01 / Applied AI / 2025–2026
Cough-based TB screening.
TB-Turing / HeAR
01 / SCOPE
Problem and
scope.
The research compared cough-audio models for TB screening and examined how demographic and symptom information affected their predictions. Each patient could contribute several cough recordings, so both prediction and evaluation operated at patient level.
I developed the HeAR pipeline: grouping cough embeddings by patient, training attention-based models, and evaluating audio-only and metadata-assisted configurations with grouped five-fold cross-validation, sensitivity-targeted thresholds and bootstrap analysis. The wider project also included a shared Wav2Vec2 workstream.
02 / IMPLEMENTATION
Implementation
details.
- 01Cough windows
- 02HeAR embeddings
- 03Patient bags
- 04Attention MIL
- 05Patient-level evaluation
Use patient-level evaluation.
Group multiple cough clips into patient bags and split at patient level. This keeps clips from the same person from appearing in both training and validation folds.
Aggregate variable-length patient bags.
Use gated-attention multiple-instance learning over frozen HeAR embeddings to aggregate variable-length patient bags.
Separate audio from context.
Compare audio-only, basic-metadata and symptom-enhanced configurations. Analyse sensitivity-targeted operating points and bootstrap uncertainty alongside AUROC.
03 / RESULTS
Results.
Pooled retrospective internal cross-validation. Symptom-enhanced AUROC: 0.829575; 95% bootstrap interval 0.801032–0.858900. The separate DS1 audio-only result was 0.850330; DS1 and DS2 are not interchangeable.
Context & limitations
These are research results, not clinical validation. Symptom availability, dataset shift and threshold selection limit how the numbers can be interpreted. Patient data and private prediction files are not published here.
OpenMP and CUDA implementations.
Count Gliders / Histogram / Emboss