Working papers · speech, health & trustworthy AIVigo, 2026

Research / Voice as a clinical signal / [6]

Can the voice alone identify Long COVID patients?

Identifying patients with Long COVID: study of the effectiveness of voice signal analysis and machine learning. JMIR Preprints, 2023.

In one sentence

Spectral features from coughs recorded after physical exertion let a random forest separate Long COVID patients from healthy people with an AUC of 93%.

1stvoice dataset for PASC
4classifiers compared
93%best AUC
86%best F1

The problem

Long COVID (PASC) has a wide and shifting set of symptoms, which makes it hard to tell apart from other conditions. Research on its diagnosis had focused on clinical symptoms, with little attention to the voice.

What we did

We collected a new dataset of voice recordings from people diagnosed with PASC, the first of its kind. Participants recorded coughs and a sustained /a/, at rest and after physical effort. The pipeline enhances the audio, extracts spectral and energy features, trains four classifiers and combines them with a voting system.

What we found

Spectral features were the most informative. Coughs worked better than vowels, and recordings made after physical exertion improved every classifier. The random forest performed best, with an AUC of 93% and an F1-score of 86%.

My contribution

I led the study: designed the recording protocol with the clinical team, built the feature-extraction and classification pipeline, and ran the experiments.

Why it matters

Its finding that physical exertion makes voice recordings more informative is the starting point of the 2026 study in Frontiers in Medicine.

0255075100 MLP 85.0 Random forest 86.0
Figure 1. Best classifiers by F1-score (%). Random forest also reached an AUC of 93%, MLP 89%.

Cite

Ramírez Sánchez J.M., Docío-Fernández L., García Mateo C., Bustillo-Casado M., García-Caballero A.A. Identifying patients with Long COVID: study of the effectiveness of voice signal analysis and machine learning. JMIR Preprints, 2023.