Page 1Coronary artery disease [606026]

Page 1Coronary artery disease
progression prediction
using neural networksIntern: Stoian Andreea Cristina
Mentor: Puiu AndreiStoian Andreea Cristina, September 2020

Page 2Introduction
Unrestricted © Siemens Medical Solutions USA, Inc., 2016•Coronary artery disease is a condition which
affects the arteries that supply the heart with
blood. It is usually caused by stenosis which is a
buildup of plaque inside the artery walls. This
buildup causes the inside of the arteries to become
narrower and slows down the flow of blood.
•Coronarography or coronary angiography
represents the main procedure to detect the
stenosis.
•Problem statement:
Having a coronarography and a set of follow-up
features, predict the progression of atherosclerosis
after x days.

Page 3Unrestricted © Siemens Medical Solutions USA, Inc., 2016Input description
•Days_diff : distance between the two coronary angiography counted in days;
•LM, LAD, LCX, RCA : Percentage of stenosis in each of the 4 sections of the
coronary arteries for the first coronary angiography;
•Dyslipidemia, diabetes, smoking, hypertension : presence of comorbidities
•Age, Gender
•Atheroma history
•Statins dose: the dose of drug treatment for this condition;
•Compliance statins, diabetes, hypertension : compliance treatment for
atherosclerosis or comorbidities

Page 4Unrestricted © Siemens Medical Solutions USA, Inc., 2016Feature Missing
valuesReplacement Feature Missing values Replacement
Days_diff 0 – Hypertension 5 Not known(-1)
LM 23 0% Age 1 Mean value
LAD 21 0% Gender 0 –
LCX 17 0% Atheroma
History1 Not known(-1)
RCA 17 0% Statins dose 11 0
Dyslipidemia 46 Not known(-1) Compliance
statins298 Not known(-1)
Diabetes 95 Not known(-1) Compliance
diabetes327 Not known(-1)
Smoking 3 Not known(-1) Compliance
hypertension41 nu se stie(-1)Data preprocessing – missing values

Page 5Unrestricted © Siemens Medical Solutions USA, Inc., 2016Class distribution

Page 6Unrestricted © Siemens Medical Solutions USA, Inc., 2016Distribution of labels correlated to days difference

Page 7Unrestricted © Siemens Medical Solutions USA, Inc., 2016Correlation matrix

Page 8Unrestricted © Siemens Medical Solutions USA, Inc., 2016Dataset splits for Stratified K-Fold cross-validation

Page 9Unrestricted © Siemens Medical Solutions USA, Inc., 2016Neural network architecture
Optimizer: Adam lr=0.0001
Epochs:50
Activation: relu, sigmoid
Batch size = 321stexperiment
Hyperparameters values
ResultsTrain Validation
Sensitivity 0.68 0.61
Specificity 0.58 0.52
Accuracy 0.61 0.54Train
Validation

Page 10Unrestricted © Siemens Medical Solutions USA, Inc., 2016Neural network architecture
Optimizer: Adam lr=0.00001
Epochs: 400
Activation: relu, sigmoid
Batch size = 322ndexperiment
Hyperparameters values
ResultsTrain Validation
Sensitivity 0.68 0.62
Specificity 0.61 0.53
Accuracy 0.63 0.56Train
Validation

Page 11Unrestricted © Siemens Medical Solutions USA, Inc., 2016Neural network architecture
Optimizer: Adam lr=0.00001
Epochs: 300
Activation: tanh, sigmoid
Batch size: 323rdexperiment
Results
Train Validation
Sensitivity 0.69 0.62
Specificity 0.57 0.56
Accuracy 0.6 0.58Hyperparameters valuesTrain
Validation

Page 12Unrestricted © Siemens Medical Solutions USA, Inc., 2016Neural network architecture
Optimizer: Adam lr=0.0001
Epochs:50
Activation: relu, sigmoid
Batch size = 324thexperiment
Hyperparameters values
ResultsTrain Validation
Sensitivity 0.65 0.6
Specificity 0.62 0.54
Accuracy 0.63 0.56Train
Validation

Page 13Unrestricted © Siemens Medical Solutions USA, Inc., 2016Conclusions
•This project shows the fact that the progression of coronary artery disease can be
predicted using neural networks
•Although the dataset was small, the accuracy was over 60%
Future directions of development
•Dynamic weighting of the loss
•Retrain the model on a larger dataset and evaluate it on a test set
•Finding more input features for the prediction of progression