Phonocardiogram Classification:

Leveraging 1D Inception Time CNN

With Explainable AI using Grad-CAM

Antony M.

Webpage: antony-gitau.github.io

​

Bjørn-Jostein Singstad¹, Antony M. Gitau²

¹Akershus University Hospital, Vestfold Hospital Trust, University of Oslo, Norway

²Kenyatta University, Nairobi, Kenya

​

TrustAI Workshop 2023

8 September

Accra, Ghana

​

Why Classify Heart Sounds?

​

• Many heart diseases cause changes in heart sounds

​

• Heart sound signal has much more information than can be assessed by the human ear or by visual inspection of the signal tracings on paper

Heart sounds, as a diagnostic tool

• 1 in 5 affected children do not survive to discharge.

• 18 and 25% will succumb in the first year of life.
• By adolescence, ~4% will succumb to congenital heart disorders.

The Scope of the methodology

• Data for Training, Validating & Testing.

• Two 1-Dimensional Convolutional Neural Networks (2 1D CNNs).

• Post hoc explanation of the two models.

Data: Training, Validating & Testing

​

Collected from a pediatric population during two mass screening campaigns conducted in Northeast Brazil in July-August 2014 and June-July 2015.

Data: auscultation positions of the chest walls

​

​

J. Oliveira et al., 2022

Data: heart sounds from one or more auscultation positions of the chest walls

​

Data: number and location of the recordings varied between patients.

​

Data: number of absent murmur labels was the highest in both females and males

​

Identifying & classifying murmurs and clinical outcome in heart signals

Identifying & classifying murmurs and clinical outcome in heart signals

​

Model performance based on PhysioNet 2022 Challenge Scoring

PhysioNet Challenge Scoring: cost metric

3-step post hoc explanations of the models

An example of a Grad-CAM heatmap plot for the murmur model

An example of a Grad-CAM heatmap plot for the clinical outcome model

Ongoing work

• Working with health specialists to better understand the explanations.

​

• Exploring more explainability tests and techniques.

​

• Improving the models' performance.

​

Conclusion

• Trained two 1D CNN models
• Heart murmur classification: present,

absent, or unknown.

• Clinical outcome classification: abnormal or normal.
• Submitted both models to PhysioNet

Challenge.

• Explainability using Grad-CAM

Phonocardiogram Classification:

Leveraging 1D Inception Time CNN

With Explainable AI using Grad-CAM

Antony M.

Webpage: antony-gitau.github.io

​

Bjørn-Jostein Singstad¹, Antony M. Gitau²

¹Akershus University Hospital, Vestfold Hospital Trust, University of Oslo, Norway

²Kenyatta University, Nairobi, Kenya

​

TrustAI Workshop 2023

8 September

Accra, Ghana

Thank you!

​

​