Proposta de um modelo de detecção de tuberculose por emissão de plasma em ar pulmonar.

Abstract

Tuberculosis (TB) remains a major global public-health concern, with approximately 10.8 million incident cases and 1.25 million deaths reported in 2023 (World Health Organization, 2024), and disproportionately high incidence in vulnerable populations. Conventional diagnostic methods (smear microscopy, microbiological culture, and molecular tests such as GeneXpert MTB/RIF) rely on laboratory infrastructure that limits their access in low-resource settings. This dissertation investigates the feasibility of a machine-learning model for TB screening based on plasma emission images generated by the BDEE-Device (Breath Diagnostic Electronic Eye Device), a portable instrument that produces microplasma from exhaled pulmonary breath and captures its optical emission as a digital image. The work is organized in the Scandinavian (article-based) format and comprises two papers. Paper I, a systematic review conducted under the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) protocol covering studies published between 2019 and 2024, identified six relevant works whose performance varies widely (sensitivities between 0.52 and 1.00) for technologies such as comprehensive two-dimensional gas chromatography time-of-flight mass spectrometry (GC×GC-TOFMS) and electronic noses combined with algorithms such as Random Forest and support vector machines (SVM); none of the reviewed studies, however, combine plasma emission, image capture, and computer vision-based classification for TB detection, defining the gap that motivates the second paper. Paper II proposes and experimentally evaluates a classification approach on the PEVA dataset (2,154 images from 60 patients; 7 positive, 53 negative), comparing classical classifiers with traditional visual descriptors and three convolutional neural network (CNN) architectures with transfer learning, under stratified cross-validation with strict patient-wise split and a hold-out test set. MobileNetV2 achieved the best performance in cross-validation (area under the ROC curve, AUC = 0.911 ± 0.166) and on the hold-out test set (AUC = 0.873; 95% confidence interval [0.820; 0.915]), with limited sensitivity (0.150). Gradient-weighted Class Activation Mapping (Grad-CAM) interpretability analysis revealed evidence of shortcut learning, and the complementary blob cropping experiment doubled test sensitivity (to 0.300) without degrading AUC. Adomain shift analysis between versions 1.0 and 2.0 of the device revealed a severe drop in performance, suggesting that part of the learned patterns may be instrumental. The reported metrics must be interpreted with caution given the high statistical variance associated with the small number of positive patients (n=7), which reinforces the proof-of-concept nature of this work. The results demonstrate the technical feasibility of this modality as a proof of concept for point-of-care TB screening and point, as central directions, to dataset expansion, prospective multicenter validation, and the investigation of domain-adaptation techniques.