Machine Learning-based Voice Assessment for the Detection of Positive and Recovered COVID-19 Patients.

The external test was very small: only 30 people, 10 in each group, and the recovered group was 60% correct. The authors say the sample size was limited and no sample size calculation was done, so firm clinical conclusions are not supported. Most positive and recovered patients had symptoms (77%), so asymptomatic infection is not represented. All participants were Italian native speakers from European backgrounds, so the model may not work for other languages or populations. Healthy controls were younger than patient groups (median 41 years versus 57 and 59 years), and groups differed in BMI and smoking, which could affect voice. Recordings were made in quiet hospital rooms with masks removed, so real-world phone use may perform differently. Recovered patients were excluded if they had lung ultrasound score above 3, so people with residual lung disease were not included.

Do not use this as a validated COVID-19 screening test. The 90.24% training-set result came from repeated checks within the same data, while the separate external test had only 30 people and was 60% correct for recovered people. The authors say positive results still need confirmation with standard testing.