Kann mithilfe von künstlicher Intelligenz die Nachsorge und Früherkennung von Glottiskarzinomen unterstützt werden?

Background: This study evaluates a machine learning algorithm for clinical diagnostics, aiming to classify and assess healthy and pathological vocal sound signals. The goal is to create an objective scale that differentiates between these signals by training algorithms on large voice datasets. This method aims to detect subtle vocal changes that may indicate medical conditions, potentially leading to earlier diagnoses.

Material and Methods: Vocal recordings from 2,676 individuals (ages 18-99; 1238 males, 1438 females) and 213 patients (ages 19-101; median age 61; 111 males, 102 females) with dysphonia of 10 etiologies were analyzed. Recordings of polysyllabic words at varying intensities were used, converted into Mel frequency cepstrum coefficients. Two models were employed: Method 1 used an unsupervised approach that reconstructed healthy vocal signals, with errors indicating deviations from normal patterns. Method 2 utilized a supervised model for cross-validation, allowing differentiation between healthy and pathological signals and categorizing clinical conditions. Discrimination capacity was evaluated using the AUROC.

Results: Method 1 achieved an AUROC of 0.98, while Method 2 exceeded 0.90 across multiple diagnostic groups, indicating high discriminatory performance. Classification with Method 2 correlated strongly with established diagnoses.

Discussion: Previous research has shown AI"s capability in assessing dysphonia severity. Our findings demonstrate that both AI approaches can reliably distinguish between healthy and pathological vocal signals, irrespective of age and gender.

Conclusion: Artificial intelligence could support glottic carcinoma diagnosis, enabling early detection and monitoring therapy progress, including biofeedback for rehabilitation.

Nein