Table 4.
Performance comparison of the baseline and machine learning–optimized matching configurations in SantéMPI in the held-out evaluation sets, for detection of possible record linkages needing manual review.
| Data set | Sensitivity (%) | Specificity (%) | Positive predictive value (%) | |||||
|
|
Baseline (95% CI) | Optimized (change; 95% CI) | Baseline (95% CI) | Optimized (change; 95% CI) | Baseline (95% CI) | Optimized (change; 95% CI) | ||
| FEBRL1a | 95.0 (90.1 to 98.9) | 100.0 (+5.0%; 1.1 to 9.6) | 100.0 (100.0 to 100.0) | 99.4 (−0.6%; −0.4 to −0.7) | 100.0 (100.0 to 100.0) | 62.9 (−37.1%; −44.4 to −29.8) | ||
| FEBRL2 | 88.7 (84.3 to 92.8) | 100.0 (+11.3%; 7.0 to 16.0) | 100.0 (100.0 to 100.0) | 99.3 (−0.7%; −0.6 to −0.7) | 100.0 (100.0 to 100.0) | 15.7 (−84.3%; −86.3 to −82.3) | ||
| FEBRL3 | 87.8 (85.1 to 90.3) | 100.0 (+12.2%; 9.9 to 15.0) | 100.0 (100.0 to 100.0) | 99.3 (−0.7%; −0.7 to −0.7) | 98.9 (97.9 to 99.6) | 26.6 (−72.3%; −74.3 to −70.3) | ||
| FEBRL4 | 90.2 (88.4 to 92.0) | 100.0 (+9.8%; 8.0 to 11.6) | 100.0 (100.0 to 100.0) | 95.9 (−4.1%; −4.1 to −4.0) | 100 (100.0 to 100.0) | 2.4 (−97.6%; −97.7 to −97.5) | ||
| Hawaii | 93.1 (88.9 to 96.6) | 100.0 (+6.9%; 3.4 to 11.1) | 100.0 (100.0 to 100.0) | 98.9 (−1.1%; −1.2 to −0.9) | 99.4 (97.9 to 100.0) | 31.9 (−67.5%; 71.6 to −63.6) | ||
aFEBRL: Freely Extensible Biomedical Record Linkage.