We validate the model by comparing the simulation output to experimental data and we show a simulation use-case of vehicular traffic monitoring in a complex urban context.

2019

1. Mobility

   Traffic analysis based on bluetooth passive scanning

   Safa Boudabous, Julian Garbiso, Bertrand Leroy, and 2 more authors

   In 2019 IEEE 89th Vehicular Technology Conference (VTC2019-Spring), Kuala Lumpur, Malaysia, Sep 2019

   Abs

   During the last decade, Bluetooth has become a widespread feature in the automobile industry, meaning that its signal activity can be correlated with road traffic. Using this technology as a means for assessing traffic is very cost-effective and has a low impact on the infrastructure. Nevertheless, unlike other techniques, it does not provide direct sensing of vehicles only. Statistical analysis methods need to be applied to the collected data. In this paper, we first propose a machine learning method for traffic flow estimation using Bluetooth sensors. We also propose a method for estimating the mean travel speed between two sensors. The performance of the proposed methods is evaluated through eight weeks of experimentation. Finally, we envision a potential solution for building real-time Origin- Destination matrices.