An optimal hybrid AI-ResNet for accurate severity detection and classification of patients with aphasia disorder

Assessing speech severity by examining the speech signals of patients with aphasia is significant for providing the best course of therapy. Though there are available various speech recognition approaches as well as assistive medical tools for speech training, there is a lack of aphasia severity assessment models which can be better rehabilitation for aphasia patients. Therefore, to deal with this situation, a novel hybrid attention inception ResNetV2-based chaotic slime mold (HAIR-CSM) approach is proposed in this paper to accurately detect and classify aphasia patient’s severity. To examine the applicability of the HAIR-CSM approach, the speech samples are collected from 91 aphasia-affected subjects. The ambient noises present in raw speech samples are removed and transformed into an appropriate format using preprocessing procedures, namely data cleaning, stop word removal, pre-emphasis, normalization, and windowing. The preprocessed data are then trained using the proposed HAIR-CSM approach which extracts deep acoustic features in speech samples and categories based on their severities into low aphasia quotient (AQ) (i.e., severe case) and high AQ (mild stage). The performance of the proposed HAIR-CSM in assessing aphasia severity is determined using evaluation measures, namely root-mean-square error, precision, mean squared error, accuracy, sensitivity, F1-score, and specificity. The evaluation results indicate that the proposed HAIR-CSM approach achieves greater aphasia severity prediction accuracy of about 98.1%.