Brief Review — Phonocardiogram Classification Using 1-Dimensional Inception Time Convolutional Neural Networks
InceptionTime on PCG
Phonocardiogram Classification Using 1-Dimensional Inception Time Convolutional Neural Networks
InceptionTime on PCG, by Simula Research Laboratory, Kenyatta University, Medsensio AS, The Arctic University of Norway, and Akershus University Hospital
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 ====
- InceptionTime is proposed to be used on PCG data for George B. Moody PhysioNet Challenge 2022.
- Two 1-dimensional convolutional neural networks (CNN) are developed on a PCG data set from a pediatric population of 1568 individuals.
- One model predicted murmurs, while the other model predicted clinical outcomes.
- (Authors’ GitHub, https://github.com/Bsingstad/Heart-murmur-detection-2022-Simulab: I have played around the GitHub and it is still workable.)
Outline
- InceptionTime on PCG
- Results
1. InceptionTime on PCG
1.1. Data
The data set used in this work consists of 5272 PCGs from a pediatric population of 1568 individuals (CirCor).
- 3163 PCGs from 942 individuals were used for training. The remaining 2109 PCGs from 149 and 477 patients, which are private.
1.2. Preprocessing
The signal is of 4000Hz and it is downsampled to 100Hz. Signal length has 6451 samples, if it is shorter, zero-padded.
1.3. Label Processing
The data set was relabeled from patient-wise labeling to recording-wise labeling, as above.
1.4. 1D CNN With Inception Time Models
Both models were 1 dimensional CNNs with an Inception Time architecture [13].
- The murmur classifier was trained using weighted categorical cross entropy, while the clinical outcome classifier was trained using weighted binary cross entropy.
- The murmur classifier was trained for 30 epochs while the clinical outcome classifier was trained for 20 epochs.
- Both models were trained using a batch size of 20.
1.5. Post-Processing
The recording-wise predictions from the model were finally converted back to patient-wise predictions. The murmur conversion is shown in Algorithm 2 and the clinical outcome conversion is shown in Algorithm 3, as above.
2. Results
2.1. Performance
- The ranking on the murmur task was determined by the weighted accuracy score. The clinical outcome task on the other hand, used clinical cost score.
The authors’ team, Simulab, trained a clinical outcome classifier that achieved a challenge cost score of 12419 (ranked 14th out of 39 teams) and the murmur classifier achieved a weighted accuracy of 0.593 (ranked 30th out of 40 teams) on the test set.
2.2. Discussions
- Since each team got 10 attempts to test their algorithm on the validation set, there can be indirect overfitting.
- Pre-training models were also tested using the 2016 PhysioNet Challenge data set [14, 15]. Different approaches on training the pre-trained models were explored. However, there were no significant improvements during CV on the public training set and the performance on the validation set was actually lower compared to no pretraining.
- Both murmur and the clinical outcome classifiers were trained using single PCG recordings, and the auscultation location was not taken into consideration.
