Vocal Age Gap as a Noninvasive Biomarker for Early Detection of Laryngeal Cancer Using Deep Learning

Early vocal changes such as hoarseness or voice fatigue are common and often attributed to normal aging or benign conditions, which can delay the diagnosis of laryngeal cancer. This study introduces the vocal age gap (VAG), an acoustic measure estimating the difference between chronological and predicted vocal age, to aid early detection.

A pretrained Convolutional Recurrent Neural Network was fine-tuned on mel spectrograms of sustained /a/ vowel phonations. The model was trained on recordings from noncancerous and healthy individuals to avoid malignancy bias and achieved a mean absolute error of 3.69 years. VAG's diagnostic potential was assessed by grouping 20 vocal pathologies into Low, Medium, and High cancer risk classes. Between-group differences were evaluated using one-way ANOVA with post hoc comparisons.

Average VAG increased significantly with cancer risk, F(2, 197) = 142.3, P < 0.001, η² = 0.59 (large effect): ∼ 1.0 years for Low Risk, ∼ 3.0 years for Medium Risk, and ∼ 6.5 years for High Risk, including confirmed laryngeal cancers. Post hoc comparisons confirmed significant pairwise differences between all risk groups (P < 0.001). These findings suggest that pathological changes accelerate perceived vocal aging, detectable computationally before conventional symptoms emerge.

VAG provides a scalable and interpretable biomarker for voice-based cancer screening. The VAG framework demonstrates potential for scalable, remote voice-based screening, though clinical implementation requires prospective validation and support for timely clinical evaluation, particularly for older adults and underserved populations.