Spectral Features Dominate BCG Respiratory-Event Detection: A Large-Scale Patient-Independent Comparison of Feature Groups in Sleep Apnea Patients
Authors: Israel Campero Jurado, Zoe Bousraou, Lara Benning, Sara Padilla Neira, Alexander Breuss, Robert Riener, Esther Irene Schwarz, Elisabeth Wilhelm
Organizations: Department of Engineering and Technology Institute Groningen, Faculty of Science and Engineering, University of Groningen, The Netherlands · Department of Pulmonology, University Hospital Zürich, Switzerland · Sensory-Motor Systems Lab, ETH Zürich, Switzerland
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.
Figures & tables
Study
Signal / modality
General stat.
Time domain
Frequency domain
Wavelet / TF
Nonlinear / complexity
Spatial / fusion
Zhao et al. [ 44 ]
BCG-derived HBI + STC-Min breathing effort
NR
✓
NR
preproc.
✓
NR
Waltisberg et al. [ 42 ]
8-sensor under-mattress strain-gauge array
✓
✓
✓
NR
NR
✓
Wang et al. [ 43 ]
BCG-derived RR intervals
✓
NR
✓
preproc.
✓
NR
Gao et al. [ 9 ]
Mattress BCG / respiratory signal; HRV + similarity features
Table 1: Feature categories used in prior BCG and bed-pressure studies for sleep-disordered breathing or OSA-related detection. ✓ = used as a classifier or decision feature; preproc.=used during preprocessing, signal extraction, or fusion but not as a final classifier-feature category; NR =not reported. PVDF stands for polyvinylidene fluoride, a piezoelectric polymer material.
Figure 1: End-to-end processing pipeline. Step 1: simultaneous acquisition from the Alice 6 PSG system and a Raspberry Pi data logger recording the 16×32 BCG at 50 Hz. Step 2: EEG sync-pulse alignment maps expert-scored events onto the BCG timeline. Step 3: six signal channels are derived and processed through standardised feature groups, yielding a 191-dimensional feature vector per window. Step 4: leave-one-patient-out cross-validated classifiers produce binary predictions (respiratory event vs. event-free window).
Feature group
Features included
Count
Applied to
StandardStat
Mean, median, standard deviation, variance, kurtosis, skewness, range
Total power; band power at 0.1–0.6, 0.1–0.4, 0.4–0.8 Hz
4
✓
HiguchiFD
Higuchi fractal dimension
1
✓
Table 2: Implemented feature groups and their per-channel application. ✓ = computed for all six channels; prep =computed only for the adaptively preprocessed channel.
Figure 2: Per-patient class distribution in the held-out test fold. (a) Event (red) and clean (blue) windows per patient, sorted by event-window count. Orange triangles mark the 29 patients whose test fold contained no clean windows after the 30 s event-exclusion buffer; these folds have undefined AUC-ROC but well-defined F1 scores. (b) Distribution of the event-window fraction per patient. The majority of patients have more event than clean windows (mean fraction =0.64 ), motivating the use of a balanced class weighting in the Random Forest and Hist-GBT.
Characteristic
Value
Participants
Sex (female / male)
52 / 103
Age, mean ± SD (years)
49.5 ± 12.5
BMI, mean ± SD (kg/m 2 )
30.3 ± 6.9
AHI, median (IQR) (events/h)
14.4 (7.8–27.6)
OSA severity (none / mild / moderate / severe)
15% / 36% / 27% / 22%
Table 3: Cohort demographics and dataset statistics. ± denotes standard deviation.
Figure 3: EEG sync-pulse protocol performance across 155 participants. (a) Empirical cumulative distribution of the detected absolute clock offset ∣Δt∣ for the NTP+EEG group. The green dashed line marks the median offset (1.7 s), and the grey dotted line marks the 90th percentile (4.0 s). Eight participants with offsets above 30 s are clipped at 30 s for display. (b) Encoding quantisation error for all participants ( N=155 ), defined as the absolute difference between the reconstructed timestamp tPi,0+s^ and the direct wall-clock timestamp logged by the RPi at pulse transmission. Median 21 ms; 96% of participants below 50 ms.
Classifier
AUC-ROC
AUC-PR
F1
Bal. Acc.
Sensitivity
Specificity
Majority-class
0.500±0.000
0.565±0.212
0.742±0.217
0.584±0.212
1.000±0.000
0.000±0.000
Logistic Reg.
0.792±0.104
0.840±0.119
0.702±0.210
0.705±0.125
0.678±0.206
0.736±0.229
Random Forest
0.967±0.044
0.977±0.036
0.905±0.153
0.895±0.119
0.953±0.065
0.807±0.245
Hist-GBT
0.969±0.052
0.979±0.040
0.915±0.151
0.912±0.100
0.944±0.077
0.862±0.193
Table 4: LOPO cross-validation performance: mean ± SD across N=155 patient folds (AUC-ROC and AUC-PR computed over the 123 folds with evaluable negative windows; F1 over all 155). Best values per metric are highlighted in bold .
Figure 4: Per-patient LOPO AUC-ROC as a function of clinical AHI and OSA severity, including no OSA, mild OSA, moderate OSA, and severe OSA. (a) Random Forest and (b) Hist-GBT: each point represents one held-out patient, coloured by OSA severity category (AASM thresholds). The dashed horizontal line marks AUC-ROC =0.95 . Vertical tick marks at the top edge indicate the 29 patients whose test fold contained no eligible negative windows after applying the 30 s event-exclusion buffer; AUC-ROC is undefined for these folds (F1 scores ranged from 0.90 to 1.00). (c) Distribution of Random Forest AUC-ROC stratified by severity category (box: IQR; white line: median; whiskers: 1.5× IQR; jittered points: individual folds). Harder cases (AUC-ROC <0.95 ) are distributed across all severity groups, with mild-OSA patients most frequently represented.
Figure 5: Per-patient LOPO performance distributions shown as box-and-whisker plots overlaid with individual fold scores (jittered points) for Logistic Regression, Random Forest, and Hist-GBT across four metrics. Boxes span the interquartile range; horizontal white lines show medians; whiskers extend to 1.5× IQR. The Majority-class Dummy baseline is omitted for clarity (fixed AUC-ROC =0.5 ). Both non-linear classifiers exhibit tight, high-valued distributions, confirming robust performance across diverse patients.
Figure 6: Per-patient sensitivity versus specificity for Random Forest (left) and Hist-GBT (right), with point colour encoding AUC-ROC (green = high, red = low). Dashed and dotted lines indicate per-classifier mean sensitivity and specificity, respectively. Random Forest achieves slightly higher mean sensitivity ( μ=0.948 ) at the cost of lower mean specificity ( μ=0.806 ), while Hist-GBT achieves a more balanced trade-off (sensitivity 0.937 , specificity 0.862 ), making the two classifiers suitable for different clinical deployment priorities.
Figure 7: Top-20 Random Forest feature importances ranked by mean decrease in impurity (MDI). Colour encodes the signal channel: preprocessed signal ( prep ) , total mat ( raw_total ) , upper zone (orange, raw_upper ), mid zone (purple, raw_mid ), lower zone (teal, raw_lower ), and top-5 sensor mean (dark blue, raw_mean5 ). The top-20 features account for 54.9% of the total MDI. FFT-derived frequency descriptors and breathing-band power of prep dominate ranks 1–3, with band-power features across all remaining signal channels occupying ranks 4–8, and AUC and wavelet-derived features providing complementary information.
Rank
Feature
Signal channel
Feature group
Importance
Cumulative
1
prep-SpectralStd
prep
FFT
0.0664
6.6%
2
prep-SpectralMean
prep
FFT
0.0604
12.7%
3
prep-Bandpower0.1-0.4
prep
PSD
0.0355
16.2%
4
raw_upper-Bandpower0.1-0.4
raw_upper
PSD
0.0350
19.7%
5
raw_mid-Bandpower0.1-0.4
raw_mid
PSD
0.0347
23.2%
6
raw_lower-Bandpower0.1-0.4
raw_lower
PSD
0.0320
26.5%
Table 5: Top-30 Random Forest feature importances. Importance denotes normalised mean decrease in impurity (MDI). The cumulative column reports the cumulative percentage of total MDI explained up to each rank.
Feature group
Importance
Share of total
PSD
0.3029
30.3%
AUC
0.2154
21.5%
FFT
0.1508
15.1%
InterQuartile
0.1206
12.1%
StandardStat
0.1048
10.5%
Wavelet
0.0687
6.9%
Table 6: Random Forest importance aggregated by implemented feature group. Values are normalised MDI summed across all features in each group.
Signal channel
Importance
Share of total
prep
0.3074
30.7%
raw_upper
0.1656
16.6%
raw_total
0.1557
15.6%
raw_mid
0.1553
15.5%
raw_lower
0.1420
14.2%
raw_mean5
0.0741
7.4%
Table 7: Random Forest importance aggregated by signal channel. Values are normalised MDI summed across all features extracted from each channel.
Feature group
Applied to
N
MDI %
Per-feat. %
Rec.
Freq-domain: PSD band power
All channels
24
30.3
1.26
Include
Time-domain: AUC / curve length
All channels
18
21.5
1.20
Include
Freq-domain: FFT shape
prep only
6
15.1
2.51
Include
Gen. stat: interquartile range
All channels
36
12.1
0.33
Consider
Gen. stat: central/spread
All channels
42
10.5
0.25
Consider
Wavelet ( bior3.9 )
prep only
45
6.9
0.15
Optional
Table 8: Feature-group MDI ranking. MDI % is the fraction of total Random Forest mean decrease in impurity attributed to each group, averaged across all 155 LOPO folds. Per-feature average = MDI % divided by the number of features in the group. Recommendation: Include = high information per feature; Consider = moderate contribution spread over many features; Optional = low per-feature contribution.
Sleep-disordered breathing, including snoring, hypopnea, and obstructive apnea, is the hallmark of obstructive sleep apnea, whose severity is quantified by the Apnea-Hypopnea Index (AHI). As conventional polysomnography is costly and clin- ically demanding, microphone-only detection has emerged as a noninvasive alternative for screening. We propose ALE- Confidence Fusion Sound Event Detection (ACF-SED), the first framework that directly injects the per-frame confidence map of an Adaptive Line Enhancer (ALE) into Transformer atten- tion biasing and feature modulation, rather than as mere pre- processing. A Multi-Resolution ALE Bank aggregates three parallel NLMS filters at distinct decorrelation delays, while a Confidence-Guided Cross-Path block couples event and noise streams via gated cross-attention. On the Audio-Polygraphy Dataset for Sleep Apnea Analysis, ACF-SED achieves state- of-the-art detection across Event-F1, Segment-F1, and PSDS, enabling end-to-end AHI estimation.
Chanwoo Park, Chanwoo Kim
Department of Artificial Intelligence, Korea University, South Korea
Respiratory acoustic foundation models (FMs) are benchmarked exclusively on smartphone recordings, yet clinical deployment increasingly targets body-coupled (BC) wearables whose sensors attenuate high-frequency content through tissue and bone, leaving FM reliability uncharacterised. We introduce BCoughBench, evaluating five FMs (OPERA-CT/CE/GT, HeAR, M2D+Resp) on nine classification tasks (AUROC, sensitivity at 95% specificity, Expected Calibration Error) and three age regression tasks (MAE vs. a mean-predictor baseline) across five EBEN-simulated BC sensor conditions on five labeled cough datasets. Mean AUROC declines from 0.785 (smartphone) to 0.689-0.723, degrading most under temple vibration pickup (Δ = -0.096) and least under the soft in-ear (Δ = -0.062). No FM meets the clinical sensitivity threshold (Se@Sp95 ≥ 0.20) on most disease tasks under any BC sensor. Sex classification on the CIDRZ cohort collapses (AUROC 0.954 to 0.596-0.628, Δ = -0.341) while COVID detection is nearly unaffected (Δ = -0.004). Age regression is robust, improving under the forehead accelerometer on CoughVID (MAE 9.61 to 8.97 yr); HeAR leads on regression and demographic tasks, M2D+Resp on disease and characteristic tasks. BCoughBench provides a reproducible framework for FM evaluation under wearable conditions.
Mayur Sanap, Prasanna Desikan, Edgar Lobaton
Centific Global Solutions Inc. · USA · North Carolina State University
Foundation models for wearable biosignals have matched or exceeded supervised specialists across a range of clinical tasks, yet all rely on modalities that require deliberate user action--wearing a device or visiting a sleep lab. We introduce BCG-FM, the first foundation model for ambient mechanical biosignals. A piezoelectric sensor embedded in the bed surface records ballistocardiography (BCG) each night without user effort; we pretrain BCG-FM with participant-level contrastive learning and using a total of 2.75 million hours of nightly recordings from 145,985 individuals, the largest raw-waveform biosignal pretraining corpus to date. Frozen BCG-FM embeddings achieve 3.26-year MAE on biological-age estimation (the lowest reported for any ambient, contactless modality) and yield clinically relevant discrimination across 15 self-reported health conditions and three independent external cohorts. Pretrained representations from only 500 labeled participants outperform a fully supervised baseline trained on 3,372, and representation quality scales log-linearly with contrastive batch size. These results establish ambient, longitudinal mechanical biosignals as a viable modality for health foundation models.
Magnus Ruud Kjaer, Haejun Han, Ashish Neupane +1
AI/ML, Eight Sleep · AI/ML, Eight Sleep (work done while at Eight Sleep)