From concept to practice: a scoping review of the application of AI to aphasia diagnosis and management

Figures & data

Figure 1. Screen capture of the data analysis dashboard.

A snapshot of an interactive dashboard created to visualise publication categories, AI techniques, keywords, research objectives and trends over time.

Figure 2. PRISMA flow chart used for the scoping review.

The PRISMA-Scoping Review flow chart showing the process followed to select studies for the review.

Figure 3. The evolution of AI in Aphasia research.

A diagram to show the evolution of AI in Aphasia over time with significant time periods/ research objectives and AI techniques used.

Figure 4. The evolution of AI techniques.

The trend chart to show the evolution of different AI techniques over time.

Figure 5. The AI landscape in Aphasia research.

A mindmap to show the AI landscape in Aphasia research in terms of different techniques.

Table 1. Publication counts by categorisation of research objectives.

Stage	Research objective	n	%
Assessment		40	51.9
Diagnosis		35	45.5
Speech analysis		3	3.9
Monitoring affect		1	1.3
Prognosis		1	1.3
Discovery		15	19.5
Computational models of language		6	7.8
New subtype classifications		3	3.9
Lesion analysis		2	2.6
Pathology analysis		1	1.3
Review paper		1	1.3
Analysis of language errors		1	1.3
Speech analysis		1	1.3
Therapy		13	16.9
Automated therapy software		5	6.5
Response verification and feedback		4	5.2
Predict therapy response		3	3.9
Review paper		1	1.3
Self management		7	9.1
Assistive systems		7	9.1
Discovery and assessment		2	2.6
Diagnosis, computational models of language		1	1.3
Diagnosis, lesion analysis		1	1.3

Table

Computational models of language	Computational models of language attempt to simulate aspects of the structure and organisation of language in the human brain. These models can then be explored to learn about language processing, storage and disruption. These models are used for word prediction tasks in Aphasia.
Convolutional neural network	Convolutional Neural Networks (CNNs) are able to “scan” spatial data such as images or spectrograms to identify multiple features. Using these features, they are then able to learn to categorise data through training. Image and video processing could be achieved through CNNs that can be used for gesture and facial expression analysis.
Decision tree	Decision trees classify or label data by “splitting” data by multiple features using a threshold. For example, data from an aphasia assessment might be split by high/low fluency, high/low comprehension scores, high/low repetition scores, etc. The thresholds in decision trees are adjustments using algorithms until the optimal categorisation is reached, ideally matching known aphasia subtypes. Decision trees could be used for tasks such as Aphasia severity prediction and type classification.
Deep neural network	Deep neural networks contain multiple hidden layers of nodes and these allow learning/processing of highly advanced and complex features, e.g. estimating recovery based on demographic and imaging data. These models could be used for tasks such as Aphasia severity prediction and type classification.
Dimensionality	Dimensionality refers to the number of features or variables in a dataset. While a higher number of dimensions increases the information available to classify the data with, when training machine learning, the number of examples required to adequately train a model increases exponentially with each additional dimension.
Fuzzy rule systems	Traditional computer/mathematical logic uses binary decision making (e.g. aphasia: true/false) based on fixed thresholds. In contrast, fuzzy rule systems allow sorting of information using multiple variables that may interact with one another.
Genetic programming	Allows a categorisation system to ‘evolve’ over many iterations from relatively random/inaccurate to an optimised categorisation.
Naïve Bayes	Uses Bayesian probability to develop classification models from training data. These models could be used for tasks such as Aphasia severity prediction and type classification.
Natural Language Processing	Natural Language Processing is the use of AI to understand and produce unstructured human language (natural language), as opposed to structured, machine-readable text such as programming code. Common examples of NLP in general use are voice assistants and search engine queries. NLP is widely to used to assess communication via automation tools and build text-based systems (therapy software/chatbots) for Aphasia.
Neural network	Neural networks are conceptually similar to the human brain in that they contain nodes and connections. Nodes activate other nodes through connections when a particular threshold is reached. Nodes are arranged in layers. An input layer is fed the data (e.g. an image), a series of hidden layers process the data using various methods, usually not comprehensible to humans, and the output layer provides a response (e.g. 0.97 probability image contains ‘thumbs down’ gesture). During training, the output from prelabelled training data is compared to the desired output, and the network adjusts the connections until its accuracy improves. These models could be used for tasks such as Aphasia severity prediction and type classification.
Optimisation	Optimisation is the adjustment process of machine learning, where the models continue to learn until the difference between the desired output (the human labelled data) and the model output is as low as possible. For example, initially, a neural network may categorise data as aphasic/non-aphasic only 50% accurately, but improve to 90% after training.
Recurrent neural network	In, recurrent neural networks, nodes are able to learn and analyse data that contains a temporal dimension (in sequences) by retaining or ‘remembering’ information from previous data in the set. For example, recurrent neural networks are particularly useful for comprehending the meaning of a sentence through retaining the meaning of each word as it is processed. These models could be used for tasks such as Aphasia severity prediction and type classification.
Self-organising maps	Self organising maps are a tool for sorting data which ultimately allows identification of categories. SOMs represent complex data on a two dimensional ‘map’ according to their similarity across many dimensions. During training, the map gradually adjusts to the data (or self organises). SOMs can be used for profiling and clustering different Aphasia types.
Supervised Learning	In supervised machine learning, labelled datasets (previously annotated by humans) are used to train models, allowing the algorithms to learn the characteristics of the data based on labels. Once trained, the model is used to predict or classify data it has not encountered before. Once trained on labelled sets of data, supervised machine learning models could be used for tasks such as Aphasia severity prediction and type classification.
Support Vector Machine	Support Vector Machines analyse training data spatially, with a dimension for each variable or feature, and calculates the optimal path to ‘slice’ through the data that separates it according to the labelled categories. New data can then be classified using the learned information. These models could be used for tasks such as Aphasia severity prediction and type classification.
Transformers	Transformers are deep learning models that have a mechanism of differentially weighting the significance of input data based on a concept called self-attention.
Unsupervised learning	Unsupervised machine learning models sort uncategorised datasets without labels or examples being provided. The models learn to explore data based on any suitable dimensions and results may or may not align with traditional, human classifications but can reveal previously hidden patterns in data. Compared to supervised machine learning models, the key benefit is that unsupervised models do not depend on annotated/pre-labelled datasets. Unsupervised learning can be used for explorations, clustering and profiling of Aphasia patients based on their data.
Word embedding	Humans largely learn word meanings by hearing them used in context, rather than through explicit definitions. Similarly, by analysing where a word is used or embedded within many sentences, the semantic meaning of words can be learned by machines by investigating the neighbouring words (context), a crucial part of processing and generating natural language. Word embedding models are used in text based assessment systems and word prediction applications for Aphasia.

Data availability statement

The data that support the findings of this study are available via an interactive dashboard hosted at https://bit.ly/3hzLZL3