zikabyte/kenari-demo-lr

Kenari demo classifier (research demo, not a clinical tool)

Logistic regression on frozen VGGish embeddings, trained on the public TBscreen dataset (Sharma et al., Science Advances 2024; Zenodo 10431329, CC BY 4.0; Pixel device, Nairobi).

• Only raises examination priority. Never states that someone is safe, TB-free, or TB positive.
• Evidence level: proof of feasibility on one public dataset with internal subject-level cross-validation (AUROC about 0.67 with simple setups). Not validated on an Indonesian population, other phones, or field recordings. The priority threshold is provisional.
• Does not include VGGish weights; load them from the TBscreen code zip (Zenodo 10431329).
• Do not use for diagnosis.