<![CDATA[Tuberculosis (TB) requires accessible pre-screening approaches for referral prioritization in resource-limited settings. Cough-acoustic studies have not always reported subject-level partitioning, patient-level evaluation, operating-point selection, and performance uncertainty consistently. This study presents RespiScan, a patient-level triage-support framework using complex Morlet Continuous Wavelet Transform scalograms and ImageNet-pretrained MobileNetV2 with corrected weighted focal loss. We curated a secondary cohort of 198 subjects (87 TB and 111 Non-TB) from TBscreen and Coswara. Audio was standardized to 16-kHz mono, one-second segments, capped at 10 clean recordings per subject, and partitioned strictly by subject into 137 training, 30 validation, and 31 internal-holdout subjects. Training-only augmentation used 10-dB environmental noise for both classes and speed perturbation for Non-TB. We selected the Non-TB: TB weight of 1.5:1.0 from three validation-only settings, then averaged patient scores before applying a validation-derived Youden threshold of 0.424084. On the internal holdout, RespiScan achieved an AUROC of 0.8866 (95% CI: 0.7479-0.9832), 77.42% accuracy, 71.43% sensitivity, and 82.35% specificity. Wide uncertainty, source-label confounding, and exploratory calibration preclude clinical-readiness claims; larger source-balanced prospective validation is required.]]>
Copyrights © 2026
