Performance Evaluation of Machine Learning Frameworks for Aphasia Assessment.

Area of Science:

• Computational linguistics
• Medical informatics
• Machine learning

Background:

• Speech assessment is vital for aphasia rehabilitation, involving complex and time-consuming tasks.
• Automating these assessments is essential for efficient patient care and diagnosis.

Purpose of the Study:

• To investigate the performance of automatic speech assessment models for patients with aphasia (PWA).
• To compare machine learning frameworks, specifically classical machine learning (CML) and deep neural networks (DNN), for aphasia assessment tasks.

Main Methods:

• Three automatic speech assessment models were evaluated using healthy subjects' datasets, aphasic patients' datasets, and combined datasets.
• A deep neural network (DNN) framework utilizing a convolutional neural network (CNN) was developed and compared against classical machine learning (CML) approaches.
• The study explored direct and indirect transformations of models for aphasia assessment tasks.

Main Results:

• The CNN-based framework demonstrated superior performance over CML frameworks across all dataset types.
• The CNN model achieved 99.23 ± 0.003% accuracy on the healthy individuals' dataset and 67.78 ± 0.047% on the aphasic patients' dataset.
• High-resolution time-frequency images were key to the CNN's enhanced performance.

Conclusions:

• Automated speech assessment for aphasia is attainable using machine learning models, particularly CNNs.
• Model performance is contingent on suitable dataset types, dataset size, and appropriate decision logic within the machine learning framework.
• The findings suggest a promising direction for developing efficient and accurate tools for aphasia diagnosis and management.