Speech-based AI-driven Diagnostics and Rehabilitation for Oral and Oropharyngeal Cancer.

Participants were instructed to articulate three sustained vowels (/a/, /i/, /u/) repeatedly at a moderate volume and pace, with three repetitions per vowel and each utterance lasting at least one second. We developed a neuromorphic computing framework that orthogonally decomposes acoustic features into ultra-dimensional omics representations, enabling the characterization of both localized lesions and systemic physiological conditions. The study collected a comprehensive spectrum of biological profiles, including sociodemographic characteristics, tumor metrics, oral function-related factors, patient-reported outcome measures (PROMs), immunoinflammatory indices, and general health status indicators, to thoroughly investigate the paralinguistic representations of transformed speech omics features. These features were then rigorously evaluated for their clinical efficacy across multiple diagnostic tasks, including screening, early detection, pathological diagnosis, disease staging, and risk factor identification.

Inclusion criteria

• diagnosed as having OC or OPC regardless of concrete pathological subtype (e.g., epithelial, adenoid, odontogenic)
• without a history of other tumor-related treatment before the initial evaluation, such as radiotherapy and chemotherapy
• native Chinese speakers

Exclusion criteria

• hearing impairment
• history of stuttering, cerebrovascular accident, brain trauma, neurodegenerative diseases
• severe dental or maxillofacial deformity
• cleft lip, cleft palate, and related post-treatment status
• severe cardiology conditions, breathing/pulmonology disorders, psychiatric disorders or other illnesses preventing patients from receiving standard surgery