Brief Review — Utilization of Deep Learning and Expert Feature Classifier for Detection of Heart Murmurs
ISIBrno-AIMT, Ranks of 9th and 12th for Murmur and Outcome Classification Tasks
Utilization of Deep Learning and Expert Feature Classifier for Detection of Heart Murmurs,
ISIBrno-AIMT, by Institute of Scientific Instruments of the Czech Academy of Sciences, Brno, 2022 CinC (Sik-Ho Tsang @ Medium)Phonocardiogram (PCG) / Heart Sound Classification
2016 … 2024 [MWRS-BFSC + CNN2D] [ML & DL Model Study on HSS] [Audio Data Analysis Tool]
Summary: My Healthcare and Medical Related Paper Readings and Tutorials
==== My Other Paper Readings Are Also Over Here ====
- A deep learning model is proposed that classifies whether there is a heart murmur.
- Furthermore, there is an expert feature classifier proposed to assess whether the patient’s status is normal or abnormal.
Outline
- Murmur Classification
- Outcome Classification
- Results
1. Murmur Classification
- The heart murmur detector is based on a deep-learning method that processes phonocardiogram data (fs = 4 kHz) converted into spectrograms by the Short Time Fourier Transform (STFT).
- Each row of the spectrogram was normalized using a z-score.
The neural network architecture consists of 5 convolutional layers with batch normalization, ReLU activation, and dropout.
- The first convolutional layer has kernel size dependent on the spectrogram height and temporal dimension of 1.
- Subsequent convolutional layers use kernel 1×N, where N is temporal domain kernel size 3, 5, 7, and 9, respectively.
- Subsequently, the gated recurrent unit layer (GRU) is utilized to process the temporal domain of the data. The output from the GRU layer is split into two output heads.
The first output head provides heart sound segmentation, and the second output head provides information on whether there is a murmur present in the recording or not. It is observed that the GRU layer was improving accuracy for the segmentation task.
For the training purpose, the model is trained using a combination of two loss functions, i.e., cross-entropy loss for murmur classification and second segmentation crossentropy loss for heart sound segmentation.
- The model is trained for 15 times, each with a different subset of training data. Then, the top 5 models are selected to form a classification ensemble (Figure 2).
2. Outcome Classification
- An outcome classification is based on the extraction of expert features and subsequent classification using a shallow neural network with two fully connected layers (82 and 128 neurons).
- The features were extracted from each auscultation point, and the features from possible missing auscultation points were filled with zeros.
- Two lists of features, the patient metadata and phonocardiographic recordings, are collected, respectively.
1. List of extracted features (patient metadata): Age, sex, pregnancy status, height, weight, number of auscultation points measured.
2. List of extracted features (from each auscultation location) time-domain mean, std, skewness, kurtosis, length in seconds, power, power of bandpass envelopes in bands (15–90Hz, 55–150Hz, 100–250Hz, 200–450Hz, 400-800Hz), and correlation coefficients of former bandpass envelopes. The band envelopes were calculated as the absolute value of the Hilbert transformed bandpass signal. The bandpass signals were computed with Butterworth 3rd order acausal filter.
- The shallow neural network was trained for 30 epochs. The model with the lowest validation score was saved for the subsequent inference phase. This approach was repeated five times to form an ensemble model.
3. Results
- The performance of the models on the first task is assessed based on a weighted accuracy score.
- The second task is assessed according to the custom challenge metrics representing the cost of treatment.
Methods proposed for this challenge evaluated on a hidden validation set achieved the murmur score of 0.709 while requiring two hours of model training. The feature extraction approach for outcome classification achieved an outcome score of 9657.
- Yet, the local cross-validation results for murmur score are approaching 0.80, and outcome score is approaching 8000. This indicates that the proposed methods are prone to overtraining.
It also obtains murmur score of 0.755 on the hidden test set. The outcome classification method based on expert feature extraction achieved the outcome score of 12313 on the hidden test set.
