Table 2.
Best-performing models (AUC) on hold-out test dataset for the across 14 drugs.
| Drug | N | Best model | AUC % | Sens. % | Spec. % | PPV % | NPV % | Accuracy % |
|---|---|---|---|---|---|---|---|---|
| INH | 33,313 | RF + F2 | 96.50 | 95.00 | 98.00 | 96.80 | 96.90 | 96.80 |
| RIF | 33,087 | RF + F2 | 96.60 | 95.50 | 97.50 | 94.60 | 98.00 | 96.90 |
| EMB | 29,950 | GBT + F1 | 91.70 | 92.00 | 91.20 | 70.20 | 98.00 | 91.40 |
| PZA | 15,107 | GBT + F2 | 88.30 | 84.44 | 92.20 | 71.60 | 96.20 | 90.70 |
| STM | 11,383 | RF + F2 | 89.60 | 88.00 | 91.30 | 84.00 | 93.50 | 90.10 |
| AMK | 17,624 | RF + F1 | 90.70 | 82.30 | 99.00 | 90.00 | 98.20 | 97.50 |
| CAP | 10,215 | RF + F2 | 84.10 | 71.80 | 96.50 | 69.80 | 96.80 | 94.00 |
| KAN | 18,328 | GBT + F2 | 91.00 | 83.50 | 98.68 | 90.50 | 97.50 | 96.70 |
| CIP | 402 | GBT + F2 | 95.40 | 92.30 | 98.50 | 92.30 | 98.50 | 97.50 |
| OFL | 3520 | GBT + F2 | 88.10 | 81.00 | 95.30 | 84.90 | 93.70 | 91.80 |
| MOX | 15,829 | GBT + F2 | 89.10 | 84.30 | 93.90 | 72.50 | 96.90 | 92.40 |
| ETH | 14,649 | GBT + F2 | 86.30 | 81.60 | 91.00 | 69.50 | 95.10 | 89.00 |
| LEV | 15,617 | GBT + F1 | 93.50 | 88.90 | 98.10 | 92.30 | 97.20 | 96.30 |
| RFB | 10,815 | GBT + F2 | 94.90 | 93.50 | 96.30 | 93.80 | 96.20 | 95.30 |
AUC area under curve, Sens sensitivity, Spec specificity, PPV positive predictive value, NPV negative predictive value, F1 variants in drug-specific regions, F2 variants across candidate genes, RF random forest, GBT gradient boosted tree, INH isoniazid, RIF rifampicin, EMB ethambuto, PZA pyrazinamide, STM streptomycin, AMK amikacin, CAP capreomycin, KAN kanamycin, CIP ciprofloxacin, OFL ofloxacin, MOX moxifloxacin, ETH ethionamide, LEV levofloxacin, RFB rifabutin.