| Artificial intelligence |
Machine-based data processing to achieve objectives that typically require human intelligence |
| Machine learning |
Subdiscipline of artificial intelligence, referring to the algorithms and statistical models used to learn how to achieve objectives just from data, without using much knowledge of the underlying domain that is learned |
| Supervised machine learning |
Uses data as input and can learn to predict a desired output. The aim is for models to ‘generalise’, that is, they can learn from (training) data so that the system can make correct predictions on unseen data. This is evaluated by using a separate test dataset. If the predicted output is categorical in nature (e.g. recognising a named disorder from ECG traces), then the problem is called classification. If the predicted output is numerical in nature (e.g. predicting potassium levels from ECG traces), then we refer to the problem as regression. Models require subsequent validation and testing using independent input data. Crucially, systems should be tested on data from different patients than the ones in the training data |
| Unsupervised machine learning |
Identification of patterns within complex data, without the specific objective of prediction. Does not require the input data to have corresponding labels nor separate training and testing data |
| Deep learning |
Artificial neural networks, algorithms inspired by the human brain, learn from large amounts of data (training datasets) to generate automated predictions from new inputs |
| Feature |
Quantifiable property of the data |
| Training dataset |
The large dataset of values for the machine learning model to learn from (model building) |
| Test dataset |
Data that have not been seen by the model during the training process, which are used to make sure that during training the model has learned useful principles that work on cases beyond the training set, rather than simply learn to recognise particular individuals within the training set |
