Table 3. Performance results of the proposed diagnostic algorithm with human reading instead of ACR.
| Reader threshold R | Selected for Xpert (%) | Sensitivity (%) | Specificity (%) | PPV (%) | NPV (%) | CSS ($) | CNTBC ($) | Throughput per day | |
|---|---|---|---|---|---|---|---|---|---|
| Reader 1 | 50 | 63 | 76.1 | 98.4 | 91.5 | 94.8 | 9.66 | 69.40 | 71 |
| Reader 1 | 64 | 32 | 57.7 | 99.1 | 93.2 | 91.3 | 5.68 | 53.73 | 140 |
| Reader 2 | 50 | 56 | 74.6 | 99.1 | 94.6 | 94.6 | 8.75 | 64.04 | 80 |
| Reader 2 | 69 | 31 | 69.0 | 99.4 | 96.1 | 93.5 | 5.47 | 53.15 | 147 |
| Reader 3 | 50 | 53 | 74.6 | 99.1 | 94.6 | 94.6 | 8.43 | 61.07 | 85 |
| Reader 3 | 70 | 28 | 63.4 | 99.4 | 95.7 | 92.4 | 5.14 | 44.30 | 160 |
Throughputs are based on an Xpert capacity of 45/day. For each reader, two thresholds, R, are shown: normal/abnormal CXR (R = 50) and a more progressive threshold (~30% Xpert).
ACR, automated chest radiography; CNTBC, cost per notified TB case; CSS, cost per screened subject; CXR, chest X-ray; NPV, negative predictive value; PPV, positive predictive value; Xpert, Xpert MTB/RIF.