TABLE 2.
Sensitivity, specificity and positive predictive value (PPV) of considered ML classification methods
| Disease | ML classification | Sensitivity (%) | Specificity (%) | PPV (%) | Author |
|---|---|---|---|---|---|
| (A) Comparative ML classification studies | |||||
| PKU | LRA | 100 | 99.793 | 17.41 | Baumgartner et al. 14 |
| LRA | 98.0 | 99.9 | – | Baumgartner et al. 13 | |
| LRA | 96.809 | 99.905 | 49.46 | Baumgartner et al. 15 | |
| MMA | NN | 98.0 | – | 98.0 | Baumgartner et al. 16 |
| MCADD | Ridge‐LRA | 100 | 99.987 | 33.90 | Van den Bulcke et al. 19 |
| LRA | 96.83 | 99.992 | 88.41 | Baumgartner et al. 14 | |
| LRA | 95.238 | 99.992 | 88.24 | Baumgartner et al. 15 | |
| 3‐MCCD* | LRA | 95.455 | 99.957 | 33.33 | Baumgartner et al. 15 |
| CH | Bagging‐SVM | 73.33 | 100 | – | Zarin Mousavi et al. 24 |
| CIT2, MET, MMA, PKU, SCADD* | SVM | 91.30 | 36.36 | 19.29 | Lin et al. 7 |
| (B) Single ML classification studies | |||||
| PKU | SVM | 100 | 99.997 (99.971) | – | Chen et al. 21 |
| SVM | 100 (100) | 99.98 (99.96) | – | Chen et al. 20 | |
| LRA | 97.66 | 31.61 | 24.59 | Zhu et al. 26 | |
| MMA | SVM | 100 (100) | 100 (99.79) | – | Hsieh et al. 18 |
| RF | 100 (100) | 89.678 (81.226) | 26.40 (16.40) | Peng et al. 25 | |
| RF | 96.117 (96.117) | 65.143 (28.286) | 28.9 (16.5) | Peng et al. 22 | |
| SVM | 95.9 (81.4) | 95.6 (76.2) | – | Hsieh et al. 17 | |
| MCADD | LRA | 100 (100) | 99.988 (99.924) | 18.2 (3.4) | Wang et al. 23 |
| RL | 100 (100) | 99.901 (98.463) | 93.75 (49.18) | Ho et al. 12 | |
| GA1 | RF | 100 (100) | 94.503 (50.751) | 22.30 (3.10) | Peng et al. 25 |
| 3‐MCCD* | SVM | 100 | 99.936 (99.711) | – | Chen et al. 21 |
| MET | SVM | 100 | 99.986 (99.958) | – | Chen et al. 21 |
| VLCADD | LRA | 100 (100) | 100 (100) | 100 (100) | Wang et al. 23 |
| RF | 100 (100) | 92.786 (92.639) | 23.40 (23.10) | Peng et al. 25 | |
| OTCD | RF | 100 (100) | 99.601 (81.983) | 62.10 (3.50) | Peng et al. 25 |
| SCADD* | LRA | 100 (100) | 99.997 (99.974) | 73.3 (22.0) | Wang et al. 23 |
| CAH | DT | 90.909 (100) | 100 (87.194) | 66.7 (20) | Lasarev et al. 27 |
Note: (A) Values of best performing ML classification methods with highest sensitivity and specificity in comparative studies. If presented in the study, these are the results from largest or unknown validation datasets. (B) All results of studies applying a single classification method. If sensitivity and specificity were not stated in the study, the results are calculated based on the published contingency table and given in italics. Results in brackets show comparison to traditional NBS, where given. Diseases with * are biochemical variations nowadays none as nondiseases. The results from Lin et al. 7 are presented in a separate row, since they only report average evaluation results for groups diseases. Most studies applied sampling algorithms, changing the sick‐to‐control ratio, and reduced datasets, such as only including false positive patients from traditional screening. Hence, the performance results and reference values of Table 2 have to be evaluated and compared carefully.
Abbreviations: CAH, congenital adrenal hyperplasia; CH, congenital hypothyroidism; CIT2, citrullinemia type II; DT, decision tree; GA1, glutaric aciduria type I; LRA, logistic regression analysis; MCADD, medium‐chain acyl‐CoA dehydrogenase deficiency; 3‐MCCD, 3‐methylcrotonyl‐CoA carboxylase deficiency; MET, hypermethioninemia; MMA, methylmalonic aciduria; NN, neural network; OTCD, ornithine transcarbamylase deficiency; PKU, phenylketonuria; RF, random forest; RL, rule learner; Ridge‐LRA, logistic ridge regression; SCADD, short‐chain acyl‐CoA dehydrogenase deficiency; SVM, support vector machine; VLCADD, very long‐chain acyl‐CoA dehydrogenase deficiency.