Edge-AI for Heart Murmur Screening in Low-Resource Settings

arXiv:2511.00966 · eess.SP · Submitted 2025-11-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Edge-AI for Heart Murmur Screening in Low-Resource Settings".

Dev: Here's the gist of the paper. Early and reliable detection of heart murmurs is essential for the timely diagnosis of cardiovascular diseases, yet traditional auscultation remains subjective and dependent on expert interpretation This work investigates artificial intelligence (AI)-based murmur detection using the CirCor Heart Sound dataset, with a focus on enabling uncertainty-aware, resource-efficient deployment on edge devices Results show that lightweight models can achieve accuracy comparable to deeper networks (≈91%) while requiring two orders of magnitude fewer parameters Incorporating uncertaintybased selective classification further improved sensitivity by 3%, enhancing robustness and clinical reliability The findings highlight the feasibility of developing computationally efficient, uncertainty-aware AI systems for heart murmur screening in low-resource and remote healthcare settings Dataset Description The CirCor DigiScope Phonocardiogram Dataset contains a large number of pediatric heart-sound recordings collected during two mass-screening campaigns in Brazil in 2014 and 2015 The dataset includes 5272 phonocardiogram (PCG) recordings from 1568 subjects with a mean age of 6.1 ± 4.3 years All PCGs were recorded using the Littmann 3200 electronic stethoscope, and individual recordings last from 4.8 to 80.4 s The dataset includes a subject description file and segmentation file for every recording, containing information regarding the start and end points of the fundamental heart sounds S1 and S2 Every audio recording is annotated with three possible labels: murmur absent, murmur present or unknown The entire dataset is partitioned into training (60%) and test sets (30%) and validation sets (10%) Feature Engineering Pipeline Frequency-domain features were extracted from the raw audio signals using the STFT technique This technique computes the Fast Fourier Transform (FFT) over short, overlapping segments of the signal after multiplying each segment by a windowing function, specifically a Hann window The segment length, denoted as NF F T, is typically chosen as a power of two to improve computational efficiency To focus on the clinically relevant frequency range, only the lower half of the frequency bins were retained, effectively applying a low-pass filter with a cutoff at 1 kHz The signal was further segmented into 2-second windows with a 1-second overlap to ensure each segment captures at least one complete systolic-diastolic cardiac cycle Segments with a Power Spectral Density ratio (P SDratio) below a predefined threshold ”psd thr” were considered low-quality and discarded Model Comparison and Selection Three CNN variants, named Light, Baseline, and Heavy, were developed to detect the presence of cardiac murmurs using the extracted 2D features The architectures are detailed in Table II The training process involved reformulating the problem as a binary classification task (Absent vs.

Rosa: Present) by excluding 'Unknown' samples from the training set Patient-level classification is derived by aggregating predictions across all segments and recording locations using a ”Voter” block The final model selection was based on cross-validation results, where NF F T = 128 and psd thr = 0.45 were selected for the final models as they obtained the better F1-score Uncertainty Quantification Strategy Monte Carlo Dropout (MCD) was applied during inference with a number of stochastic forward passes N = 10 to estimate prediction reliability The predictive uncertainty was quantified using the expected entropy of the softmax outputs in the MCD samples, defined as E = −Xpi ∗ log(pi) A prediction coherence score (C) was computed as C = 1 − 4 ∗ var(pi), where pi is the vector containing the predicted class of the 10 MCD samples An overall confidence score (CS) was introduced that combines uncertainty and coherence as CS = α ∗ (1 − E) + (1 − α) ∗ C with α = 0.5 This confidence score enabled a selective classification strategy where unreliable predictions were discarded, resulting in improved robustness and recall Resource Efficiency and Deployment The analysis of model complexity showed significant resource savings achieved by the Light model, which reduces FLASH memory usage by a factor of 102× compared to the Heavy model The Heavy model cannot be deployed on any of the selected boards, highlighting the challenges of implementing large networks on hardwareconstrained devices Quantization, converting network weights from 32-bit floating point (float32) to 8-bit integers (int8), reduces model size by approximately a factor of four and lowers inference latency by more than 4× The combination of the Light model and quantization represents the optimal solution for on-the-edge implementation, providing significant advantages in terms of memory and latency Clinical Impact and Future Work The integration of uncertainty estimation through MCD demonstrated a positive impact on the robustness and sensitivity of the murmur detection The application of selective classification based on the uncertainty threshold led to an increase in Recall (+3%), which implies a reduction in missed murmur detections The uncertainty analysis on the confidence score revealed significantly lower confidence scores for the unknown test samples, mirroring diagnostic uncertainty among clinical experts Future research will focus on model generalization through external validation and full edge deployment via firmware integration Explainability techniques could be combined with uncertainty quantification to provide clinicians and non-expert users with interpretable visual feedback The study acknowledges limitations, such as the reliance on the CirCor dataset and the approximate nature of Monte Carlo Dropout The overall performance remains stable and reliable regardless of the individual model variant, in line with state-of-the-art approaches Comparison Against Existing Work The proposed trade-off model, achieving an accuracy of 91%, is in line with the state of the art when compared against existing heart murmurs classifiers Specifically, the Light model's performance is comparable to other methods like Niizumi

34: and Guan

21: , while significantly outperforming approaches like McDonald

17: in terms of parameter reduction The overall pipeline demonstrates a robust methodology for murmur detection, with performance that remains stable and reliable regardless of the individual model variant The authors conclude that an accurate detection is feasible even without relying on high-performance computers Evaluation Metrics Summary Model performance was assessed using standard metrics for binary classification, specifically the classification accuracy and the F1-score These metrics were computed at the location level and then aggregated to assess overall model performance on a patient-level basis Precision quantifies the proportion of correctly identified positive cases among all samples predicted as positive, while Recall (or Sensitivity) measures the proportion of correctly identified positive cases among all actual positives The final results show that for the Light model, patient-level classification employing confident segments achieved an F1-score of 81.

Title and authors: Rosa: So, this paper is about taking those complex heart sound readings and making them work on devices that don't have much power or memory.

Dev: Yeah, they’re talking about using AI to detect heart murmurs where you aren't in a fancy hospital with all the big machines.

Rosa: Exactly. The main point is showing that you can get a decent accuracy, like ninety one percent, with a very small model that takes up way less space than the bigger ones.

Dev: And they’re not just talking about size; they’re talking about making sure the system actually works reliably when things go wrong on those smaller chips.

Rosa: They introduced this whole uncertainty thing, using something called Monte Carlo Dropout, to see how confident the AI is in its own prediction.

Dev: That uncertainty measurement lets them do selective classification, which means if the AI isn't sure about a reading—maybe it’s an unknown sample—it just skips it instead of guessing.

Rosa: They found that this filtering actually made their sensitivity better by three percent, which is important because you don't want to miss a murmur when you need to catch it early.

Dev: It also gave the system a way to tell you how reliable its answer is through this overall confidence score they built, which helps in making safer decisions about the screening results.

Rosa: The practical side of this is that they showed how using a light model and then shrinking it down with quantization—turning those big floating point numbers into smaller integer numbers—is the best way to get it running fast on things like microcontrollers.

Dev: That combination cuts memory usage by over one hundred times compared to the heavier models, which means you can actually put this kind of detection onto hardware that couldn't handle anything bigger before.

Rosa: So, for someone listening, this paper means that accurate heart sound screening isn't just a thing for big labs anymore; it’s something that can potentially be deployed in remote clinics or even on mobile devices.

Dev: It moves the focus away from needing massive computing power to achieve high accuracy, focusing instead on making the AI smart enough to handle real-world limitations and uncertainty gracefully.

Rosa: And because they tested this method outside of their main lab setup, it gives us some insight into how this stuff might hold up when you take it out into the field where things get messy.

Dev: Which brings us to what we need to watch next, how these kinds of efficient AI models handle actual failure modes in those tough environments.

The paper's summary: Rosa: So, looking at what they suggested for making this AI system better, they focused on how to truly leverage that uncertainty we talked about earlier to make the decisions smarter.

Dev: Right, they aren't just throwing a number out there; they are suggesting you use that confidence score to actually decide which recordings you trust and which ones you have to flag for a human expert.

Rosa: That selective classification strategy is key, because it means the system doesn't just give you an answer; it tells the user exactly when its knowledge might be shaky.

Dev: And they show that by using this uncertainty filtering, they actually gained three percent more sensitivity in catching those murmurs they would have otherwise missed.

Rosa: That improvement is significant because when you’re screening for something like a heart murmur, reducing false negatives means you’re catching more of the actual cases.

Dev: It also showed that this whole confidence score approach gives the user a much clearer picture of what the AI knows and what it doesn't.

Rosa: They found that when they looked at those confidence scores, the samples they labeled as unknown actually had way lower scores than the ones where the system was certain.

Dev: That’s a strong signal that you can use this to build a very robust safety net for your deployment, because it helps you spot exactly where the system is likely to fail or be unreliable.

Rosa: They are pushing toward using these models on actual microcontrollers, like the F401RE board mentioned in their analysis, and they show that with quantization applied, you can get a latency of two hundred thirty milliseconds.

Dev: That speed is what makes it viable for real-time monitoring; it means the system reacts fast enough to be useful in a clinical setting where timing matters.

Rosa: The big implication here is that this isn't just some theoretical exercise; they’re showing a path toward having these kinds of reliable, lightweight detection tools actually running on the hardware that would be in a remote clinic.

Dev: It means the next step for this research isn't just proving it works on paper, but actually getting it integrated into firmware so it can run continuously without needing a giant computer nearby.

Rosa: And once you get that deployed, we need to think about what happens when things go sideways outside of the perfect testing conditions they used in the CirCor dataset.

The paper's improvements: Taro: So we're wrapping up this look at "Edge-AI for Heart Murmur Screening in Low-Resource Settings" by Rosa and Dev, summarizing what they found about making heart murmur detection practical on small devices.

Rosa: Basically, they proved that you can use a very light AI model to detect murmurs with good accuracy without needing huge computers or massive amounts of memory.

Dev: And the whole point is that they built in a way for the system to know when it’s not sure about a reading so it can be selective and reliable.

Rosa: They showed that this approach, using lightweight models and uncertainty estimation, leads to an improvement in sensitivity, which means catching more of those potential heart murmurs.

Dev: It sounds like a solid path for getting these kinds of screening tools into places where they actually need to be deployed quickly.

Taro: I’m still thinking about how this uncertainty filtering works when the world misbehaves; if the AI gets unsure, what's the graceful recovery mechanism?

Rosa: Well, it handles uncertainty by discarding unreliable predictions, which is a form of graceful recovery because it prevents a wrong answer from being delivered.

Dev: That’s right. It’s not about trying to guess when you don't know anything; it’s about knowing when to stop and ask for more help or just skip that reading entirely.

Rosa: Overall, the paper "Edge-AI for Heart Murmur Screening in Low-Resource Settings" gives us a blueprint for building screening tools that don't need massive supercomputers to function effectively in remote or low-resource healthcare settings.

Dev: It shifts the focus from just getting a high accuracy number on a huge server to creating something that actually runs reliably on smaller, lower-power hardware in real time.

Taro: It’s promising, but I still want to see how this system handles the actual physical noise and distortion you get when you take it out into the field where things aren't perfectly controlled.

Rosa: That’s exactly what we need to look at next; moving from the lab results to real-world trials where those environmental factors actually hit the system.

Conclusion: Rosa: So we're wrapping up this look at "Edge-AI for Heart Murmur Screening in Low-Resource Settings" by Rosa and Dev, summarizing what they found about making heart murmur detection practical on small devices.

Dev: Basically, they proved that you can use a very light AI model to detect murmurs with good accuracy without needing huge computers or massive amounts of memory.

Rosa: And the whole point is that they built in a way for the system to know when it’s not sure about a reading so it can be selective and reliable.

Dev: They showed that this approach, using lightweight models and uncertainty estimation, leads to an improvement in sensitivity, which means catching more of those potential heart murmurs.

Taro: I’m still thinking about how that uncertainty filtering works when the world misbehaves; if the AI gets unsure, what's the graceful recovery mechanism?

Rosa: Well, it handles uncertainty by discarding unreliable predictions, which is a form of graceful recovery because it prevents a wrong answer from being delivered.

Dev: That’s right. It’s not about trying to guess when you don't know anything; it’s about knowing when to stop and ask for more help or just skip that reading entirely.

Taro: Overall, the paper "Edge-AI for Heart Murmur Screening in Low-Resource Settings" gives us a blueprint for building screening tools that don't need massive supercomputers to function effectively in remote or low-resource healthcare settings.

Rosa: It shifts the focus from just getting a high accuracy number on a huge server to creating something that actually runs reliably on smaller, lower-power hardware in real time.

Dev: It’s about making the technology practical for deployment, not just theoretical performance on a benchmark dataset like CirCor.

Taro: I still want to see how this system handles the actual physical noise and distortion you get when you take it out into the field where things aren't perfectly controlled.

Rosa: That’s exactly what we need to look at next; moving from the lab results to real-world trials where those environmental factors actually hit the system.

Politecnico di Torino · Aalborg University

eess.SP

Submitted: 2025-11-02

Updated: 2026-11-10

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 77/100

The gist: Early and reliable detection of heart murmurs is essential for the timely diagnosis of cardiovascular diseases…

Key concepts

CirCor Heart Sound Dataset
This is a large collection of pediatric heart sound recordings from mass screenings in Brazil. It includes information about the start and end points of heart sounds S1 and S2 for each recording, allowing researchers to train and test their AI models effectively.
Frequency-domain Features (STFT)
These features are extracted from raw audio signals using the Short-Time Fourier Transform. This technique breaks down a continuous sound wave into different frequencies over short time segments. The process filters out irrelevant low-quality sounds, focusing the analysis on clinically important frequency ranges.
Monte Carlo Dropout (MCD)
MCD is a method used during testing to estimate how certain an AI model is about its prediction. By running the model multiple times with random variations, it quantifies the predictive uncertainty. This helps determine if a detection result is highly reliable or just a guess.

Terminology

Summary

Early and reliable detection of heart murmurs is essential for the timely diagnosis of cardiovascular diseases, yet traditional auscultation remains subjective and dependent on expert interpretation This work investigates artificial intelligence (AI)-based murmur detection using the CirCor Heart Sound dataset, with a focus on enabling uncertainty-aware, resource-efficient deployment on edge devices Results show that lightweight models can achieve accuracy comparable to deeper networks (≈91%) while requiring two orders of magnitude fewer parameters Incorporating uncertaintybased selective classification further improved sensitivity by 3%, enhancing robustness and clinical reliability The findings highlight the feasibility of developing computationally efficient, uncertainty-aware AI systems for heart murmur screening in low-resource and remote healthcare settings

Dataset Description

The CirCor DigiScope Phonocardiogram Dataset contains a large number of pediatric heart-sound recordings collected during two mass-screening campaigns in Brazil in 2014 and 2015 The dataset includes 5272 phonocardiogram (PCG) recordings from 1568 subjects with a mean age of 6.1 ± 4.3 years All PCGs were recorded using the Littmann 3200 electronic stethoscope, and individual recordings last from 4. p. 2

Where the paper says it

All PCGs have been recorded using the Littmann 3200 electronic stethoscope. Individual recordings last 4.

8 to 80.4 s The dataset includes a subject description file and segmentation file for every recording, containing information regarding the start and end points of the fundamental heart sounds S1 and S2 Every audio recording is annotated with three possible labels: murmur absent, murmur present or unknown The entire dataset is partitioned into training (60%) and test sets (30%) and validation sets (10%)

Feature Engineering Pipeline

Frequency-domain features were extracted from the raw audio signals using the STFT technique This technique computes the Fast Fourier Transform (FFT) over short, overlapping segments of the signal after multiplying each segment by a windowing function, specifically a Hann window The segment length, denoted as NF F T, is typically chosen as a power of two to improve computational efficiency To focus on the clinically relevant frequency range, only the lower half of the frequency bins were retained, effectively applying a low-pass filter with a cutoff at 1 kHz The signal was further segmented into 2-second windows with a 1-second overlap to ensure each segment captures at least one complete systolic-diastolic cardiac cycle Segments with a Power Spectral Density ratio (P SDratio) below a predefined threshold ”psd thr” were considered low-quality and discarded

Model Comparison and Selection

Three CNN variants, named Light, Baseline, and Heavy, were developed to detect the presence of cardiac murmurs using the extracted 2D features The architectures are detailed in Table II The training process involved reformulating the problem as a binary classification task (Absent vs. Present) by excluding 'Unknown' samples from the training set Patient-level classification is derived by aggregating predictions across all segments and recording locations using a ”Voter” block The final model selection was based on cross-validation results, where NF F T = 128 and psd thr = 0.45 were selected for the final models as they obtained the better F1-score

Uncertainty Quantification Strategy

Monte Carlo Dropout (MCD) was applied during inference with a number of stochastic forward passes N = 10 to estimate prediction reliability The predictive uncertainty was quantified using the expected entropy of the softmax outputs in the MCD samples, defined as E = −Xpi ∗ log(pi) A prediction coherence score (C) was computed as C = 1 − 4 ∗ var(pi), where pi is the vector containing the predicted class of the 10 MCD samples An overall confidence score (CS) was introduced that combines uncertainty and coherence as CS = α ∗ (1 − E) + (1 − α) ∗ C with α = 0.5 This confidence score enabled a selective classification strategy where unreliable predictions were discarded, resulting in improved robustness and recall

Resource Efficiency and Deployment

The analysis of model complexity showed significant resource savings achieved by the Light model, which reduces FLASH memory usage by a factor of 102× compared to the Heavy model The Heavy model cannot be deployed on any of the selected boards, highlighting the challenges of implementing large networks on hardwareconstrained devices Quantization, converting network weights from 32-bit floating point (float32) to 8-bit integers (int8), reduces model size by approximately a factor of four and lowers inference latency by more than 4× The combination of the Light model and quantization represents the optimal solution for on-the-edge implementation, providing significant advantages in terms of memory and latency

Clinical Impact and Future Work

The integration of uncertainty estimation through MCD demonstrated a positive impact on the robustness and sensitivity of the murmur detection The application of selective classification based on the uncertainty threshold led to an increase in Recall (+3%), which implies a reduction in missed murmur detections The uncertainty analysis on the confidence score revealed significantly lower confidence scores for the unknown test samples, mirroring diagnostic uncertainty among clinical experts Future research will focus on model generalization through external validation and full edge deployment via firmware integration Explainability techniques could be combined with uncertainty quantification to provide clinicians and non-expert users with interpretable visual feedback The study acknowledges limitations, such as the reliance on the CirCor dataset and the approximate nature of Monte Carlo Dropout The overall performance remains stable and reliable regardless of the individual model variant, in line with state-of-the-art approaches

Comparison Against Existing Work

The proposed trade-off model, achieving an accuracy of 91%, is in line with the state of the art when compared against existing heart murmurs classifiers Specifically, the Light model's performance is comparable to other methods like Niizumi [34] and Guan [21], while significantly outperforming approaches like McDonald [17] in terms of parameter reduction The overall pipeline demonstrates a robust methodology for murmur detection, with performance that remains stable and reliable regardless of the individual model variant The authors conclude that an accurate detection is feasible even without relying on high-performance computers

Evaluation Metrics Summary

Model performance was assessed using standard metrics for binary classification, specifically the classification accuracy and the F1-score These metrics were computed at the location level and then aggregated to assess overall model performance on a patient-level basis Precision quantifies the proportion of correctly identified positive cases among all samples predicted as positive, while Recall (or Sensitivity) measures the proportion of correctly identified positive cases among all actual positives The final results show that for the Light model, patient-level classification employing confident segments achieved an F1-score of 81.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can be deployed on resource-constrained edge devices by using lightweight models like Light, which requires two orders of magnitude fewer parameters than deeper networks while maintaining accuracy comparable to heavier models.

  2. The system can achieve improved sensitivity by incorporating uncertainty quantification and selective classification, as this approach further improved sensitivity by 3%, enhancing robustness and clinical reliability.

  3. The AI system can provide users with a confidence score using the overall confidence score CS, which is defined as CS = α ∗ (1 − E) + (1 − α) ∗ C to indicate greater model confidence and reliability.

  4. Selective classification enables the system to flag or discard uncertain samples, leading to safer and more robust decision-making by differentiating between known and unknown recordings.

  5. The final optimized AI system can be deployed on microcontrollers like the F401RE with a latency of 230ms, which is feasible given the resource constraints analyzed through tools like ST Edge AI Developer Cloud.

Abstract

Early and reliable detection of heart murmurs is essential for the timely diagnosis of cardiovascular diseases, yet traditional auscultation remains subjective and dependent on expert interpretation. This work investigates an artificial intelligence (AI)-based approach for heart murmur detection using the CirCor Heart Sound dataset, with a particular focus on deployment in low-cost and resource-constrained embedded platforms. A lightweight convolutional neural network (CNN) is developed using spectrogram features extracted from heart sound recordings. The complete processing pipeline, from feature extraction to final prediction, is implemented and validated on an STM32F401 microcontroller (MCU). To enable real-time execution within the limited hardware resources of the target platform, memory- and latency-oriented optimizations together with fixed-point processing are employed. Experimental results demonstrate that accurate murmur detection can be achieved despite the severe computational constraints of the embedded device. The proposed system achieves a patient-level accuracy of 92.5% and an F1-score above 82.5%, while using far fewer parameters than conventional high-resource deep-learning approaches. The embedded implementation exhibits only marginal performance degradation, with performance drops below 2 percentage points compared to the original full-precision implementation. These findings demonstrate the feasibility of real-time AI-based heart murmur screening on low-cost microcontrollers. By reducing hardware requirements while preserving predictive performance, the proposed approach can facilitate the development of portable, affordable, and widely deployable diagnostic tools for remote and resource-limited healthcare settings.

Related papers