SESSION TITLE: Chest Infections Posters: Tuberculosis
SESSION TYPE: Original Investigation Poster
PRESENTED ON: Wednesday, October 28, 2015 at 01:30 PM - 02:30 PM
PURPOSE: Use of chest radiography (CXR) in the evaluation of pulmonary tuberculosis (PTB) suspects in resource-limited settings has been limited by prohibitive costs of film development, and lack of human resources for accurate interpretation. Inter-observer variability is also a concern. Digital CXR and computer aided diagnostics (CAD) have the potential to overcome these obstacles.
METHODS: A systematic review of studies evaluating accuracy of CAD software for diagnosis of PTB, and comparing CAD to CXR interpretation by humans. Two readers searched 4 databases (Embase, PubMed, Scopus & Engineering village) and used pre-specified criteria to select studies for inclusion.
RESULTS: Out of 386 citations, 3 studies met inclusion criteria (2 retrospective, 1 prospective). All 3 evaluated the “CAD4TB” software (Diagnostic Image Analysis Group, Radboud Univ. Med. Centre) which calculates a radiographic abnormality score ranging from 0 to 100; higher scores reflect greater abnormality. All studies were conducted in sub-Saharan Africa. Prevalence of HIV among the TB suspects ranged from 44% to 68%. Reference tests consisted of culture (2 studies) or Xpert MTB/RIF (1 study); and the prevalence of PTB ranged from 33% to 60%. CAD4TB’s accuracy was dependent on the ‘threshold score’ (minimum score for classifying the CXR as abnormal); sensitivity (SN) ranged from 0.47 to 1.0 and specificity (SP) from 0.33 to 0.94. AUC ranged from 0.71 to 0.84. The 2 studies comparing CAD to human readers used different approaches for the comparison. In 1 study, when CAD4TB’s threshold score was set such that SP was equivalent to a non-expert field officer reporting TB-compatible abnormalities (SP=0.41), CAD4TB’s SN (0.86, 95CI:0.75-0.94) was similar to the average of 4 clinical officers (0.83, 95CI:0.75-0.91). In the other study, when the threshold score was set to achieve the same SN as humans reporting TB-specific abnormalities, the CAD4TB’s SP was higher versus a non-expert reader (p=0.02), and lower compared to an expert (p=0.02) (point estimates not reported).
CONCLUSIONS: Published evidence on CAD for TB diagnosis is limited. Future studies should prioritize assessment of diagnostic accuracy compared to human readers. To improve generalizability of the evidence base, CAD needs to be studied in a variety of settings and patient populations, including those of low HIV prevalence.
CLINICAL IMPLICATIONS: The evidence base is too limited to support adoption of existing CAD software as either a diagnostic or a triage test for pulmonary TB.
DISCLOSURE: The following authors have nothing to disclose: Tripti Pande, Chad Cohen, Madhukar Pai, Faiz Ahmad Khan
No Product/Research Disclosure Information