Accepted Publications

Abstract:
Introduction: The electrocardiogram (ECG) still acquires an important role in the diagnosis of heart diseases to this day. However, most patterns of diseases were based on old data sets and outdated stepwise algorithms that proved to have limited accuracy. To further benefit from the ECGs of old cohorts that are difficult to bring back again for a repeat ECG, we needed to transform the ECG signal, either scanned or printed, to the raw digital (time, millivolts) form that permit us to use them in machine learning algorithms aimed at improving diagnostic accuracy of the ECG. Objectives: Our aim is to develop an application that can digitize the printed or scanned format of the ECG. Methods: 50 random ECG scanned or paper-based images are utilized in our study. An image processing method is implemented to identify the ECG signal by detecting the regions of interest and extracting the ECG signal. Then serial steps will follow to digitize and verify the results. Results: The digitized ECG signals were initially visually validated to ensure the correctness of the image processing algorithm. Also, a comparison against parameters such as heart rate and QRS interval, from the scanned images was performed with the accuracy being more than 90%.
Conclusion: Digitized ECG signals from previously stored paper or scanned ECGs can be obtained with high accuracy and precision. This can permit integrating the digitized ECG signal with clinical data and cardiac imaging data in machine learning algorithms to aid in the diagnostic and prognostic evaluation of patients with cardiovascular disease.

Abstract:
Introduction: The gold standard for the diagnosis of ischemic cardiomyopathy is coronary angiography. However, risk score models are used to stratify patients who are most likely to benefit from the procedure. These score models have assumed traditional regression analyses to associate between risk factors and cardiovascular outcomes. Data mining and machine learning allows for improved accuracy in predicting the outcome. Objectives: Our aim is to implement machine learning algorithms to interpret clinical, ECG, and echocardiography data to determine need for coronary angiography. Methods: We identified a cohort of 204 consecutive patients with reduced ejection fraction (EF<50% on echocardiography) and who received a subsequent coronary angiogram and ECG within 90 days between July 1st 2013 and December 1st 2015. Several algorithms were tested to obtain the best subset of features along with the corresponding ranking of the importance of each clinical, ECG and echocardiography (including speckle tracking) variables. These included Decision Tree Classification, ensemble learning techniques (boosting and bagging), and Deep Learning Neural Networks. Results: The resulting algorithms contained between 9-17 variables. The post-systolic shortening (PSS) and PSS deceleration time, especially of the basal anterior and infero-septal segments respectively, and Right bundle branch block were highly ranked variables in many of the algorithms. The first ranked feature in all algorithms was regional wall motion (RWM) and had comparable AUC across all approaches. However, the accuracy results of the downstream variables diverged within each algorithm. AUC was lowest for the decision tree (0.72) and the Deep Learning NN (0.767) and highest for the Logitboost (0.839) when RWM was excluded. As such, the power of the algorithms to sway the final result from an ischemic to non-ischemic output or vice versa was shown to be the greatest with Deep learning ANN algorithm.
Conclusion: Machine learning algorithms, particularly the AdaboostM1 and GentleBoost models, are easy algorithms to implement and interpret with high accuracy. They outperform the traditional regression analyses, in predicting ischemic cardiomyopathy from clinical, ECG and echocardiographic data. These algorithms may be integrated as tools to determine need for coronary angiography to diagnose this disease.