Semi-supervised Encoding for Outlier Detection in Clinical Observation Data
Background and ObjectiveTo evaluate the utility of encoding for outlier detection in clinical observation data from Electronic Health Records (EHR).\n\nMethodsThis article presents a semi-supervise encoding approach (super-encoding) for constructing a non-linear exemplar data distribution from EHR data and detecting non-conforming observations as outliers. Two hypotheses are tested using experimental design and non-parametric hypothesis testing procedures to evaluate the outlier detection performance of the semi-supervised encoding approach and increasing demographic precision in encoding.\n\nResultsThe experiments involved applying 492 encoders to 30 laboratory tests extracted from the Research Patient Data Registry (RPDR) from Partners HealthCare. We report results obtained from 14,760 encoders. The semi-supervised encoders (super-encoders) outperformed conventional autoencoders in outlier detection. Adding age at observation to the baseline encoder (that only included observation value as the feature) slightly improved outlier detection. Top-nine performing encoders are introduced. The best outlier detection performance was from a semi-supervised encoder, with observation value as the single feature and a single hidden layer, built on one percent of the data and one percent reconstruction error. At least one encoder had a Youdens J index higher than 0.9999 for all 30 observations.\n\nConclusionGiven the multiplicity of distributions for a single observation in EHR data (i.e., same observation represented with different names or units), as well as non-linearity of human observations, encoding offers huge promises for outlier detection in large-scale data repositories.