Towards an explainable Artificial intelligence system for voice pathology identification and post-treatment characterisation

The voice pathology identification task has recently gained great attention. However, several research questions remain open. This study proposes an explainable AI framework to address the implicit role of age in voice pathology recognition and to investigate vocal quality improvement after surgical treatment in organic voice disorders. The aim is also to define an optimal features subset through predictor importance analysis. A set of 287 patients diagnosed with benign lesions of vocal folds (BLVF) and unilateral vocal fold paralysis (UVFP) was enrolled. Classification experiments were performed for female (F) and male (M) groups: they aimed at distinguishing BLVF from UVFP in age-unbalanced (E1) and age-balanced (E2) datasets, differentiating BLVF subclasses (E3), and detecting pre- and post-treatment conditions (E4). The comparison between E1 and E2 suggests that age does not influence the classification performance. In E1, 76% (F) and 81% (M) accuracies were obtained. The best features concerned vocal fold dynamics and articulator positioning for F and M datasets. In E3, an accuracy of 60% was achieved, suggesting that larger datasets are required. In E4, the best models showed 76% (F) and 72% (M) accuracy, with a good sensitivity in detecting pre-treatment patients. The error rate analysis proved that UVFP was the most misclassified group. Moreover, an agreement between the AI outcome and perceptual evaluations was detected for misclassified recordings. These results suggest their clinical relevance to highlight key aspects of voice quality recovery and to define acoustic parameters that otolaryngologists could employ to monitor the patient's follow-up.