Murmur: Clinical and Computational Overview

Updated 4 November 2025

Murmur is defined as an abnormal acoustic phenomenon observed in clinical contexts—such as heart murmurs from turbulent blood flow—and analogous physical systems like astrophysical emissions.
Its detection in clinical settings involves skilled auscultation and digital methods using grading scales and machine learning algorithms for accurate classification.
The concept extends to astrophysics and computational graphics, where murmur describes low-amplitude emissions and noise patterns analyzed with advanced signal processing techniques.

A murmur is broadly defined as an abnormal or sustained acoustic phenomenon, occurring either in physiological systems—such as the cardiovascular system—or, by analogy, as a low-level, persistent signal or fluctuation in physical systems or computational contexts. In clinical medicine, a heart murmur refers to atypical sounds arising from turbulent blood flow within the heart, typically heard via auscultation and discernible from normal heart sounds (S1, S2). In the context of laboratory physics and computer science, “murmur” may refer to phenomena such as low-amplitude variable emissions (“X-ray murmur” of black holes) or computational constructs (such as noise patterns using the Murmur hash function). This article presents a rigorous, multi-domain overview of “murmur,” emphasizing its principal usage in clinical auscultation, its characterization and detection, the emerging role in data-driven medicine, and applications in other scientific disciplines.

1. Acoustic and Physiological Basis of Murmur

Heart murmurs are produced by turbulent or non-laminar blood flow through the cardiac chambers, across valves, or via structural defects. Key acoustic features include:

Timing: Occurring during systole, diastole, or both; further subclassified (e.g., holosystolic, early/mid/late systolic/diastolic) (Oliveira et al., 2021).
Pitch and Frequency Content: Determined by the velocity and volume of flow; frequency distributions often in the 20–200 Hz range, distinguishable by appropriate filtering (e.g., bell mode of electronic stethoscopes) (Torabi et al., 4 Oct 2024).
Duration & Intensity: Quantified using grading scales (e.g., Levine I/VI to VI/VI), with digital data restricting upper grades due to lack of tactile feedback (Oliveira et al., 2021).
Shape and Quality: Morphological descriptors (crescendo, decrescendo, plateau, diamond) and qualitative terms (blowing, harsh, musical) encode the murmur’s acoustic envelope and spectral profile (Oliveira et al., 2021, Guo et al., 2022).

The precise characterization of murmurs is crucial for identifying underlying pathologies, such as valvular stenosis, regurgitation, or septal defects.

2. Clinical Detection and Grading of Murmurs

Traditional detection relies on skilled human auscultation, with accuracy highly operator-dependent (83–91% for clinical diagnosis compared to gold-standard echocardiography) (Gupta et al., 2019). Grading is typically based on the Levine scale:

Grade	Description
I/VI	Barely audible
II/VI	Soft but obvious
III/VI	Moderately loud or louder

With the prevalence of digital stethoscopes and large annotated datasets (e.g., CirCor DigiScope: 5,282 recordings, >200,000 annotated events) (Oliveira et al., 2021), algorithmic and automated grading and detection have become feasible and are increasingly reproducible. Automated systems now routinely encode temporal (timing), spectral (pitch), and morphological (shape) features for diagnostic and stratification purposes (Guo et al., 2022, Elola et al., 2022).

3. Signal Processing, Machine Learning, and Data Resources

Algorithmic analysis of murmurs utilizes time-domain, spectral, and time-frequency representations:

Feature Extraction: Mel spectrograms, wavelet scattering transforms, MFCCs, wavelet entropy, fractal and multifractal features are all in active use (Vimalajeewa et al., 2023, Patwa et al., 2023, Alam et al., 2018).
Modeling Approaches:
- End-to-end deep learning: Convolutional and recurrent neural networks (e.g., LSTM, RCNN, BiLSTM), providing both temporal and spectral feature learning (Alam et al., 2018, Patwa et al., 2023, Nie et al., 25 Jul 2024).
- Hybrid/Modular methods: Systems combining symbolic (e.g., logical or feature-based filtering) with neural modules to capture both discrete and continuous properties (Saha et al., 2022).
- Self-supervised and transfer learning: Leveraging large unlabeled datasets and contrastive objectives to improve robustness and generalizability (Ballas et al., 2022).
- LLMs for Audio: Recent works finetune large pre-trained audio-LLMs (Qwen2-Audio) for simultaneous multi-attribute murmur classification, with advanced segmentation frontends to enhance noise robustness and long-tail feature detection (Florea et al., 23 Jan 2025).

A summary of recent model architectures and their performance is provided below:

Model	Key Feature Extraction	Accuracy/F1	Data Source
1D-CNN + Wavelet Scattering (Patwa et al., 2023)	WST, denoising	F1 up to 79%	CirCor, CinC 2016
RCNN (Gupta et al., 2019)	Denoising, segmentation	F-beta 0.95, 95.5%	Clinical samples
Parallel CNN-BiLSTM (Alam et al., 2018)	Spectrograms, MFCCs	F1: 98%, Sens: 96%	Pooled (10,892 seg.)
Deep CardioSound (Guo et al., 2022)	Waveform, DenseNet ensemble	F1: 0.99 (sample)	CirCor
Qwen2-Audio LLM (Florea et al., 23 Jan 2025)	Segmented PCG, audio LLM	>95% features	CirCor, multiple datasets
Parallel-Attentive + Uncertainty (Zhang et al., 7 May 2024)	Mel-spectrograms, attention/conv	Weighted Acc: 79.8%	CirCor

Datasets such as the CirCor DigiScope (Oliveira et al., 2021), PCG 2016, PASCAL, and various manikin-recorded sets (Torabi et al., 4 Oct 2024) serve as the primary testbeds, offering high-fidelity clinical, simulated, and annotated murmur data.

4. Emerging Techniques: Uncertainty, Interpretability, and Multi-Step Reasoning

Interpretability and clinical trust in automated murmur detection are advanced through:

Uncertainty Estimation: Bayesian deep networks, Monte Carlo Dropout, and temperature scaling generate probabilistic outputs and calibrated confidence intervals; high uncertainty can triage cases for expert review (Zhang et al., 7 May 2024, Walker et al., 2023).
Multi-label and Multi-class Annotation: Multilabel architectures (e.g., Deep CardioSound) permit annotation across orthogonal axes: timing, pitch, grading, quality, and shape (Guo et al., 2022).
Modular Reasoning: Neuro-symbolic systems, such as MURMUR, explicitly construct reasoning paths for semi-structured data-to-text generation, enabling logical consistency and semantic coverage in interpretations (Saha et al., 2022).
Long-tail Feature Classification: Recent LLM-based approaches demonstrate accurate classification for rare or underrepresented murmur features, outperforming conventional neural pipelines (Florea et al., 23 Jan 2025).

These methodological advances enable superior discrimination, interpretability, and clinical alignment compared to both hand-crafted feature and monolithic neural approaches.

5. Non-Clinical Contexts: Murmur in Physics and Signal Processing

In astrophysics, the term "murmur" denotes the persistent, low-level emission, variability, or “flaring” observed in black hole systems. For example, Chandra’s decadal monitoring of M31* revealed an extended quiescent X-ray period, a dramatic outburst, and a subsequent state of heightened variable emission—interpreted as an “X-ray murmur” indicative of low-level accretion activity and episodic jet formation (Li et al., 2010).

In computational graphics, "murmur" refers to the Murmur hash function, which is implemented in shader pipelines to generate procedural noise fields without reliance on texture lookup tables, thus trading computational operations for reduced memory bandwidth (Valdenegro-Toro et al., 2019). The Murmur hash is chosen for cryptographic resistance to collisions and efficient 32-bit integer implementation, with effective application in noise-based rendering.

6. Present Challenges and Future Directions

Key challenges include:

Generalization to Unseen Patient Populations: Although many models deliver high accuracy on held-out datasets, performance in diverse, real-world, "user-independent" scenarios remains an active area.
Low-Prevalence (“long-tail”) Murmur Feature Detection: Accurate characterization of rare murmur subtypes, especially diastolic or low-grade murmurs, requires further augmentation and model robustness (Florea et al., 23 Jan 2025).
Clinical Integration and Human-AI Collaboration: Automated murmur analysis systems are increasingly positioned as decision-support tools, not replacements for expert cardiologists; quantification of uncertainty and interpretability remain vital for adoption.
Synthetic Data and Simulation: Manikin-based datasets and synthetic augmentation of rare murmur events are increasingly important both for model training and for explainable classifier evaluation (Torabi et al., 4 Oct 2024).
Physical and Computational Murmur Analogs: The application and significance of “murmur” in other domains (e.g., neutron transition detection in hidden sectors (Stasser et al., 2018, Stasser et al., 2020), astrophysical variability, or hash-based procedural noise) illustrate the concept's cross-disciplinary relevance.

7. Summary Table: Murmur Across Domains

Domain	Definition/Use	Core Methodologies
Clinical cardiology	Abnormal heart sounds due to turbulent blood flow	Auscultation, PCG recording, ML, DL
Bioacoustics/datasets	Annotated heart/lung sounds, simulated/synthetic	Digital stethoscope, expert labeling
Signal processing & ML	Acoustical event detection, interpretation, grading	Wavelet/MFCC/spectrogram, CNN, LLM
Astrophysics	Low-level X-ray variability from compact objects	Chandra X-ray analysis, timing studies
Computational graphics	Procedural noise using (Murmur) hash in GPU shaders	Hash-based gradient gen., GLSL
Particle physics	Faint neutron sector transitions ("murmur" detection)	Low-background, lead regenerator