Evaluating semantic similarity methods for comparison of text-derived phenotype profiles

Background Semantic similarity is a valuable tool for analysis in biomedicine. When applied to phenotype profiles derived from clinical text, they have the capacity to enable and enhance 'patient-like me' analyses, automated coding, differential diagnosis, and outcome prediction. While a large body of work exists exploring the use of semantic similarity for multiple tasks, including protein interaction prediction, and rare disease differential diagnosis, there is less work exploring comparison of patient phenotype profiles for clinical tasks. Moreover, there are no experimental explorations of optimal parameters or better methods in the area. Methods We develop a platform for reproducible benchmarking and comparison of experimental conditions for patient phentoype similarity. Using the platform, we evaluate the task of ranking shared primary diagnosis from uncurated phenotype profiles derived from all text narrative associated with admissions in the medical information mart for intensive care (MIMIC-III). Results 300 semantic similarity configurations were evaluated, as well as one embedding-based approach. On average, measures that did not make use of an external information content measure performed slightly better, however the best-performing configurations when measured by area under receiver operating characteristic curve and Top Ten Accuracy used term-specificity and annotation-frequency measures. Conclusion We identified and interpreted the performance of a large number of semantic similarity configurations for the task of classifying diagnosis from text-derived phenotype profiles in one setting. We also provided a basis for further research on other settings and related tasks in the area.

Automated summary

Semantic similarity is a valuable tool in biomedical analysis, especially for patient phenotype analysis in clinical tasks. This study developed a reproducible benchmarking platform to evaluate patient phenotype similarity in uncurated phenotype profiles. The results showed that term-specificity and annotation-frequency measures performed the best among the evaluated configurations.