Table 1.
Relevant Machine Learning Terms
| Term | Definition |
|---|---|
| Artificial intelligence (AI) | A computer’s ability to learn from experience |
| Machine learning (ML) | A type of artificial intelligence in which computers draw conclusions from data without being directly programmed |
| Supervised learning | Models in which the outcome is known for each observation |
| Unsupervised learning | Models in which the outcome is not known for each observation |
| Semisupervised learning | Models in which the outcome is known for some observations but not others |
| Label | The patient outcome (dependent variable) |
| Feature | An attribute/characteristic of the patient (dependent variable) |
| Sensitivity | Ability of a model to correctly identify true cases |
| Specificity | Ability of a model to directly identify negative cases |
| Accuracy | Measure of correctly labeled data instances over the total number of instances |
| Precision | Fraction of relevant instances among the retrieved instances (ie, positive predictive value) |
| Recall | Fraction of relevant instances that were retrieved correctly (ie, sensitivity) |
| Training data set | Data used to develop a model |
| Validation data set | Data used to test a model’s performance while training |
| Test data set | Data used to test the accuracy, precision, or recall against real-world data |
| Out-of-sample data | In a study cohort, the data not used as training data |
| Bias-variance tradeoff | In supervised learning, overfitting and underfitting can result in loss of performance |
| Bias | Difference between the average prediction of a model and the correct value |
| Variance | Variability of a model prediction for a given data point |
| Overfitting | When the model follows noise, resulting in low bias and high variance |
| Noise | Nonpredictive features in the data set |
| Underfitting | When the model fails to capture the underlying patterns in the data, resulting in low variance and high bias |
| Decision tree | A model that separates data into smaller and smaller partitions until each observation is classified according to the outcome of interest |
| Stopping criteria | Criteria used to stop further partitioning of data in a decision tree. Can prevent overfitting |
| Ensemble model | An ML technique combining multiple individual models |
| Random forest | A type of ensemble model that combines decision trees to produce a probabilistic prediction for the outcome |
| Receiver operator characteristic (ROC) curve | A way to graph the sensitivity and specificity (or precision) of a model |
| Area under the curve (AUC) | A technique to compare model results (with other models or other measurement tools) by calculating the area under an ROC curve |
| Natural language processing (NLP) | A type of AI in which the algorithm learns how to ‘understand’ language, including contextual nuances |