Artificial Intelligence to Detect Voice Disorders: An AI-Supported Systematic Review of Accuracy Outcomes

Background: The objective of the present systematic review is to identify which artificial intelligence (AI) approaches have been used to successfully detect voice disorders. The review examines studies involving patients with non-neurological voice disorders and controls, where AI was applied to detect voice disorders. The primary outcome of interest is the accuracy of these AI models. Additionally, this review demonstrates how the procedures of conducting a systematic review can be supported by AI. Methods: Studies were eligible for inclusion if they implemented an AI approach to detect non-neurological voice disorders from healthy voice samples. A comprehensive search was conducted using PubMed/MEDLINE, Science Direct, Web of Science, EBSCO, and Scopus databases. Risk of bias was assessed via the Quality Assessment Tool for Diagnostic Accuracy Studies. The occurrences of the most common AI techniques utilized in the literature are presented, and a summary of their abilities to accurately detect a voice disorder is reported. Results: In total, 79 publications met the inclusion criteria. These studies included patient recordings from a variety of voice databases. The most common AI techniques implemented were Support Vector Machines (SVMs) (n = 28) and Convolutional Neural Networks (CNNs) (n = 22). The mean accuracy of the models in detecting voice disorders was 92% across all studies. Nine studies reported 100% accuracy, and 32 studies reported between 95% and 99%. Discussion: Strengths of the evidence include high accuracies across diverse models and datasets. Limitations include a limited variety of datasets and a trend of hyperoptimization without sufficient external validation. Clinicians and researchers should recognize that while current AI models show promise, future studies should prioritize robust external validation and more representative datasets.