ARTIFICIAL INTELLIGENCE-BASED ACOUSTIC VOICE ANALYSIS FOR EARLY DETECTION OF BENIGN AND MALIGNANT LARYNGEAL DISORDERS: A PROSPECTIVE DIAGNOSTIC ACCURACY STUDY

Abstract

Background: Persistent dysphonia may reflect benign, premalignant, or malignant laryngeal disease. Acoustic voice analysis combined with machine learning offers a non-invasive approach to risk stratification, but prospective clinical validation remains limited. Objective: To evaluate the diagnostic performance of an artificial intelligence (AI)-based acoustic voice analysis system for differentiating malignant from non-malignant laryngeal disorders and to examine associations between selected demographic and lifestyle factors and malignant pathology. Methods: In this prospective diagnostic accuracy study, 112 adults with dysphonia lasting longer than two weeks underwent standardised voice recording, VHI-30 and GRBAS assessment, AI-based acoustic analysis, and videolaryngoscopy. Histopathology was performed when biopsy was clinically indicated. The AI system extracted spectral, cepstral, prosodic, and perturbation features and generated diagnostic classifications. Diagnostic performance was assessed using sensitivity, specificity, predictive values, accuracy, area under the receiver operating characteristic curve (AUC), and Cohen's kappa. Results: Final diagnoses comprised 68 benign lesions (60.7%), 12 premalignant lesions (10.7%), and 32 squamous cell carcinomas (28.6%). The manuscript reports sensitivity of 93.8%, specificity of 94.5%, positive predictive value of 92.4%, negative predictive value of 95.6%, accuracy of 93.8%, AUC of 0.956, and kappa of 0.89. Smoking, alcohol use, age older than 60 years, male sex, professional voice use, and gastro-oesophageal reflux disease were associated with malignant lesions in unadjusted analyses. Conclusion: AI-based acoustic voice analysis showed promising discrimination between malignant and non-malignant laryngeal disorders. Because histopathology was available only for a subset and the AI model and validation strategy were incompletely described, the reported estimates should be considered preliminary until verification bias is addressed and the model is externally validated.