PaperlayerView exact revision on arXiv

Source-linked AI summary

Spectral Features Dominate BCG Respiratory-Event Detection: A Large-Scale Patient-Independent Comparison of Feature Groups in Sleep Apnea Patients

Israel Campero Jurado, Zoe Bousraou, Lara Benning, Sara Padilla Neira, Alexander Breuss, Robert Riener, Esther Irene Schwarz, Elisabeth Wilhelm

arXiv:2608.28242v1cs.LG

TL;DR

The paper addresses uncertainty about which BCG feature groups best discriminate respiratory events under patient-independent validation. It compares feature groups using strict leave-one-patient-out evaluation and finds that frequency-domain features dominate, while a compact frequency- and time-domain subset retains near-maximal performance.

  • Problem

    Which BCG signal features are most discriminative for respiratory-event detection, given substantial differences across prior studies in hardware, feature definitions, prediction targets, and validation?

  • Method

    The study systematically compares ten BCG feature groups under strict leave-one-patient-out cross-validation in a cohort of 155 patients.

  • Results

    Frequency-domain features account for 45.4% of Random Forest discriminative information, including 30.3% from breathing-band power and 15.1% from FFT spectral-shape descriptors.

  • Takeaways & Limitations

    A detector can achieve near-maximal discriminative performance using frequency-domain features and time-domain AUC groups comprising 48 features and 67% of MDI.

  • Takeaways & Limitations

    Feature-importance rankings may overestimate high-cardinality features, and pooling respiratory event types or changing the prediction target may yield different rankings.

Abstract

from arXiv · show

Unobtrusive ballistocardiographic (BCG) sensing is a promising modality for long-term sleep-apnea monitoring, yet it remains unclear which signal features are most discriminative for respiratory-event detection. We present a literature-guided, patient-independent comparison of ten BCG feature groups using a 512-sensor capacitive pressure mat recorded simultaneously with respiratory polygraphy in 155 patients (52 female, 103 male) undergoing in-hospital evaluation for obstructive sleep apnea. Features were extracted from six spatially distinct signal channels, yielding a 191-dimensional feature vector spanning general statistical, time-domain, frequency-domain, wavelet, frame-energy, and nonlinear complexity descriptors. Under strict leave-one-patient-out cross-validation for binary classification of respiratory-event windows versus event-free reference windows, Random Forest and Histogram Gradient Boosting achieved AUC-ROC of 0.967 and 0.969 and AUC-PR of 0.977 and 0.979, respectively. Feature-importance analysis revealed that frequency-domain features dominate discrimination: breathing-band power in the 0.1-0.4 Hz range accounted for 30.3% of total discriminative information across all spatial channels, and Fast Fourier Transform spectral-shape descriptors of the adaptively preprocessed channel contributed a further 15.1%. AUC and curve-length features provided the main complementary time-domain evidence (21.5%), whereas wavelet-derived and nonlinear features contributed smaller secondary effects (10.4% combined across 59 features). Frequency-domain and time-domain features together accounted for 67% of total discriminative information, demonstrating that a compact, interpretable subset of the full feature library achieves clinically relevant performance under patient-independent validation and providing an empirical basis for feature selection in future BCG systems.

1. Introduction

BCG enables comfortable, repeated overnight monitoring, but prior studies use differing hardware, features, targets, and validation protocols, leaving feature-group contributions unclear. This study compares ten literature-guided feature groups across six channels under patient-independent validation in 155 patients.

  • BCG captures respiratory effort, posture, and gross body movement without body-attached sensors, supporting comfortable repeated overnight monitoring.
  • Prior BCG studies differ substantially in sensing hardware, feature definitions, prediction targets, and validation protocols.
  • The unresolved question is which feature groups contribute most to distinguishing respiratory-event windows from normal-breathing windows.
  • Feature ranking can identify event-indicative BCG properties and guide feature selection for real-time or resource-constrained systems.
  • The study extracts 191 literature-guided features from ten groups across six channels of a 512-sensor capacitive BCG recorded synchronously with respiratory polygraphy in 155 patients.

2. Background and Related Work

BCG combines mechanical signals from cardiovascular activity, respiration, posture, and body movement, making feature engineering important for detecting respiratory events. Prior work uses varied feature categories and sensing configurations, but has not jointly compared their contributions under patient-independent validation.

  • BCG records a superposition of mechanical sources, including cardiovascular activity, respiratory effort, posture, and gross body movement.
  • Respiratory events can alter BCG frequency content and waveform morphology through changed breathing patterns, recovery breaths, and arousal-related heart-rate changes.
  • Prior BCG, bed-pressure, ECG, and acoustic studies informed feature selection because these modalities target the same underlying physiological events.
  • Table 1 distinguishes classifier features from preprocessing or signal-extraction uses and marks unreported categories, while defining PVDF as a piezoelectric polymer material.
  • No single prior study used all feature categories simultaneously or compared their relative discriminative contributions under patient-independent validation.
  • Earlier studies commonly used statistical, time-domain, and frequency-domain features, while wavelet and nonlinear methods were more often applied to signal separation or cardiac-interval analysis.
  • Many systems rely on one sensor or a global aggregate, whereas high-resolution mats can represent spatial differences in thoracoabdominal motion, posture coupling, and breathing patterns.

3. Materials and Methods

The study synchronised polygraphy and BCG recordings, derived spatial signal channels, and extracted a 191-feature representation for patient-independent respiratory-event classification. Adaptive sensor selection and preprocessing were used alongside global and regional aggregates to preserve respiratory structure across the pressure mat.

  • Polygraphy–BCG Synchronisation: EEG sync pulses aligned independently clocked polygraphy and Raspberry Pi BCG recordings before respiratory-event labels were mapped onto the BCG timeline.The recordings came from a clinical setting and used expert-scored polygraphy events.
  • Signal Channels: Six channels represented the 512-sensor BCG through global, upper-, middle-, and lower-body aggregates plus raw and adaptively preprocessed top-sensor averages.The spatial aggregates preserved coarse anatomical information, while top sensors were selected for respiratory periodicity.
  • Adaptive Sensor Selection: The adaptive sensor criterion ranked sensors using respiratory periodicity and spectral concentration, then selected the five highest-scoring sensors for each window.The score combined autocorrelation and spectral-concentration terms after high-pass filtering.
  • Signal Channels: The raw top-sensor channel averaged the selected sensors pointwise without filtering, preserving slow baseline components in the pressure signal.This channel complemented the adaptively preprocessed representation.
  • Feature Representation: The complete per-window representation contained 191 features, comprising 110 features from five aggregate channels and 81 from the preprocessed channel.The feature vector combined the channel-specific implementations of the standardised feature groups.

4. Results

Across patient-independent LOPO evaluation, non-linear classifiers achieved high respiratory-event detection performance, while feature importance was concentrated in frequency-domain and complementary time-domain descriptors. Synchronisation was also precise across the cohort, supporting the event-mapping pipeline.

  • 155 participants comprised the final dataset spanning the full clinical range of obstructive sleep-apnea severity.
  • AUC-ROC was 0.967 for Random Forest and 0.969 for Hist-GBT, while AUC-PR was 0.977 and 0.979, respectively.These LOPO results substantially exceeded Logistic Regression, which achieved AUC-ROC = 0.792.
  • 79% of Random Forest folds and 83% of Hist-GBT folds exceeded AUC-ROC > 0.95 among 123 evaluable patient folds.93% of folds exceeded AUC-ROC > 0.90 for both models, and harder folds occurred across OSA-severity groups.
  • Random Forest and Hist-GBT had compact AUC-ROC distributions, with interquartile ranges of 0.961–0.994 and 0.968–0.996, respectively.Their corresponding medians were 0.982 and 0.988.
  • Random Forest showed mean sensitivity of 0.948 and specificity of 0.806, whereas Hist-GBT showed sensitivity of 0.937 and specificity of 0.862.The figure presents Random Forest as more sensitivity-oriented and Hist-GBT as offering a more balanced trade-off.
  • PSD features accounted for 30.3% of total importance, AUC features for 21.5%, and FFT-derived frequency descriptors for 15.1%.Together, these three groups accounted for 66.9% of total Random Forest mean decrease in impurity; the top-20 features accounted for 54.9%.

5. Discussion

Frequency-domain features provide the strongest discrimination for respiratory-event detection, while AUC and curve-length features add complementary information. Lower-ranked wavelet and nonlinear groups contribute less under this validation protocol, supporting compact feature selection but with important scope limitations.

  • Dominant Role of Frequency-Domain Features: 45.4% of total MDI comes from frequency-domain features, combining FFT spectral-shape descriptors at 15.1% and PSD band-power features at 30.3%.FFT descriptors average 2.51% MDI per feature, the highest group-level average.
  • Physiological Interpretation: The 0.1–0.4 Hz breathing band captures event-related changes in respiratory modulation, while FFT descriptors capture shifts in spectral centroid, spread, and overall shape.Band-power features quantify magnitude across channels; spectral-shape descriptors capture the broader distributional shift.
  • Complementary Time-Domain Evidence: 21.5% of total MDI is supplied by AUC and curve-length features, making them the main complementary time-domain category.Their 1.20% average contribution per feature is comparable to band-power features at 1.26%.
  • Effect of Preprocessing: Adaptive preprocessing concentrates the respiratory signature in the prep channel, whose top individual features include spectral shape and 0.1–0.4 Hz band power.Raw spatial aggregates also rank highly through breathing-band power, indicating that this signature remains spatially distributed.
  • Practical Feature-Group Recommendations: Frequency-domain and AUC features together account for 67% of discriminative information using 48 features across six channels, supporting compact implementations.General statistical descriptors add 22.6% across 78 features but are described as nonessential when computation is constrained.
  • Limitations: The ranking is constrained by Random Forest MDI, binary pooling of respiratory-event types, and controlled in-hospital recording conditions.The authors caution that the ranking may differ for finer event labels or home settings with varying mattresses, positions, and co-sleeping.

6. Conclusion

The patient-independent comparison identifies frequency-domain features as the dominant discriminative category, with adaptive preprocessing strengthening spectral features. Time-domain AUC and curve-length features provide the main complement, supporting a compact 48-feature detector.

  • 45.4% of total Random Forest discriminative information comes from frequency-domain features, led by breathing-band power at 30.3% and FFT spectral-shape descriptors at 15.1%.Breathing-band power spans all six channels, while FFT spectral-shape descriptors are computed on the preprocessed channel alone.
  • 2.51% mean decrease in impurity per feature is achieved by preprocessed-channel FFT spectral-shape descriptors, approximately twice the raw-channel average.The top three individual features all belong to the preprocessed channel.
  • 21.5% of MDI comes from AUC and curve-length features, making them the main complementary group through sensitivity to cumulative pressure displacement changes during events.
  • 10.4% of MDI is contributed by wavelet-derived, Higuchi fractal-dimension, frame-energy, and entropy features across 59 features, the weakest tier.
  • 48 frequency-domain and time-domain AUC features account for 67% of MDI and can achieve near-maximal discriminative performance in resource-constrained implementations.General statistical features may provide a modest further gain, while wavelet-derived, frame-energy, and nonlinear features can be omitted without substantial performance loss.
Loading 2608.28242v1…