JMIR Publications

ABSTRACT

Background:

De-identification of clinical records is a critical step before data can be made publicly available to the research community. This task is usually treated as a sequence labeling issue and ensemble learning is one of the best performing solutions. The significance of the classical rule-based method remains an open issue as a candidate learner.

Objective:

The main objective of this study is to investigate whether a rule-based learner is useful in a hybrid de-identification system and bring suggestions on how to build and integrate a rule-based learner.

Methods:

We choose a data-driven rule-learner named TBED and integrate into the best performed hybrid system in this task.

Results:

On the popular i2b2 de-identification data set, experiments show that TBED can generate high performance with the rules learned. And integrating the rule-based model into an ensemble framework achieves the best performance reported in the community, which reached an F1 score of 96.76%.

Conclusions:

We not only prove the contribution of rule-based method to the current ensemble learning approach for the de-identification of clinical records, but also validate such a rule system could be automatically learned by TBED mechanism, avoiding the high cost and low-reliability manual rule development method. In particular, we boost the ensemble model with rules to the top performance of the de-identification of clinical records.