Voice analysis to flag early glottic cancer, as proof of concept

Design

Machine-learning diagnostic model with data augmentation

Population

63 patients (25 early glottic carcinoma, 38 laryngeal immobility) plus 40 controls

Primary outcome

Accuracy distinguishing early glottic carcinoma from laryngeal immobility

Effect

72% accuracy

A hoarse voice can mean an early glottic carcinoma or a benign cause of a fixed vocal fold, and telling them apart usually needs endoscopy. This study tested whether machine learning applied to a sustained-vowel recording could help distinguish early glottic carcinoma from unilateral laryngeal immobility.

Using 63 patients (25 early carcinomas, 38 laryngeal immobility) plus 40 controls, with the dataset expanded by augmentation to 650 samples, a combined model analysing fundamental frequency, jitter and shimmer reached 72% accuracy in separating the two. The authors frame it as a non-invasive triage aid, particularly where specialist laryngoscopy is hard to access.

The honest status is early. A 72% accuracy on a small, augmented dataset is a proof of concept, not a screening tool, and it cannot replace laryngoscopy, which remains essential to diagnose and biopsy a suspicious lesion. The idea is promising for triage in low-resource settings but needs larger, real-world validation before any clinical role.

• Machine-learning model used sustained-vowel voice parameters to separate early glottic carcinoma from laryngeal immobility.
• Based on 63 patients (25 carcinoma, 38 immobility) plus 40 controls, augmented to 650 samples.
• The model reached 72% accuracy in distinguishing the two.
• Positioned as a non-invasive triage aid where specialist access is limited.
• Any dysphonia suspicious for cancer still needs laryngoscopy and biopsy.

Why it matters

It points to a cheap, non-invasive way to prioritise who needs urgent laryngoscopy, relevant where specialists are scarce.

Don't overread it

This is a small, data-augmented single-centre model at 72% accuracy; it is not validated for clinical use and cannot substitute for laryngoscopy and biopsy.

The statistics, in plain English

72% accuracy means roughly one in four cases misclassified, which is far short of a diagnostic test; the small sample inflated by augmentation also risks optimistic performance that may not hold in real patients.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.