thymia T2D speech model

Roughly a million adults in the UK are living with type 2 diabetes that nobody has diagnosed. The pathway meant to find them — the NHS Health Check, a 20-to-30-minute appointment with a blood draw — reaches only about 40% of eligible people. That is a friction problem, not a modelling one, and the disease is doing something measurable in the meantime: poor glycaemic control and neuropathy are associated with vocal straining, hoarseness and a shortened maximum phonation time. Twenty seconds of someone reading a paragraph aloud already carries part of the signal the missed appointment was meant to collect.

This model, built by thymia, turns that observation into a screening score. A single 10-to-20-second recording of a standard read-aloud passage is resampled, silence-trimmed and passed through a frozen in-house paralinguistic speech encoder derived from Google's TRILLsson5; a calibrated logistic-regression head over the resulting 1,024-dimensional embedding returns the probability that the speaker currently has type 2 diabetes. Nothing is refitted per deployment — the same checkpoint scores new speakers unchanged. It sits in a product line built on shared paralinguistic representations that also carries Helios, thymia's non-clinical wellbeing model.

What sets the work apart is the endpoint of its validation. Earlier speech-based diabetes studies ran on a few hundred participants, under laboratory recording conditions or against self-reported labels alone. Here a separately recruited cohort of 7,319 UK adults was scored, and 801 of them were mailed home HbA1c finger-prick kits so predictions could be checked against a blood biomarker rather than against what people recalled about their own charts.

#Key Features

• Twenty seconds of speech, no questionnaire: prediction needs one short read-aloud recording and no demographic, anthropometric or clinical inputs, so it can run remotely on any device with a microphone.
• Blood-biomarker validation: predictions were compared with HbA1c tests run by a UKAS-accredited laboratory within three months of the recording, reaching AUC 0.75 (±0.07) at the diabetic threshold of 48 mmol/mol and 0.73 (±0.06) at the prediabetic threshold.
• Detects current disease, not future risk: unlike the demographic risk equations used in primary care, it separated type 2 from type 1, gestational and drug-induced diabetes (AUC 0.79 against QDiabetes' 0.62).
• Subgroup performance reported in full: discrimination was stable across sex (0.79 male, 0.81 female) and age, and fell in participants with hypertension (0.65), cardiovascular disease (0.69) and obesity (0.73).
• Calibrated probabilities: Platt scaling gives an expected calibration error of 0.019, so a score reads as a rate rather than only as a ranking.

#Technical Details

Training used a proprietary corpus of 63,283 voice samples from 21,129 UK and US residents collected on thymia's research platform, comprising one read-aloud task (the Aesop fable "The North Wind and the Sun") and two free-speech prompts. Roughly 10% of samples were dropped for insufficient speech, poor recording quality or missing questionnaire fields, with Deepgram's base transcription model used only for silence trimming and quality control. Training labels came from a health questionnaire that did not distinguish diabetes type, giving a 6.3% mixed-type prevalence of which an estimated 80–90% is type 2; retraining on clean type 2 labels moved AUC from 0.740 to 0.775, short of significance (p = 0.34). The classifier is L2-regularised logistic regression (C = 0.001, lbfgs, balanced class weights) over standard-scaled 1,024-dimensional embeddings, calibrated by 3-fold sigmoid Platt scaling. On the 7,319-person cohort (3.0% type 2 prevalence) it reached AUC 0.80 ± 0.03 with sensitivity 0.76 and a 31% false-positive rate, against 0.86 ± 0.03 for QDiabetes-2018, the NICE-recommended risk tool that requires age, BMI, ethnicity, family history and clinical variables.

#Applications

The intended role is a low-friction first filter ahead of blood testing: a recording taken before or during a primary-care appointment, over the phone, or inside a voice-agent interaction, producing a risk tier that decides who is invited for a confirmatory HbA1c test. The authors argue the approach is not disease-specific, so one recording could be scored for several conditions at once. thymia lists type 2 diabetes among the clinical capabilities it groups under Apollo, its UKCA Class I voice-biomarker device, whose shipping API currently returns depression and anxiety scores only.

#Impact

Pairing predictions with concurrent blood sampling at this scale moves the question from whether diabetes leaves an acoustic trace to whether that trace is usable on consumer devices outside a lab. The honest reading is narrower than the headline. QDiabetes still discriminates better overall, and the paper positions speech as complementary rather than superior. The 801-person HbA1c subsample was selected using the model's own scores — the top and bottom deciles plus a random middle band — so its AUC is not a population-prevalence estimate. Performance drops in exactly the comorbid groups most likely to be screened. The preprint has not been peer reviewed, and no code, weights or training recordings have been released: the speech corpus cannot be shared for consent reasons, and the model is distributed only through thymia's commercial platform.