Table 2.
Algorithm | Description | Advantages and Disadvantages | Suitable Data in Cardiac Electrophysiology |
---|---|---|---|
Deep learning | Mimics human neuronal structure with many processing layers | Represents state-of-the-art performance with raw input data in complex tasks and does not require any feature engineering but often requires extremely large datasets, intensive computational power, significant processing time, and algorithms are difficult to interpret | Raw 12-lead ECG, imaging (CT, MRI, echocardiography), clinical text, or other diagnostic and monitoring data (telemetry, Holter/wearables, electrograms, electrophysiological studies) |
Traditional supervised ML algorithms | |||
Logistic regression | Linear combinations of log-odds | Transparent and fast but performs poorly with large numbers of variables and does not automatically capture interactions | Well-selected features obtained from the electronic medical record or manually engineered from processed data (eg, features extracted from ECG signal processing, specific measurements made on images, etc) |
Support vector machine | Identifies hyperplane that separates classes, can use a linear or nonlinear kernel | Fast can be flexible with kernel adjustment but is difficult to interpret and features may need normalization and scaling | |
Naïve Bayes | Bayes’ theorem of conditional probabilities | Fast, scalable, does not require large amounts of data, but is not directly interpretable and assumes independence between variables | |
Random forest | Large ensembles of decision trees | Resistant to noise and overfitting even with large amounts of variables, flexible with continuous and categorical variables, capable of capturing variable interactions | High quantities of features obtained from the electronic medical record or manually engineered from processed diagnostic data (ECG, imaging, etc) |
CT indicates computed tomography; ML, machine learning; and MRI, magnetic resonance imaging.