Source-linked AI summary
Phonocardiographic Sensing using Deep Learning for Abnormal Heartbeat Detection
Siddique Latif, Muhammad Usman, Rajib Rana, Junaid Qadir
TL;DR
Automatic cardiac auscultation must achieve reliable, accurate abnormal-heartbeat detection despite practitioner burden and noisy recordings. The paper evaluates RNN-based PCG classification, finding significant improvements over conventional and non-recurrent deep-learning models, with potential for real-time remote monitoring.
Problem
Manual auscultation requires extensive expertise, has limited accuracy among less-trained practitioners, and heart recordings contain background noise.
Method
The study evaluates various RNN architectures for automated abnormal-heartbeat classification using segmented PCG signals and extracted features.
Results
RNNs significantly outperform conventional classifiers and non-recurrent deep-learning models, including the Rank 1 AdaBoost-CNN model from the 2016 PhysioNet challenge.
Takeaways & Limitations
RNNs provide promising PCG heartbeat classification and may support real-time abnormal-heartbeat detection for remote monitoring.
Abstract
from arXiv · showhide
Cardiac auscultation involves expert interpretation of abnormalities in heart sounds using stethoscope. Deep learning based cardiac auscultation is of significant interest to the healthcare community as it can help reducing the burden of manual auscultation with automated detection of abnormal heartbeats. However, the problem of automatic cardiac auscultation is complicated due to the requirement of reliability and high accuracy, and due to the presence of background noise in the heartbeat sound. In this work, we propose a Recurrent Neural Networks (RNNs) based automated cardiac auscultation solution. Our choice of RNNs is motivated by the great success of deep learning in medical applications and by the observation that RNNs represent the deep learning configuration most suitable for dealing with sequential or temporal data even in the presence of noise. We explore the use of various RNN models, and demonstrate that these models deliver the abnormal heartbeat classification score with significant improvement. Our proposed approach using RNNs can be potentially be used for real-time abnormal heartbeat detection in the Internet of Medical Things for remote monitoring applications.
I. INTRODUCTION
Cardiovascular disease diagnosis relies on auscultation, but accuracy varies with practitioner expertise and echocardiography is costly. PCG enables sensor-based automated detection, motivating RNNs for noisy, sequential heart-sound classification.
- CVDs are a major global health problem, making earlier diagnosis important for reducing potential risk factors.
- Auscultation accuracy ranges from 20–40% for medical students and primary-care physicians to roughly 80% for expert cardiologists.
- Echocardiography visualizes cardiac activity but costs an average of $1500, leaving a need for reliable, cost-effective early diagnosis.
- PCG records heart sounds with a digital stethoscope for automated normal-versus-abnormal assessment and potential real-time mobile or cloud diagnosis.
- Because heart sounds are temporal signals affected by symptoms and noise, the study evaluates RNNs on four-location PCG recordings from the 2016 PhysioNet challenge dataset.
- Prior PCG studies used CNNs, fully connected networks, or AdaBoost-CNN ensembles, while LSTM and GRU architectures had not been applied to PCG analysis.
III. PROPOSED APPROACH AND RNN MODELS
The proposed pipeline preprocesses and segments heart sounds, extracts features, and feeds sequential representations into RNNs for normal-versus-abnormal classification.
- The proposed approach detects S1 and S2, segments heart sounds into smaller chunks, extracts features, and classifies them with RNNs.
- RNNs process sequences by updating a hidden state from the current input and previous hidden state, then projecting outputs to class probabilities.
- The RNN formulation uses weight matrices, a bias vector, and a hidden-layer function to compute the hidden representation.
1) Long Short-Term Memory (LSTM) Units:
LSTM units extend RNNs with memory and gates that regulate information entering, leaving, and remaining in each recurrent memory block.
- An LSTM memory block stores representations over extended intervals using input, output, and forget gates.
- Each LSTM unit maintains a memory c_t for a heart-signal sequence at time t.
- The output gate modulates memory content, while the forget gate updates memory by discarding existing information and incorporating incoming information.
- The input gate controls the extent of incoming information entering the recurrent memory.
- The three gates jointly control the final update of the existing memory.
2) Gated Recurrent Units (GRUs):
GRU is a simplified recurrent architecture that combines LSTM’s input and forget gates into an update gate while adding a reset gate.
- GRU combines the LSTM input and forget gates into one update gate and adds a reset gate.
- Unlike LSTM, GRU does not use a forget gate to control the amount of incoming memory.
B. Bidirectional Recurrent Neural Networks
The paper uses bidirectional recurrent models to incorporate context from both directions when processing heart-sound sequences. It evaluates both standard and bidirectional RNNs for abnormal heartbeat detection.
- Bidirectional RNNs process information forward and backward, allowing them to exploit future context unavailable to standard RNNs.The hidden sequence is computed in both directions before updating the output layer.
- Bidirectional GRUs produce forward and backward hidden-state sequences that are concatenated at the output.
- The study compares standard RNNs and bidirectional RNNs for abnormal heartbeat detection.
IV. EXPERIMENTAL PROCEDURE
The experimental procedure evaluates different RNN models on publicly available datasets, with dataset details and procedures presented in the section.
- Different RNN models are evaluated on publicly available datasets.
- The section presents the details of the datasets used for evaluation.
- The section presents the experimental procedure used to evaluate the methodology.
A. Database Description
The experiments use PhysioNet Challenge 2016 heart-sound recordings and Logistic Regression-HSMM preprocessing to identify heart-cycle states before classification. The preprocessing combines multiple envelope features and detects four states in the cardiac cycle.
- Database Description: The PhysioNet Challenge 2016 dataset contains 3240 raw heart-sound recordings collected in six databases.The recordings include normal and abnormal heart sounds from subjects with varied ages and heart conditions.
- Preprocessing: Segmentation identifies S1 and S2 locations so incoming heartbeats can be aligned before classification.Alignment significantly improves recognition scores according to the cited preprocessing description.
- Preprocessing: Logistic Regression-HSMM uses emission probabilities to identify heart states and improves discrimination compared with Gaussian- or Gamma-distribution approaches.
- Preprocessing: The preprocessing combines homomorphic, Hilbert, wavelet, and power spectral density envelope features to detect probable S1 and S2 states.
- Preprocessing: Figure 3 shows S1, S2, systole, and diastole as the four detected heart-cycle states using Logistic Regression-HSMM.The accompanying description identifies the blue line as the heart signal and the black line as the detected states.
C. Segment Extraction
Because the dataset contains only 3240 recordings, the study segments complete PCG waveforms into shorter sequences of heart cycles for RNN training and validation.
- Segment Extraction: The overall PCG waveform is segmented because 3240 recordings are considered inadequate for evaluating RNNs robustly.
- Segment Extraction: The study extracts sequences containing 2, 5, or 8 heart cycles from the recordings.
- Segment Extraction: Segment extraction divides overall heart sounds into smaller chunks for model training and validation.The passage notes that prior studies also used shorter segments, including 3-second segments for CNN training and validation.
D. Feature Selection
The study represents PCG signals with compact MFCC features and evaluates gated recurrent architectures, finding that two gated layers provide the best LSTM and GRU performance while larger models offer no improvement.
- Feature representation: MFCCs represent PCG signals compactly using 25 ms windows, 10 ms steps, and the first 13 coefficients.The authors select a compact feature space because larger feature spaces do not always improve performance.
- Model architecture: Two gated layers produced the best classification results for both LSTM and GRU models.One-layer models performed poorly, while deeper gated and dense configurations failed to improve performance, possibly because of overfitting.
- Model architecture: The LSTM and BLSTM models use two LSTM layers with tanh activation, followed by dense and softmax classification layers.For each heartbeat, recurrent-layer outputs are passed to a dense layer and then to softmax.
V. RESULTS AND DISCUSSION
The experiments use segmented heart-sound recordings to test cycle-length choices and compare RNNs with conventional and prior deep-learning models. Five heart cycles were selected for reporting because all RNN models performed consistently well at that length, and RNNs outperformed the tested baselines and prior deep models.
- Experimental setup: Five heart cycles were selected for reporting because all RNN models consistently performed well at that sequence length.The dataset contains 3,240 recordings, split into 75% training, 15% validation, and 10% testing data.
- Experimental setup: Abnormal heart cycles have longer duration than normal heart cycles in the extracted five-cycle segments.
- Baseline comparison: RNNs significantly outperformed Logistic Regression, Support Vector Machines, and Random Forest in every reported performance measure.The baseline models were trained on MFCCs computed from five heart cycles.
- Comparison with prior work: RNNs outperformed all deep-learning models used for heart-sound classification, including the Rank 1 AdaBoost-CNN approach.The authors report better recognition rates for both LSTM and GRU models across every performance measure.
C. Performance Comparison of RNNs
The study compares recurrent architectures for PCG heartbeat classification and finds broadly comparable performance across RNN variants. BLSTM performs consistently well, while GRU remains promising despite its simpler architecture.
- Performance comparison: Different RNN architectures achieved comparable performance on PCG heartbeat classification.
- Performance comparison: BLSTM performed consistently well and was preferred for heartbeat classification using PCG.
- Performance comparison: GRU showed promising performance on PCG data despite having a simpler architecture than LSTM.
- Comparison with prior models: The study reports that RNNs significantly outperform conventional deep-learning models and the Rank 1 AdaBoost-CNN model.The authors associate this advantage with RNNs' ability to capture temporal statistics and dynamics in heartbeat sequences.