Detecting Duplicates at Hospital Admission: Comparison of Deterministic and Probabilistic Record Linkage

Effective detection of corresponding or duplicate records in medical data sets is vital for a high quality health care system. We evaluate the efficacy of several current and novel record linkage approaches by modeling a hospital-admission scenario, wherein an incoming patient may or may not have been previously treated. Our work is to develop recommendations for how an automated system could operate in such a scenario, especially regarding comparison and classification. By using a large, anonymous, real-world data set, we can gain insight into the robustness of these methods in a way that artificial data sets cannot provide. Preliminary results show that even minor confounders have deleterious effects on our ability to classify matches. We aim to evaluate and refine a semi-supervised classification technique to cope with these influences.