Dataset Information

Representation of EHR data for predictive modeling: a comparison between UMLS and other terminologies.

ABSTRACT:

Objective

Predictive disease modeling using electronic health record data is a growing field. Although clinical data in their raw form can be used directly for predictive modeling, it is a common practice to map data to standard terminologies to facilitate data aggregation and reuse. There is, however, a lack of systematic investigation of how different representations could affect the performance of predictive models, especially in the context of machine learning and deep learning.

Materials and methods

We projected the input diagnoses data in the Cerner HealthFacts database to Unified Medical Language System (UMLS) and 5 other terminologies, including CCS, CCSR, ICD-9, ICD-10, and PheWAS, and evaluated the prediction performances of these terminologies on 2 different tasks: the risk prediction of heart failure in diabetes patients and the risk prediction of pancreatic cancer. Two popular models were evaluated: logistic regression and a recurrent neural network.

Results

For logistic regression, using UMLS delivered the optimal area under the receiver operating characteristics (AUROC) results in both dengue hemorrhagic fever (81.15%) and pancreatic cancer (80.53%) tasks. For recurrent neural network, UMLS worked best for pancreatic cancer prediction (AUROC 82.24%), second only (AUROC 85.55%) to PheWAS (AUROC 85.87%) for dengue hemorrhagic fever prediction.

Discussion/conclusion

In our experiments, terminologies with larger vocabularies and finer-grained representations were associated with better prediction performances. In particular, UMLS is consistently 1 of the best-performing ones. We believe that our work may help to inform better designs of predictive models, although further investigation is warranted.

SUBMITTER: Rasmy L

PROVIDER: S-EPMC7647355 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:Background: Electronic health records (EHR) play an important role for the redefinition of phenotypes in view of the wealth and heterogeneity of information now available from disparate data sources. A recent cross-sectional retrospective study has described the potential of EHR toward type 2 diabetes mellitus (T2D) screening when ad hoc models are used. About 10,000 US patients have been analyzed through a variety of inference techniques applied to all records with a variable degree of completeness. The analyses conducted in the reference study have indicated that EHR phenotypes significantly improved T2D detection. Methods: With these US patients and the T2D data evidenced in the above study, we propose an integrative inference approach that leverages the prediction power of EHR features selected by two well-known methods, Random Forests and Lasso. The goal is 2-fold: reducing the Big Data redundancies potentially harmful to the predictive learning task and exploiting the interconnectivity of EHR features. A mutual information (MI) network is the inference tool used to identify communities useful to prioritize significant T2D features underlying the similarity between patients. Results: Endowed with a different degree of granularity, the communities detected after the application of both methods were centered especially on T2D comorbidities and risk factors. As such, they appear very relevant for assessment of two main issues, T2D disease burden, and prevention. Conclusions: Our analytical approach offers a solution for managing the EHR scale factor in a complex disease context. EHR are rich sources of phenotypic diversity through which novel stratifications of patients are expected. To enable these results, both pre-screening of variables and calibration of risk prediction methods become necessary steps in EHR analyses. We have presented networks identifying major T2D communities. The specific significance assigned to comorbidities and risk factors in relation to T2D can be inferred with accuracy from just a suitably reduced number of EHR features.