Abstract
<title>Abstract</title> <p>Background Acromegaly has an insidious onset and prolonged diagnostic delay. Vocal changes are a recognized clinical feature, the diagnostic value of deep learning architectures remains underexplored. We aimed to assess Wav2Vec2.0-based voice analysis as a non-invasive digital biomarker for acromegaly. Methods This single-center case-control study included 70 patients with acromegaly and 68 age- and sex-matched controls. Clinical status was defined by standard biochemical criteria, symptoms, comorbidities, and residual tumor. Voice samples were recorded with a standard mobile phone and resampled to 16 kHz. Wav2Vec2.0 was used to extract high-dimensional embedding vectors. These embeddings, were used to train and validate four different machine learning classifiers using an 80%/20% train–test split with five-fold cross-validation. Model calibration was assessed with the Brier score, and clinical utility with decision curve analysis (DCA). Results Demographics were similar between acromegaly (mean age 45.5 years; 57.1% female) and control (mean age 47.5 years; 61.8% female) groups. In five-fold cross-validation, Logistic Regression(LR) performed best, with an AUC of 0.86 (95% CI: 0.79–0.93), sensitivity 0.73 (0.57–0.90), and specificity 0.79 (0.72–0.87). On the test set, LR maintained high accuracy (AUC 0.84; specificity 0.92; sensitivity 0.71) and showed good calibration (Brier score 0.1753). DCA confirmed that LR provided the highest benefit across a clinically relevant range of threshold probabilities (~ 15–85%). Conclusion Voice analysis Wav2Vec2.0-based model demonstrates good diagnostic accuracy for acromegaly. These preliminary findings suggest that voice may serve as a promising non-invasive digital biomarker, pending validation in larger and more diverse cohorts.</p>