Table 3.
Genetic Predictive Model Performance
| Mutations Included | All Variables on Learning |
Selected Mutation | Selected Variables on Learning Isolate Set |
Selected Variables on Validation Isolate Set |
||||
|---|---|---|---|---|---|---|---|---|
| Sensitivity (%) | Specificity (%) | Number | Sensitivity (%) | Specificity (%) | Sensitivity (%) | Specificity (%) | ||
| INH | 220 | 96 ± 1 | 98 ± 2 | 18 | 95 ± 1 | 98 ± 2 | 94 ± 1 | 94 ± 3 |
| RIF | 85 | 93 ± 1 | 98 ± 1 | 14 | 92 ± 1 | 98 ± 1 | 93 ± 1 | 95 ± 2 |
| PZA | 127 | 72 ± 2 | 97 ± 1 | 124 | 72 ± 3 | 96 ± 1 | 64 ± 3 | 92 ± 3 |
| EMB | 126 | 84 ± 2 | 91 ± 2 | 18 | 83 ± 2 | 89 ± 2 | 80 ± 2 | 82 ± 3 |
| STR* | 176 | 65 ± 2 | 97 ± 1 | 37 | 61 ± 2 | 97 ± 1 | 54 ± 3 | 94 ± 2 |
| ETH | 110 | 65 ± 2 | 92 ± 2 | 20 | 55 ± 3 | 90 ± 2 | 54 ± 3 | 89 ± 3 |
| KAN | 19 | 66 ± 4 | 98 ± 1 | 2 | 62 ± 4 | 99 ± 0.5 | 66 ± 5 | 98 ± 1 |
| CAP | 66 | 43 ± 3 | 96 ± 1 | 5 | 38 ± 2 | 96 ± 1 | 38 ± 3 | 95 ± 2 |
| AMK | 47 | 85 ± 3 | 98 ± 1 | 2 | 82 ± 3 | 98 ± 1 | 79 ± 5 | 97 ± 1 |
| CIP | 26 | 56 ± 4 | 98 ± 1 | 7 | 52 ± 5 | 99 ± 0.4 | 51 ± 5 | 100 ± 0.0 |
| LEVO | 18 | 77 ± 5 | 99 ± 0.3 | 8 | 74 ± 5 | 99 ± 0.4 | 63 ± 9 | 99 ± 1 |
| OFLX† | 19 | 83 ± 5 | 88 ± 3 | 6 | 77 ± 5 | 90 ± 2 | 74 ± 15 | 90 ± 6 |
| PAS† | 13 | 18 ± 5 | 99 ± 0.3 | 4 | 14 ± 5 | 99 ± 0.2 | 13 ± 9 | 99 ± 1 |
Definition of abbreviations: AMK = amikacin; CAP = capreomycin; CIP = ciprofloxacin; EMB = ethambutol; ETH = ethionamide; INH = isoniazid; KAN = kanamycin; LEVO = levofloxacin; OFLX = ofloxacin; PAS = paraaminosalicylic acid; PZA = pyrazinamide; RIF = rifampicin; STR = streptomycin.
Bootstrap SEs are reported.
*
For STR we also ran the prediction model after removal of gid_E92D. This resulted in a decrease in the sensitivity of prediction model by 2% but no change in the specificity.
†
Tenfold cross-validation results shown for OFLX and PAS in seventh and eighth columns.