Analysis of a probabilistic record linkage technique without human review

Abstract

We previously developed a deterministic record linkage algorithm demonstrating sensitivities approaching 90% while maintaining 100% specificity. Substantially better performance has been reported using probabilistic linkage techniques; however, such methods often incorporate human review into the process. To avoid human review, we employed an estimator function using the Expectation Maximization (EM) algorithm to establish a single true-link threshold. We compared the unsupervised probabilistic results against the manually reviewed gold-standard for two hospital registries, as well against our previous deterministic results. At an estimated specificity of 99.95%, actual specificities were 99.43% and 99.42% for registries A and B, respectively. At an estimated sensitivity of 99.95%, actual sensitivities were 99.19% and 98.99% for registries A and B, respectively. The EM algorithm estimated linkage parameters with acceptable accuracy, and was an improvement over the deterministic algorithm. Such a methodology may be used where record linkage is required, but human intervention is not possible or practical.

Figure 1. Typical two-threshold scheme for probabilistic scores using human review.

Record pairs between the upper and lower thresholds are manually reviewed for true- or false-link status.

Figure 2. Single probabilistic score threshold without human review.

No scores are tagged for human review.

Figure 3. Registry A sensitivities and specificities as a function of match likelihood score.

The EM-estimates are compared with manually reviewed (observed) values.

Figure 4. Registry B sensitivities and specificities as a function of match likelihood score.

The EM-estimates are compared with manually reviewed (observed) values.