Table 1.
Descriptions of ML tasks and examples of associated clinical use cases
| Machine learning tasks | Description | Common types of models* | Clinical example |
|---|---|---|---|
| Classification | Classify data into pre-specified groups | Logistic regression, support vector machines, random forests, neural networks | Predict critical care usage and hospitalization in children presenting to emergency departments [20] |
| Regression | Predict a continuous variable outcome | Linear regression, random forests, neural networks | Predicting fracture healing time in children [21] |
| Clustering | Completed without labeled data in order to categorize patients into either known groupings or novel subgroups | K-means clustering, Gaussian mixture models | Discovering phenotypes of pediatric asthma on the basis of asthma control [22] |
| Outlier detection | Unsupervised approaches to detecting deviations from distributions in data | One-class support vector machines, Bayesian networks, hidden Markov models, isolation forests | Detecting copy number variations in patients with Tetralogy of Fallot [23] |
| Time-series models | Modeling longitudinal data sampled over multiple time points | Recurrent neural networks, long short-term memory networks | Prediction of cardiac arrest from continuous physiological signals over time in pediatric intensive care unit patients [24] |
| Reinforcement learning | Modeling the optimal action to take in order to maximize a reward in response to environmental changes | Markov decision processes, Q-learning | Determining optimal and individualized treatment suggestions for septic patients in intensive care units [25] |
| Image-based models | Models involving medical image data (i.e., X-rays, MRIs, and CT) | Convolutional neural networks, generative adversarial networks | Diagnosing and providing treatment suggestions for congenital cataracts [26] |
| Natural language processing | Processing and analyzing language-based data (i.e., electronic medical records) | Latent Dirichlet allocation, long short-term memory networks, transformer models | Extracting text from an EMR to predict a patient’s diagnosis, assisting in triage or aiding in complex cases [27] |
*Does not include all types of models for each category as new modeling techniques are rapidly evolving