JMIR Publications

ABSTRACT

Background:

To provide quality care, modern health care systems must match and link data about the same patient from multiple sources, a function often served by master patient index software. Record linkage in the master patient index is typically performed manually by health care providers, guided by automated matching algorithms. These matching algorithms must be configured in advance, such as by setting the weights of patient attributes, usually by someone with knowledge of both the matching algorithm and the patient population being served.

Objective:

We aimed to develop and evaluate a machine learning-based software tool which automatically configures a patient matching algorithm by learning from pairs of patient records previously linked by humans already present in the database.

Methods:

We built a free and open-source software tool to optimize record linkage algorithm parameters based on historical record linkages. The tool uses Bayesian optimization to identify the set of configuration parameters that lead to optimal matching performance in a given patient population, by learning from prior record linkages by humans. The tool is written assuming only the existence of a minimal HTTP application programming interface, and so is agnostic to the choice of master patient index software, record linkage algorithm, and patient population. As a proof of concept, we integrated our tool with SanteMPI, an open-source master patient index. We validated the tool using several synthetic patient populations in SanteMPI by comparing the performance of the optimized configuration in held-out data to SanteMPI’s default matching configuration using sensitivity and specificity.

Results:

The machine learning-optimized configurations correctly detect over 90% of true record linkages as definite matches in all datasets, with 100% specificity and positive predictive value in all datasets, whereas the baseline detects none. In the largest dataset examined, the baseline matching configuration detects possible record linkages with a sensitivity of 90.2% (95% CI: 88.4, 92.0) and specificity of 100%. By comparison, the machine learning-optimized matching configuration attains a sensitivity of 100%, with a decreased specificity of 95.9% (95% CI: 95.9, 96.0). We report significant gains in sensitivity in all datasets examined, at the cost of only marginally decreased specificity. The configuration optimization tool, data, and dataset generator have been made freely available.

Conclusions:

Our machine learning software tool can be used to significantly improve the performance of existing record linkage algorithms, without knowledge of the algorithm being used or specific details of the patient population being served.