AI Tool for Diabetes Detection Uses Twenty Seconds of Speech

Speech Analysis Identifies Type 2 Diabetes Risk in Seconds

An artificial intelligence (AI) tool for diabetes detection flagged signs of type 2 diabetes from 20 seconds of recorded speech. Researchers presented the data at the annual meeting of the European Association for the Study of Diabetes (EASD) in Milan, Italy. The evaluation evaluated short audio samples to screen for metabolic risk without drawing blood.

Deep tech company thymia developed the model alongside investigators from RMIT University in Melbourne, Australia. The team trained the algorithm on 63,283 voice samples from 21,129 adults across the United Kingdom and the United States. Participants in the training set provided voice data alongside self-reported diabetes diagnoses.

Acoustic Changes in Vocal Tracks Signal Metabolic Strain

Vocal changes linked to type 2 diabetes include elevated roughness, increased hoarseness, and reduced breath control during normal speech. The software measures these acoustic shifts while a person reads aloud. Study participants recorded themselves remotely reading one of Aesop's fables, which provided a standard vocal sample for automated comparison.

Traditional screening requires in-person clinic visits or blood draws. In the United Kingdom, the National Health Service (NHS) offers health checks to adults aged 40 and older every five years, but attendance sits at 40.4%. The study authors noted that undiagnosed cases make up roughly 30% of total diabetes diagnoses in that population.

Clinical Testing Evaluated Self-Reports and Laboratory Results

Researchers validated the speech model across two separate clinical testing groups. In the first group of 7,319 adults in the United Kingdom, the model assigned higher risk scores to people with reported diabetes 80% of the time. The software demonstrated consistent output across male and female participants and across varying age brackets.

A second evaluation assessed 801 individuals who completed home hemoglobin A1c (HbA1c) blood tests within three months of their audio recording. Compared to these objective lab values, the model assigned higher risk scores to individuals with type 2 diabetes 75% of the time. The tool produced an 82% sensitivity rate alongside a 47% false-positive rate.

• Model training drew on 63,283 voice recordings from 21,129 adult participants.
• First cohort testing reached 80% concordance with self-reported diabetes status among 7,319 people.
• Blood-validated testing achieved 82% sensitivity and correctly separated low-risk individuals from diabetic ranges.

Comorbid Conditions and Demographic Gaps Reduced Accuracy

Diagnostic accuracy dropped when the software analyzed recordings from Black participants. The researchers attributed this variation to a low count of Black individuals reporting type 2 diabetes within the dataset. Performance also declined among participants diagnosed with obesity, high blood pressure, or heart disease.

These co-occurring conditions can cause vocal shifts similar to those observed in diabetes. The investigators noted that speech screening is intended to sit alongside standard blood testing rather than replace laboratory diagnostics. Clinicians could potentially use short voice recordings to triage patients before ordering confirmatory blood panels.