Machine learning detects acromegaly using voice recordings

A recent study has shown how digital voice analysis using short voice recordings from a mobile phone can accurately identify patients with acromegaly.

“We conclude that high-resolution digital analysis of voice samples collected by a mobile phone and machine learning training resulted in a preliminary algorithm with a high accuracy for acromegaly,” the researchers said. [J Clin Endoc Metab 2025;110:983-990]

One hundred fifty-one Swedish patients with acromegaly (13 percent biochemically active and 10 percent newly diagnosed) and 139 matched controls were included in this study. The machine learning model performed better than experienced endocrinologists in identifying patients with acromegaly (area under the receiver operating curve [AUROC]: 0.84 vs 0.69, respectively).

Patients with acromegaly showed a more pronounced self-reported voice problems (median Voice Handicap Index [VHI]: 6 vs 2; p<0.01) and a higher prevalence of clinically significant voice handicap (VHI ≥20: 22.5 percent vs 3.6 percent).

“Patients with acromegaly experience more voice disorders than matched controls,” according to the researchers.

“This study is the first step to further investigate voice as a biomarker of acromegaly, but more training data and further validation in other populations and languages is essential to create a model with higher sensitivity and specificity,” they said.

Biomarker

Several studies have investigated voice as a potential biomarker for the detection of cardiovascular, neurological, and mental disorders that may affect voice production or speech regulation. [Mayo Clin Proc 2023;98:1353‐1375; Digit Biomark 2021;5:78‐88]

The current study is the first of its kind to assess voice as a biomarker for acromegaly detection.

“Our data support previous findings of persistent voice changes in active and controlled disease with no significant difference in voice parameters after treatment despite regression of mucosal edema in the vocal tract,” the researchers said. [Eur Arch Otorhinolaryngol 2013;270:1391‐1396; J Voice 2020;35:932.e13‐932.e27]

“This indicates that reversible soft tissue alterations are not the only factors affecting the voice in acromegaly,” they added.

Changes in speech and voice are potentially driven by the remaining orodental pathologies, prognathism, and other changes in the viscerocranium after biochemical control of acromegaly. [Endocrine 2018;60:323‐328; Exp Clin Endocrinol Diabetes 2011;119:9‐14]

“In our acromegaly cohort, 66 percent had IGF-1 levels within the reference range, but 77 percent were assessed as biochemically controlled by an experienced endocrinologist based on full clinical, biochemical, and radiological assessment at the including centre,” the researchers said.

“Considering the assay variability in IGF-1 measurements, the compound clinical assessment of biochemically active or controlled was used for subgroup analysis in the manuscript as a more robust indicator of clinical activity,” they added. [Front Endocrinol (Lausanne) 2023;14:1266339]

In the current study, mobile phones were used to collect voice recordings from patients with acromegaly and matched controls at Swedish university hospitals. The researchers then assessed the anthropometric and clinical data, as well as the VHI.

Digital voice analysis of a sustained and stable vowel generated a total of 3,274 parameters, which were used to train machine learning models in classifying the speaker as “acromegaly” or “control.”

Seventy-six percent of the data was used to train these models, while the remaining 24 percent was used to assess their performance. For comparison, 12 experienced endocrinologists evaluated the voice recordings of 50 pairs of participants.