Table 2.
Confusion matrices and performance characteristics (GLMRidge algorithm).
| Table 2a: Algorithms calibrated to a sensitivity of 0·90 (high coverage - screening) | ||||||
| Age, sex and STIs | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 1,172 | 797,138 | 798,310 | Sensitivity | 0·911 (0·894 - 0·926) | |
| Test- | 115 | 613,489 | 613,604 | Specificity | 0·435 (0·434 - 0·436) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0015 (0·0014 - 0·0016) | |
| NPV | 0·9998 (0·9998 - 0·9998) | |||||
| Age, sex, origin of birth, educational attainment, marital status, place of residence and main source of income | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 1,108 | 487,880 | 488,988 | Sensitivity | 0·861 (0·841 - 0·879) | |
| Test- | 179 | 922,747 | 922,926 | Specificity | 0·654 (0·653 - 0·655) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0023 (0·0021 - 0·0024) | |
| NPV | 0·9998 (0·9998 - 0·9998) | |||||
| Age, sex, origin of birth, educational attainment, marital status, place of residence, main source of income and medical history | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 1,110 | 424,654 | 425,764 | Sensitivity | 0·862 (0·842 - 0·881) | |
| Test- | 177 | 985,973 | 986,150 | Specificity | 0·699 (0·698 - 0·700) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0026 (0·0025 - 0·0028) | |
| NPV | 0·9998 (0·9998 - 0·9998) | |||||
| Table 2b: Algorithms calibrated to a specificity of 0·999 (high risk population) | ||||||
| Age, sex and STIs | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 35 | 1,667 | 1,702 | Sensitivity | 0·027 (0·019 - 0·038) | |
| Test- | 1,252 | 1,408,960 | 1,410,212 | Specificity | 0·999 (0·999 - 0·999) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0206 (0·0144 - 0·0285) | |
| NPV | 0·9992 (0·9991 - 0·9992) | |||||
| Age, sex, origin of birth, educational attainment, marital status, place of residence and main source of income | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 77 | 1,042 | 1,119 | Sensitivity | 0·060 (0·048 - 0·074) | |
| Test- | 1,210 | 1,409,585 | 1,410,795 | Specificity | 0·999 (0·999 - 0·999) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0688 (0·0547 - 0·0853) | |
| NPV | 0·9992 (0·9991 - 0·9992) | |||||
| Age, sex, origin of birth, educational attainment, marital status, place of residence, main source of income and medical history | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 104 | 1,156 | 1,260 | Sensitivity | 0·081 (0·067 - 0·097) | |
| Test- | 1,183 | 1,409,471 | 1,410,654 | Specificity | 0·999 (0·999 - 0·999) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0825 (0·0679 - 0·0991) | |
| NPV | 0·9992 (0·9991 - 0·9992) | |||||
| Table 2c: Algorithms calibrated to maximal Youdens Index (Sensitivity + Specificity – 1) | ||||||
| Age, sex and STIs | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 952 | 424,326 | 425,278 | Sensitivity | 0·740 (0·715 - 0·763) | |
| Test- | 335 | 986,301 | 986,636 | Specificity | 0·699 (0·698 - 0·700) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0022 (0·0021 - 0·0024) | |
| NPV | 0·9997 (0·9997 - 0·9998) | |||||
| Age, sex, origin of birth, educational attainment, marital status, place of residence and main source of income | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 919 | 246,857 | 247,776 | Sensitivity | 0·714 (0·689 - 0·739) | |
| Test- | 368 | 1,163,770 | 1,164,138 | Specificity | 0·825 (0·824 - 0·826) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0037 (0·0035 - 0·004) | |
| NVP | 0·9997 (0·9996 - 0·9997) | |||||
| Age, sex, origin of birth, educational attainment, marital status, place of residence, main source of income and medical history | ||||||
| HIV+ | HIV- | Total | ||||
| Test+ | 981 | 211,973 | 212,954 | Sensitivity | 0·762 (0·738 - 0·785) | |
| Test- | 306 | 1,198,654 | 1,198,960 | Specificity | 0·850 (0·849 - 0·850) | |
| Total | 1,287 | 1,410,627 | 1,411,914 | PPV | 0·0046 (0·0043 - 0·0049) | |
| NPV | 0·9997 (0·9997 - 0·9998) |
The tables depict confusion matrices and actual sensitivities, specificities, positive predictive values (PPVs) and negative predictive values (NPVs) of the best performing GLMridge model. The models are calibrated according to the risk score that yields the desired value in the training data, i.e. when sensitivities are calculated on the validation set the actual sensitives and specificities may differ slightly.