Table 1.

Glossary of key machine learning terminology.

Terminology	Definition
Artificial neural networks:	Deep machine learning inspired by the biological neural network of an animal brain and Hebbian learning (1).
Black box:	A short-term ethical challenge in machine learning where the process by which the computer reaches an outcome is not easily interpretable and is hidden from consumers and engineers alike (2).
Decision tree learning:	A supervised machine that visually resembles a tree with nodes, branches, and leaves. Trees are adept at identifying clusters of homogenous variables and predicting outcomes. Most commonly a classification and regression tree (3).
Deep learning:	Computers that utilize representation learning or hidden layers to characterize unlabeled input variables without much manual human engineering. Commonly used for natural language processing, self-driving automobiles, pharmaceutical drug research, among others (1).
Distributional shift:	A short-term ethical challenge in machine learning where the training dataset poorly represents the true test set, secondary to racial or socioeconomic biases, or outdated information (4).
Feature values:	Individual characteristics or variables that are associated with the outcome of interest. Feature engineering can either be manually conducted or automated (5).
Hebbian theory:	Based on neuropsychology work by Dr. Donald O. Hebb from his book, The Organization of Behavior. Dr. Hebb's work on neuronal plasticity contributed greatly to the initial architecture of artificial neurons and networks (6).
Insensitivity to impact:	An ethical challenge in machine learning where the algorithm is unaware of the consequences of a false-positive or false-negative test (4).
Linear classification:	A task that involves predicting categorical outcomes (i.e., type of fruit or species of animal).
Linear regression	A task that involves predicting discrete or numeric outcomes that are integers or serial numbers (i.e., patient reported outcome scores).
Machine learning:	The study of using algorithms and mathematics to predict outcomes or accomplish tasks with little instruction or explicit programming. A subset of artificial intelligence (7).
Reward hacking:	A long-term ethical challenge of machine learning where algorithms self-learn how to maximize favorable outcomes but do so by circumventing rules or cheating the system (4).
Supervised learning:	Learner attempts to describe the input-output relationship based on input variables that are labeled and have a grounded truth (5).
Support vector machine:	A machine learning modality that can either solve classification tasks by creating a maximum margin hyperplane between two outcomes, or regression tasks by plotting a best-fit plane. Involves significant human engineering through kernel functions to transform data into higher dimensions (8).
Unsupervised learning:	Learner attempts to describe the input-output relationship based on input variables that are unlabeled. Typically associated with deep learning (9).