Researchers at Klick Labs converted the recordings into spectrograms, visual representations of sound frequencies, and analyzed them with an Audio Spectrogram Transformer. This approach allowed the model to learn combinations of acoustic patterns directly from the recordings rather than relying primarily on predefined voice features. Participants read randomly selected short English sentences, while their sex was used to stratify the analyses. The study did not require participant age, body mass index, or a fixed voice sample as study criteria, according to the release.

For women, the voice-based models reported a balanced accuracy of 71.3%, an area under the receiver operating characteristic curve of 0.788, and an AUC expressed as a percentage of 78.8%. For men, the corresponding results were 72.4% balanced accuracy, an AUC of 0.779 and 77.9% when expressed as a percentage. Balanced accuracy averages sensitivity and specificity, giving equal weight to identifying people with and without type 2 diabetes. These results measure performance within the analyzed dataset and do not establish that vocal patterns cause diabetes or that the models can determine an individual’s diagnosis.

The study extends an earlier Klick Labs Voice to Diabetes study published in 2023 in Mayo Clinic Proceedings: Digital Health. That earlier work analyzed a fixed one-sentence smartphone recording alongside participants’ age, body mass index and other basic health information, and identified 14 predefined acoustic features associated with type 2 diabetes, including pitch, intensity, shimmer and jitter. It reported conventional accuracy of 89% for women and 86% for men.

The two studies’ reported accuracy figures should not be compared directly. The 2023 study used conventional accuracy, which is the percentage of all predictions that are correct, while the newer study reported balanced accuracy. The investigators said the difference also reflects different inputs, methods, validation approaches and datasets, rather than evidence that the newer models are more accurate. The supplied material does not describe the participants’ diabetes distribution, demographic composition, external validation or performance in routine clinical care, so the reported model performance should not be interpreted as proof of clinical diagnostic utility.

Subject of Research: Smartphone voice-recording analysis for associations with type 2 diabetes in adults

Article Title: Scientists see new vocal patterns associated with type 2 diabetes

Article References: Scientists see new vocal patterns associated with type 2 diabetes. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: type 2 diabetes, vocal biomarkers, smartphone recordings, spectrograms, audio transformer, acoustic patterns, machine learning, biomedical informatics, Scientists, vocal, patterns, associated