Main figure captions
Figure 1. Hierarchy of disciplines of artificial intelligence.The discipline of AI is often categorized as part of computer science,
although it also builds upon other fields, such as mathematics and
cognitive sciences. ML is a subdiscipline of AI, whereas deep learning
is a further specialization within ML, generally characterized by
large-scale artificial neural networks consisting of many layers, hence
the term deep. The right part of the figure displays typical
terms that one encounters within the (sub)discipline.
Figure 2. A conceptual framework for AI applications in the
biomedical domain. The framework is structured by learning strategy,
learning goal and data modality. The included studies are selected as
illustrations of how AI is used within the medical field and how its
applications can be conceptually categorized. They are not necessarily
selected based on the inclusion criteria stated in the introduction.
References for the shown studies are included in the Supplementary
References.
Figure 3. Difference between ordinary least squares (OLS),
machine learning (ML), and deep learning (DL) methodology. (a) In OLS,
features (or predictors) are modeled manually, and their relationship is
assumed linear to the output variable unless specified differently.
Interpretation of the model and learned patterns (inference) is
straightforward. (b) A similar procedure is followed with ML, but the
algorithm can learn more complex patterns from the provided features.
Nevertheless, thorough feature engineering by the practitioner is a
critical step for delivering a performant model. (c) With DL, especially
when applied to unstructured data, feature engineering is an inherent
behavior of the interconnected neural network layers. The relationship
between input features (tabular data fields, image pixels, text
snippets, etc.) to the predicted output is more opaque and harder to
interpret.
Figure 4. Workflow of developing a machine learning model to
predict disease risk. The best practice in machine learning modeling is
using distinct training, validation (or tuning), and test datasets. The
modeling steps till testing are generally executed in sequential order,
where it is common to iterate multiple times based on validation results
that inform model improvements. It is discouraged to assess and improve
test performance iteratively, as this can lead to overfitting. The steps
preceding model development are excluded, primarily consisting of
problem definition and study design, data collection, and preprocessing.
Figure 5. Application domains of artificial intelligence within
the allergy field. Domains identified as currently most active and
discussed in the main text. Other areas, such as clinical trial
optimization, have been excluded due to the limited number of impactful
applications.