Table 1 Common AI algorithm types

Regression Algorithms

Centred around basic statistical principles and adopted early within ML frameworks. Usually used to predict one variable based on the known value of other variables. Used often in areas that require numerical estimation such as forecasting and trend analysis.

Decision Tree Algorithms

Represents a decision pathway based on comparison to a known set of pre-determined ‘rules’. Application of each rule occurs at a ‘node’. Depending on the determined answer, the decision is progressed to the next linked node, and another answer is determined. The nodes therefore ‘branch’ with sequential layers of assessment from a starting point – hence the name ‘decision tree’. For example, ‘does the feature of interest have A or B characteristics? If B, does the feature then have C or D characteristics? If C, does the feature then have E characteristics?’, and so on. Complexity increases as the number of interconnected decision-making steps increases.

Clustering Algorithms

Is an approach based around grouping features of interest into relatively homogenous classes. This is done based on recognized element similarities. Are often used for preliminary data analysis and the isolation of like subpopulations. These smaller cohort fragments can then be separately explored for identifiable within-group commonalities.

Instance-Based Algorithms

Does not require ‘training’, per se, rather stores a series of exemplars in memory and compares new instances to these with the goal of establishing a ‘best match’ based on similarity. Each new case is analysed independently and can often be time consuming. Often work best in instances whereby the target function is complex but can be simplified into less complex generalisations.

Association Rule Learning Algorithms

Is a common means for initial data mining of relatively ‘raw’ datasets. It involves analysis of specific attributes looking for repetitive dependencies (i.e. What precursor features or elements are consistently associated with an observed outcome?) Often used to determine cause and effect relationships between critical events captured within the dataset. When linked with Bayesian theorem, event or outcome probability prediction can often be achieved with high reliability.

Ensemble Algorithms

Is an umbrella term to describe the practice of using multiple independently trained ‘weaker learning algorithms’ and merging the combined analysis output. Highly susceptible to inaccuracies within the individual algorithms combined (much akin to the integrity of a systematic literature review being undermined by included poor quality RCTs). Performed well however, is regarded as one of the most effective and ‘powerful’ algorithmic styles.

Artificial Neural Network (ANN) Algorithms

As discussed previously, ANNs are interconnected iterative sequences based conceptually upon human (biological) neural networks. They are commonly used for regression and classification. Acknowledged to be an extremely complex analytical subfield, consisting of many variations and algorithms for specific problems. Usually time-consuming to establish and require high computer processing capacity. A rapidly growing field both within and outside medicine.

Deep Learning (DL) Algorithms

Is currently the newest form of neural networks employed in healthcare. Use large modelling domains with a complex and hierarchical structure usually composed of many interconnected, nonlinear ‘layers’. Have been applied with great effect in areas such as image and feature recognition (see ‘face recognition’ technologies in the lay world and ‘diagnosis’ from digital imaging within medicine). Deep convolutional neural networks represent an evolved DL platform centred around the established mathematical principle of convolution which considers the fluid impact of one variable interacting with another to generate a third, separate function.