⚡ electrical engineering

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

This research explores the integration of advanced AI models—ranging from generative and explainable AI to quantum neural networks—with next-generation biosensing technologies to enhance the separation, clustering, and anomaly detection of cardiorespiratory signals for intelligent healthcare diagnostics.

Original authors: Yasaman Torabi

Published 2026-02-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yasaman Torabi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are at a loud, crowded party. You are trying to have a conversation with a friend, but there is a loud bassline thumping from the speakers and someone else is laughing loudly nearby.

In this scenario, your brain is performing a miracle: it filters out the music and the laughter so you can focus only on your friend's voice. This is exactly what this PhD thesis is trying to do, but instead of a party, the "noise" is the sound of your lungs breathing, and the "voice" is your heart beating.

The researcher, Yasaman Torabi, has developed a high-tech "AI toolkit" to help doctors listen to the body more clearly. Here is a breakdown of the four main "superpowers" she built:

1. The "Smart Translator" (LingoNMF)

The Problem: When a doctor uses a stethoscope, the heart and lung sounds are often mashed together into one messy sound wave.
The Solution: Imagine you have a recording of a song where the singer and the drummer are playing at the same time. Usually, it's hard to separate them. Torabi created an AI that uses a Large Language Model (like a smarter version of ChatGPT) to help.
The Analogy: It’s like giving the AI a textbook on music. The AI "reads" the sound and says, "Wait, I know from my textbook that a heartbeat usually has this specific rhythm, while breathing is more like a long, flowing wave." Because the AI "understands" the language of biology, it can much more accurately untangle the heart from the lungs.

2. The "X-Ray Vision" for Data (XVAE-WMT)

The Problem: AI is often a "black box." It might tell a doctor, "This sound is abnormal," but it can't explain why. Doctors don't like being told "just trust me."
The Solution: She built a special type of AI called a Variational Autoencoder that is "Explainable."
The Analogy: Imagine a chef making a soup. A normal AI just gives you the bowl and says, "It's delicious." This AI is like a chef who provides a transparent recipe card, showing exactly how much salt, pepper, and garlic went into it. It shows the doctor exactly which parts of the sound wave triggered the "abnormal" alarm, making the AI a partner rather than a mystery.

3. The "Chemical Catalyst" (Chem-NMF)

The Problem: When teaching an AI to group similar sounds together (clustering), the AI often gets "stuck." It finds a pattern, thinks it's finished, but it's actually just stuck in a shallow rut (a "local minimum").
The Solution: She looked at Chemistry to solve a math problem.
The Analogy: Think of a ball trying to roll into a deep valley, but it gets stuck in a tiny pothole halfway down the hill. In chemistry, a catalyst is something that lowers the energy needed to get over a hump so the reaction can finish. She added a "mathematical catalyst" to the AI that helps it "jump out" of those tiny potholes so it can keep rolling until it finds the deepest, most accurate valley.

4. The "Quantum Detective" (QuPCG)

The Problem: As medical devices get smaller (like wearable patches), they need to be incredibly efficient. Traditional computers can sometimes be too "heavy" or slow for tiny, ultra-fast sensors.
The Solution: She explored Quantum Computing—the next frontier of technology.
The Analogy: A regular computer is like a librarian who looks through books one by one to find an answer. A Quantum computer is like a librarian who can read every book in the library all at the same time. She designed a "Quantum Neural Network" that takes a tiny, compressed "snapshot" of a heartbeat and uses the strange, lightning-fast rules of quantum physics to spot abnormalities (like heart murmurs) with incredible speed and accuracy.

The Big Picture

By combining better sensors (the "ears"), smarter math (the "brain"), and quantum speed (the "reflexes"), this research is paving the way for a future where a tiny, wearable patch could listen to your heart and lungs 24/7, instantly telling you if something is wrong before you even feel a symptom.

Technical Summary: AI-Driven Cardiorespiratory Signal Processing

Title: AI-Driven Cardiorespiratory Signal Processing: Separation, Clustering, and Anomaly Detection
Author: Yasaman Torabi (McMaster University)
Field: Electrical and Computer Engineering / Signal Processing / Artificial Intelligence

1. Problem Statement

The accurate assessment of cardiopulmonary health relies heavily on auscultation (listening to heart and lung sounds). However, clinical diagnosis faces significant technical hurdles:

Signal Overlap: Heart and lung sounds are often recorded simultaneously, creating mixed signals that are difficult to interpret.
Noise and Artifacts: Ambient noise, motion, and biological interference obscure critical physiological features.
Data Scarcity: There is a lack of high-quality, annotated, and diverse datasets for training robust AI models.
Interpretability: Traditional "black-box" AI models lack the transparency required for clinical trust.
Hardware Constraints: Real-time monitoring requires efficient algorithms that can eventually run on low-power, edge-computing, or even quantum-enhanced hardware.

2. Methodology

The thesis adopts a multi-disciplinary approach, spanning from hardware sensing to advanced generative and quantum machine learning. The methodology is divided into four primary pillars:

A. Data Acquisition (The HLS-CMDS Dataset):
The author developed the Heart and Lung Sounds Clinical Manikin Dataset (HLS-CMDS). Using a digital stethoscope and a clinical simulation manikin, the researcher recorded 535 high-fidelity audio files (22 kHz). This dataset is unique because it provides both "pure" (separated) and "mixture" recordings, providing a reliable ground truth for training separation algorithms.

B. Blind Source Separation (BSS) via Generative AI:

LingoNMF: A novel algorithm that integrates Large Language Models (LLMs) with Non-negative Matrix Factorization (NMF). It uses a Retrieval-Augmented Generation (RAG) pipeline to feed physiological knowledge (e.g., expected fundamental frequencies) back into the NMF cost function as a penalty term, guiding the separation process.
XVAE-WMT: An Explainable Variational Autoencoder that utilizes wavelet-based inputs and a Temporal Consistency (TC) loss. It employs "output masking" to ensure the reconstructed signals sum back to the original mixture, maintaining physical consistency.

C. Unsupervised Clustering via Physics-Inspired Algorithms:

Chem-NMF: A multi-layer $\alpha$ -divergence NMF algorithm inspired by physical chemistry. It treats the optimization process like a chemical reaction, using a "bounding factor" (analogous to a catalyst) to lower the "activation energy" (optimization barriers), thereby preventing the algorithm from getting trapped in local minima.

D. Anomaly Detection via Quantum Machine Learning:

QuPCG: A Quantum Convolutional Neural Network (QCNN). The methodology involves a classical preprocessing pipeline (Wavelet transform $\rightarrow$ downsampling $\rightarrow$ binarization) that compresses 2D scalograms into an 8-pixel map. These 8 pixels are then encoded into 8 qubits for processing through quantum convolution and pooling layers.

3. Key Contributions

Algorithmic Innovation: Bridging the gap between Natural Language Processing (LLMs) and traditional signal decomposition (NMF).
Theoretical Rigor: Providing a formal mathematical convergence analysis for multi-layer NMF using Boltzmann probability and energy barrier analogies.
Explainability (XAI): Integrating SHAP (SHapley Additive exPlanations) and latent-space visualization to make the internal representations of generative models interpretable for clinicians.
Quantum-Classical Hybridization: Demonstrating a viable pathway for using near-term quantum computing (NISQ era) for specialized biomedical pattern recognition.
Hardware-Software Co-design Perspective: A comprehensive review linking the evolution of biosensors (from MEMS to Integrated Quantum Photonics) to the requirements of AI algorithms.

4. Results

Separation Performance: LingoNMF improved the Signal-to-Distortion Ratio (SDR) and Signal-to-Interference Ratio (SIR) by 3–4 dB compared to standard NMF. XVAE-WMT achieved an impressive SDR of 26.8 dB and SIR of 32.8 dB.
Clustering Efficiency: Chem-NMF demonstrated superior stability and accuracy in clustering tasks (ORL and MNIST datasets) compared to standard $\alpha$ -NMF, particularly when using the bounding factor to stabilize convergence.
Quantum Accuracy: The QuPCG model achieved a test accuracy of 93.33 ± 2.9% in classifying cardiac abnormalities (S3 vs. murmurs), proving that highly compressed quantum circuits can effectively capture complex physiological correlations.

5. Significance

This research represents a significant leap toward Intelligent Auscultation. By combining the reasoning capabilities of LLMs, the generative power of VAEs, the mathematical stability of physics-inspired NMF, and the computational potential of Quantum AI, the work provides a roadmap for the next generation of diagnostic tools. These tools promise to be more accurate, more interpretable, and capable of running on highly integrated, chip-scale quantum-photonic biosensors, ultimately enabling more accessible and personalized healthcare.