Predicting Severity in People with Aphasia: A Natural Language Processing and Machine Learning Approach.

Area of Science:

• Computational linguistics
• Artificial intelligence in healthcare
• Speech-language pathology

Background:

• Accurate assessment of aphasia severity is crucial for effective speech therapy.
• Manual evaluation by clinicians is time-consuming and relies on limited expertise.
• Developing automated methods can improve efficiency and consistency in severity assessment.

Purpose of the Study:

• To develop and validate natural language processing (NLP) and machine learning (ML) models for predicting aphasia severity.
• To differentiate aphasia types using engineered language features.
• To assist clinicians in planning and monitoring treatment for people with aphasia (PWA).

Main Methods:

• Analysis of transcripts from three discourse elicitation methods.
• Engineering language features from PWA tasks.
• Unstructured k-means clustering for aphasia type identification.
• Development of regression and classification models for severity prediction.

Main Results:

• Clustering revealed distinct aphasia types, validating selected language features.
• The best ML regression model (deep neural network) achieved MAE of 0.0671 and RMSE of 0.0922.
• The best classification model (random forest) achieved 73% overall accuracy, with 87.5% accuracy for mild severity.

Conclusions:

• NLP and ML offer an accurate and cost-effective method for evaluating aphasia severity.
• Automated assessment can significantly aid clinicians in determining rehabilitation strategies.
• This approach enhances the precision and efficiency of care for people with aphasia.