The strongest signal came from frequency-based features in ballistocardiography, or BCG, a pressure signal recorded from a sensor mat. In a preprint analysis of 155 patients undergoing hospital evaluation for obstructive sleep apnea, two nonlinear machine-learning models reached area-under-the-curve scores near 0.97 when tested on patients they had not seen during training. The result points toward compact BCG detectors using fewer feature types, although the study did not test the proposed smaller feature set directly.
The strongest results came from Random Forest and Histogram Gradient Boosting. Their receiver-operating-characteristic AUC scores were 0.967 and 0.969, respectively, while their precision-recall AUC scores were 0.977 and 0.979. AUC summarizes how well a model separates event windows from clean windows across decision thresholds, with higher values indicating better separation in this comparison. Logistic Regression reached 0.792 on AUC-ROC, and a baseline that always chose the majority class scored 0.500.
What the models were seeing
The researchers built a library of 191 features across ten groups and six signal channels. The feature groups included frequency and spectrum measures, along with time-domain and other signal descriptions. The authors interpret breathing-band power in the frequency spectrum and fast-Fourier-transform descriptors from an adaptively prepared channel as the leading sources of information. AUC and curve-length measures added complementary information, while lower-ranked wavelet, energy, entropy and fractal groups were less useful in the reported ranking.
That prepared channel was not simply a fixed average of the mat’s sensors. For each window, the method selected five sensors with the strongest respiratory periodicity, aligned their signal polarity and averaged them after high-pass filtering. The five ordinary aggregate channels supplied 110 features, while the prepared channel supplied 81, making 191 features for each window in the complete representation. Frequency-transform, energy, entropy and wavelet features were calculated only for the prepared channel.
The study compared respiratory-event windows with clean windows taken from the same recordings. It contained 28,786 event windows and 21,088 clean windows, a positive-to-negative ratio of 1.37. Windows shorter than ten seconds were padded symmetrically to ten seconds. Clean windows had to last at least 30 seconds and be separated from annotated events by a 30-second safety buffer.
A patient-by-patient test
The key test was leave-one-patient-out cross-validation, meaning that all windows from one participant were held back while the model learned from the others, and the process was repeated across the cohort. This is a stricter test of patient-independent discrimination than randomly mixing windows from the same people between training and testing. Four classifiers were compared: a majority-class baseline, Logistic Regression, a class-balanced Random Forest and Histogram Gradient Boosting.
The AUC summaries used 123 evaluable patient folds because some held-out folds did not contain both classes. F1 scores, which balance precision and recall, were calculated over all 155 patient folds. These details matter because the reported headline AUC values do not represent a complete 155-fold AUC summary.
A result with a narrow test bed
The cohort included 52 female and 103 male participants. Mean age was 49.5 years, with a standard deviation of 12.5 years, and mean body mass index was 30.3 kg/m2, with a standard deviation of 6.9. The recordings came from a 16-by-32 capacitive pressure mat containing 512 sensors and sampled at 50 hertz, used alongside a respiratory polygraphy system. Ethics approval and written informed consent were reported.
The recording streams were synchronized for all 155 participants using decoded electroencephalography sync pulses. One hundred recordings also had network-time synchronization, while 55 used the electroencephalography signal alone. The median clock offset was 1.7 seconds, 90 percent of participants were below four seconds, and the median encoding precision was 21 milliseconds, with 96 percent below 50 milliseconds.
Those design choices define what the findings do and do not establish. They support window-level prediction and feature ranking within this hospital cohort and this particular 512-sensor mat setup. They do not establish performance in homes, on other mattresses, in different body positions or with different sensor hardware, and they do not show that the recommended reduced feature set would retain the same performance because that direct test was not reported.
A guide for smaller detectors, not a clinical verdict
The authors argue that a compact detector should give priority to breathing-band spectral power, prepared-channel FFT measures, and AUC or curve-length features. That recommendation is a modeling conclusion, not proof that these features are causally responsible for respiratory events. The feature ranking came from Random Forest mean-decrease-in-impurity scores, a model-based measure that can favor features with many possible values.
The target also pooled obstructive apneas, hypopneas and central events into one binary category, so the most useful features could differ if those event types were separated. The preprint does not report an external-cohort test or a clinical diagnosis or treatment outcome. Further testing would be needed to learn whether the ranking survives changes in setting, hardware and event definition.
Paper data and sources
Original title: 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 et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text