Skip to content

AI-Driven Cardiorespiratory Signal Processing - Separation, Clustering, and Anomaly Detection

FieldValue
JournalOpen MIND
Published2026-02-09
AuthorsYasaman Torabi
DOI10.48550/arxiv.2602.09210

An 8-qubit hybrid quantum-classical network classified third heart sound and murmur recordings with 93.33 ± 2.9% test accuracy. The same work reports 26.8 dB signal-to-distortion ratio and 32.8 dB signal-to-interference ratio for its wavelet, masked, temporally consistent variational autoencoder. These results combine controlled manikin recordings with algorithms for single-channel separation, clustering, and binary classification of abnormal sounds. The quantum result is preliminary because it was evaluated on a simulator and a limited dataset.

Controlled manikin data supplies paired mixtures and reference sounds

Section titled “Controlled manikin data supplies paired mixtures and reference sounds”

The authors recorded HLS-CMDS with a 3M Littmann CORE Digital Stethoscope at a sampling frequency of 22 kHz. Each capture lasted 15 s and used the stethoscope’s heart, lung, or mixed-sound filters. Amplification reached ×40, while active noise cancellation suppressed ambient interference.

The CAE Juno clinical manikin reproduced heart and lung sounds through anatomically positioned speakers. The manikin was positioned upright in a quiet room, and the stethoscope remained at standard auscultation landmarks. Mixtures were recorded with both sources enabled. The authors then recorded each heart and lung source separately without moving the stethoscope. Because the manikin repeated prerecorded sounds, these files provided aligned reference signals for separation tests.

The dataset contains 535 WAV recordings, with 268 labeled female and 267 labeled male. The set includes 50 heart-only recordings, 50 lung-only recordings, and 3 × 145 mixed recordings with corresponding individual sources. Metadata covers 10 heart-sound types, 6 lung-sound types, and 12 anatomical locations.

The prerecorded manikin sounds originated from real patients, but acquisition was simulator based. Controlled recording removed biological noise and motion artifacts. It supports controlled algorithm comparisons, but simulator-based acquisition does not establish performance on human auscultation.

LingoNMF adds language-guided separation to periodic NMF

Section titled “LingoNMF adds language-guided separation to periodic NMF”

PL-NMF first decomposes each mixture through parallel multilayer NMF blocks. Autocorrelation selects the shorter-period output as heart sound and the longer-period output as lung sound. Affine scaling and offsets make signed recordings non-negative before factorization.

Extracted features include spectral centroid, RMS energy, zero-crossing rate, variance, mean frequency, and maximum amplitude. LLaMA uses these features to propose fundamental frequencies that update the NMF penalty. A retrieval-augmented generation pipeline grounds the prompts in cardiorespiratory acoustics literature. The model also returns feature observations and possible diagnoses as adjunct output.

Evaluation used 100 normal two-mixture cases and 210 manikin recordings containing normal and abnormal sounds. All metrics were calculated against the separately recorded reference sources. SDR measures signal-to-distortion ratio, SIR measures signal-to-interference ratio, and SAR measures signal-to-artifact ratio.

LingoNMF reported 22.3 dB SDR, 22.4 dB SIR, and 25.2 dB SAR in Table A.1. Its SDR exceeded all eight listed comparison methods, whose reported SDR values ranged from 7.1 to 17.3 dB. ECNet reached a higher SIR at 31.1 dB, while LSTM reached 26.7 dB.

LingoNMF improved heart-sound SDR and SIR over PL-NMF on both datasets at p < 0.01. Against conventional NMF baselines, lung-sound SDR and SIR improved at p < 0.01 on Dataset Two and p < 0.05 on Dataset One. The work reports heart-sound SDR and SIR improvements of up to 3 to 4 dB.

XVAE-WMT separates sources through masked latent representations

Section titled “XVAE-WMT separates sources through masked latent representations”

The model receives a 1000 × 128 spectrogram and uses two parallel encoders and decoders. Each encoder has three convolutional layers and produces a 128-dimensional latent vector. The decoders generate soft masks between 0 and 1 that partition mixture energy across the two sources. Temporal consistency loss penalizes abrupt changes between adjacent reconstructed frames. Wavelet inputs preserve time-frequency structure, and the mask keeps separated outputs tied to the observed mixture.

Dataset One combined Kaggle Respiratory Sound, CirCor DigiScope, and Chest Wall Recording databases. At 16 kHz, it supplied 24,383 heart and 18,144 lung segments in 1 s frames across training iterations. Dataset Two formed 25,000 random mixtures from resampled HLS-CMDS segments.

XVAE-WMT achieved 26.8 dB SDR, 32.8 dB SIR, and 28.6 dB SAR on Dataset One. On Dataset One, VAE-WT reached 25.8 dB SDR and 30.0 dB SIR, while plain VAE reached 8.7 dB SDR. The full method also exceeded all listed baselines on the three separation metrics on Dataset One.

SHAP ranked latent dimensions, and t-SNE visualized source grouping. The authors retained the top 75% of dimensions for final analysis. Its Dataset One Silhouette score was 0.345, while Dataset Two gave 0.324. Dataset Two’s highest Silhouette was 0.345 for VAE-M, so the clustering advantage was not uniform across datasets.

Chem-NMF uses bounded initialization for multilayer clustering

Section titled “Chem-NMF uses bounded initialization for multilayer clustering”

Chem-NMF factorizes data through multiple NMF layers and clusters the resulting activation maps. K-means supplies the cluster assignments, and feature basis matrices reconstruct the inputs. A bounding factor blends each new random basis with the mean basis from the preceding layer.

The authors use a chemical reaction analogy in which catalysts lower activation barriers. Their analysis shows a non-increasing cost under multiplicative updates. It also argues that bounded initialization lowers the probability of remaining in a local minimum.

Tests used 400 ORL face images resized to 32 × 32 and 400 balanced MNIST samples resized to 28 × 28. Cardiovascular clustering used 50 lung recordings in 6 classes and 50 heart recordings in 10 classes. The authors compared K-means, Gaussian mixture models, agglomerative clustering, and spectral clustering with and without NMF features. The experiments report better clustering than regular NMF and α-NMF, but no tabulated headline accuracy is provided.

QuPCG classifies two abnormal heart-sound classes with eight qubits

Section titled “QuPCG classifies two abnormal heart-sound classes with eight qubits”

The QuPCG model classified S3 against murmur segments from HLS-CMDS. Recordings were resampled to 4 kHz and segmented by cardiac cycle. A complex Morlet transform with 128 scales produced 32 × 32 scalograms. Max-pooling with a 4 × 4 kernel reduced them to 8 × 8 maps. Binarization and further reduction produced eight values for an eight-qubit circuit.

Qiskit version 0.45 ran three alternating quantum convolution and pooling stages. The final qubit was measured through the Pauli-Z expectation value. Table 6.1 reports mean ± standard deviation for four preprocessing variants.

MethodTrain accuracyTest accuracyTest loss
I-QuPCG51.06 ± 2.3%47.62 ± 2.8%1.10 ± 0.54
M-QuPCG74.29 ± 5.2%53.33 ± 4.0%0.91 ± 0.25
W-QuPCG91.43 ± 2.9%80.00 ± 6.1%0.73 ± 0.03
W-QuPCG+97.14 ± 4.6%93.33 ± 2.9%0.45 ± 0.12

W-QuPCG+ obtained its reported accuracy after 1000 training epochs. Its test loss was 0.45 ± 0.12, compared with 1.10 ± 0.54 for I-QuPCG. The wavelet variant reached 80.00 ± 6.1% test accuracy, versus 53.33 ± 4.0% for Mel and 47.62 ± 2.8% for grayscale.

Simulated backends cannot reproduce all physical noise sources of quantum hardware. The authors also cite dataset size and sample diversity as constraints. The test set contains only S3 and murmur classes, so broader cardiac conditions remain untested.

Sensor quality and integration constrain the signal pipeline

Section titled “Sensor quality and integration constrain the signal pipeline”

MEMS microphones use a deformable diaphragm and fixed backplate to convert acoustic pressure into electrical signals. Digital MEMS outputs can use I2S, time-division multiplexing, or pulse-density modulation. The review identifies dynamic range, motion interference, environmental noise, fabrication cost, and long-term biocompatibility as constraints.

The review describes NV-center diamond sensors for magnetic, thermal, and chemical sensing. A green laser excites the negatively charged center, microwave fields manipulate its spin state, and a photodetector reads spin-dependent fluorescence. Resonance shifts report magnetic, thermal, or chemical perturbations.

The results favor coordinated choices of transducer, sampling rate, preprocessing, and model. Separation tests rely on repeated paired recordings that clinical deployment may not provide. QuPCG depends on preserving temporal and spectral patterns before compression. Clinical claims require human validation across patient groups and cardiac conditions. Quantum deployment must also be tested on physical hardware with its noise and decoherence.

NV-center biosensing requires diamond with optically active nitrogen-vacancy defects for spin-state readout of magnetic, thermal, and chemical variations. 6CCVD’s diamond materials catalog includes single-crystal and polycrystalline diamond grown by MPCVD in plate and wafer formats. The paper does not specify NV defect density or readout requirements, so use the engineering support page to confirm those parameters.

View Original Abstract

This research applies artificial intelligence (AI) to separate, cluster, and analyze cardiorespiratory sounds. We recorded a new dataset (HLS-CMDS) and developed several AI models, including generative AI methods based on large language models (LLMs) for guided separation, explainable AI (XAI) techniques to interpret latent representations, variational autoencoders (VAEs) for waveform separation, a chemistry-inspired non-negative matrix factorization (NMF) algorithm for clustering, and a quantum convolutional neural network (QCNN) designed to detect abnormal physiological patterns. The performance of these AI models depends on the quality of the recorded signals. Therefore, this thesis also reviews the biosensing technologies used to capture biomedical data. It summarizes developments in microelectromechanical systems (MEMS) acoustic sensors and quantum biosensors, such as quantum dots and nitrogen-vacancy centers. It further outlines the transition from electronic integrated circuits (EICs) to photonic integrated circuits (PICs) and early progress toward integrated quantum photonics (IQP) for chip-based biosensing. Together, these studies show how AI and next-generation sensors can support more intelligent diagnostic systems for future healthcare.