Ballistocardiography (BCG) with mattress-embedded sensors is attractive for long-term sleep monitoring because it does not require attaching sensors to the body. However, previous studies varied in hardware, features, and validation protocols, making it unclear which feature groups are most informative for detecting respiratory events. This work stands apart: the nearest neighbors in our archive concern ECG, cardiogram-based sleep, black holes, and neural network methods for lungs, but not a systematic comparison of BCG features on a large clinical cohort. This study fills that gap by comparing all major feature groups for the first time in a unified patient-independent protocol.
Data were collected at the Sleep Disorders Center of the University Hospital Zurich from 155 patients (52 women, 103 men, mean age 49.5 years, median apnea-hypopnea index 14.4 events/hour). A mat of 512 capacitive pressure sensors (16×32, 50 Hz) was placed under the bedsheet, while standard respiratory polygraphy (Philips Respironics Alice 6) was recorded simultaneously. Synchronization between systems was achieved using a pseudo-random sequence fed into the EEG channel of the polygraph; median synchronization accuracy was 21 ms. Six channels were derived from the signal: four zonal aggregates (upper, middle, lower, sum), the average of the five best sensors, and an adaptively preprocessed channel with high-frequency filtering, sensor selection based on respiratory periodicity quality, and polarity alignment. For each window, 191 features were extracted from ten groups: general statistics, interquartile range, zero crossings, area under the curve (AUC), power spectral density (PSD), fast Fourier transform (FFT), energy, entropy, wavelets, and Higuchi fractal dimension. Binary classification (event vs. clean window) was evaluated using leave-one-patient-out (LOPO) cross-validation with Random Forest and Histogram Gradient Boosting.
Random Forest achieved AUC-ROC 0.967 (SD 0.044) and AUC-PR 0.977; Histogram Gradient Boosting achieved AUC-ROC 0.969 (SD 0.052) and AUC-PR 0.979. For 79% of patients, AUC-ROC exceeded 0.95 for Random Forest, and for 83% for Hist-GBT. Feature importance analysis using mean decrease in impurity (MDI) revealed the dominance of frequency-domain features: PSD features (power in the 0.1–0.4 Hz band) account for 30.3% of total importance, and FFT descriptors of spectral shape on the preprocessed channel add another 15.1%. Together, frequency groups contribute 45.4%. Time-domain AUC features (area and curve length) add 21.5%, and general statistics contribute 22.6% (but spread across 78 features, averaging 0.29% per feature). Wavelets, fractal dimension, frame energy, and entropy together contribute only 10.4% across 59 features. The top-3 features are spectral standard deviation, spectral mean, and power in the 0.1–0.4 Hz band on the preprocessed channel. Frequency and AUC features together explain 67% of the information, requiring only 48 features.
This work provides the first quantitative ranking of BCG feature groups for detecting respiratory events in a large cohort with strict patient-independent validation. Unlike neighboring studies that focus on individual modalities (ECG, neural networks for images) or use smaller samples, here all major feature groups are systematically compared on a single hardware platform. The result shows that a compact set of frequency and time-domain features is sufficient for clinically meaningful accuracy, simplifying the development of resource-constrained systems. This is an empirical basis for feature selection, not just a demonstration of high accuracy.
The topic could evolve toward validating the feature ranking in home settings with different mattresses, sleeping positions, and co-sleeping. It is also promising to investigate the influence of event type (obstructive, central, hypopnea) on feature importance, since this study pools all events. Optimization of wavelet representations is possible: other bases may yield different contributions. Additionally, it would be interesting to see whether frequency features remain dominant under lower signal quality, for example with fewer sensors.
The results will influence the development of wearable and embedded sleep monitoring systems, allowing selection of a minimal feature set without loss of accuracy. This is also useful for clinical practice, as it simplifies model interpretation and reduces computational requirements.
Conduct validation on an independent home dataset with diverse sleep conditions. Investigate feature importance for subtypes of respiratory events and compare with alternative feature selection methods such as SHAP.
This work addresses an unresolved problem in sleep medicine: how to achieve reliable detection of respiratory events outside the laboratory without attached sensors. Existing ambulatory systems suffer from variability in conditions and the need for manual tuning. This study shows that even simple spectral features based on BCG can provide high accuracy, bringing the solution of home apnea screening closer. However, the question of generalization to different populations and conditions remains open.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This study systematically compares ten groups of ballistocardiographic signal features for detecting respiratory events in 155 patients with suspected obstructive sleep apnea. A 512-capacitive pressure sensor array was used, synchronized with respiratory polygraphy. Under strict leave-one-patient-out cross-validation, Random Forest and Histogram Gradient Boosting achieved AUC-ROC of 0.967 and 0.969, respectively. Feature importance analysis showed that frequency-domain characteristics dominate, explaining 45.4% of the information, while time-domain AUC features add another 21.5%.
Where its strength is
A large-scale comparison of feature groups on a clinical cohort with patient-independent validation provides an empirically grounded benchmark for feature selection in future BCG-based sleep monitoring systems.
Where it could go next
Worth trying
Why this may matter
Such a systematic comparison of features sets a benchmark for developing non-invasive home sleep monitoring systems, which could expand access to the diagnosis of sleep-disordered breathing.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.52, against a usual 0.64. It looks like few have drilled nearby.
The median synchronization accuracy between the polygraph and the mat was 21 milliseconds, which is less than the duration of one frame at 50 Hz. This accuracy was achieved by transmitting a pseudo-random code through the EEG channel, like Morse code between two independent clocks.