Accuracy of computer-aided chest X-ray in community-based tuberculosis screening: Lessons from the 2016 Kenya National Tuberculosis Prevalence Survey

Abstract

Community-based screening for tuberculosis (TB) could improve detection but is resource intensive. We set out to evaluate the accuracy of computer-aided TB screening using digital chest X-ray (CXR) to determine if this approach met target product profiles (TPP) for community-based screening. CXR images from participants in the 2016 Kenya National TB Prevalence Survey were evaluated using CAD4TBv6 (Delft Imaging), giving a probabilistic score for pulmonary TB ranging from 0 (low probability) to 99 (high probability). We constructed a Bayesian latent class model to estimate the accuracy of CAD4TBv6 screening compared to bacteriologically-confirmed TB across CAD4TBv6 threshold cut-offs, incorporating data on Clinical Officer CXR interpretation, participant demographics (age, sex, TB symptoms, previous TB history), and sputum results. We compared model-estimated sensitivity and specificity of CAD4TBv6 to optimum and minimum TPPs. Of 63,050 prevalence survey participants, 61,848 (98%) had analysable CXR images, and 8,966 (14.5%) underwent sputum bacteriological testing; 298 had bacteriologically-confirmed pulmonary TB. Median CAD4TBv6 scores for participants with bacteriologically-confirmed TB were significantly higher (72, IQR: 58–82.75) compared to participants with bacteriologically-negative sputum results (49, IQR: 44–57, p<0.0001). CAD4TBv6 met the optimum TPP; with the threshold set to achieve a mean sensitivity of 95% (optimum TPP), specificity was 83.3%, (95% credible interval [CrI]: 83.0%—83.7%, CAD4TBv6 threshold: 55). There was considerable variation in accuracy by participant characteristics, with older individuals and those with previous TB having lowest specificity. CAD4TBv6 met the optimal TPP for TB community screening. To optimise screening accuracy and efficiency of confirmatory sputum testing, we recommend that an adaptive approach to threshold setting is adopted based on participant characteristics.

Diagnostic Accuracy of Computer-Aided Detection During Active Case Finding for Pulmonary Tuberculosis in Africa: A Systematic Review and Meta-analysis

“…Characteristics of included articles are presented in Table 1 . The diagnostic accuracy of CAD was evaluated retrospectively in 3 studies [ 30 , 32 , 34 ], among which 2 included data from national prevalence surveys [ 30 , 32 ]. The remaining 2 studies utilized a prospective study design [ 31 , 33 ], of which 1 reported the data of participants recruited from both community screening and health care facilities (ie, both active and passive case finding) [ 34 ].…”

“…One study stratified accuracy measures based on HIV status [ 31 ], reporting improved AUC in individuals with vs without HIV (CAD4TB v5: 0.80 [95% CI, 0.72–0.87] vs 0.75 [95% CI, 0.68–0.83], respectively, and CAD4TB v6: 0.81 [95% CI, 0.74–0.88] vs 0.76 [95% CI, 0.68–0.84], respectively). Finally, 1 study assessed the differences in CAD accuracy when grouping participant characteristics (sex, age, cough duration, and history of previous TB), with the threshold fixed to 55 (CAD4TB v6) to achieve an overall sensitivity of 0.95 [ 32 ]. The authors reported the lowest specificity in male participants who were older (>41 years of age), had a cough for >2 weeks, and had a history of previous TB (0.38 [95% CI, 0.30–0.46]).…”

“…Additionally, the single study that analyzed data using both active and passive case finding recruitment methods (reported ∼20% proportion recruited through active case finding) was deemed to have high risk of bias [ 34 ]. The CAD threshold was only predefined in 2 studies [ 31 , 33 ], with the 3 remaining studies reporting threshold scores postanalysis [ 30 , 32 , 34 ], and therefore determined to have high risk of bias. All studies had low risk of bias with regards to the reference standard.…”

Computer-aided detection thresholds for digital chest radiography interpretation in tuberculosis diagnostic algorithms

“…However, because it is in essence used as a probabilistic imputation here, it also requires the (implausible) missing at random assumption like the MI/MI approach. Bayesian LCA has only been used in one previous study of CAD thresholds [ 11 ], reflecting its complexity and specialist knowledge required. We have shown that, in our setting, the assumption that individuals who did not undergo testing – usually due to not having symptoms nor abnormalities on chest radiography – did not have tuberculosis is reasonable, even though this also included a smaller group that was eligible for testing, but for other reasons did not have bacteriological test results available.…”

“…Neither approach correctly accounts for those not undergoing microbiological testing, which may be a substantial proportion of individuals in large community screening settings. To date, only one study has attempted to address this problem in a principled way, using latent class analyses (LCA) to account for the missingness in tuberculosis status [ 11 ]. There remains a paucity of guidance for threshold determination when the majority of individuals did not undergo microbiological testing [ 30 ].…”