OBJECTIVES

Data collected in electronic medical records (EMR) or electronic health records (EHR) are subject to various limitations for the assessment of clinical effectiveness and cost-effectiveness as they are sparse, highly fragmented and require breakdown of temporal events to provide valuable insight in the history of the disease and the treatment. Our aim was to review the methods used to process and analyze routinely collected EMR/EHR data to create model input variables for decision-analytic modelling in the field of oncology.

METHODS

A systematic literature review was conducted and reported in compliance with the PRISMA Statement. After the screening, potentially relevant articles were analyzed in full text, and information were systematically extracted. The studies were categorized according to the methods used and their application in the data analysis process.

RESULTS

The search resulted in 112 articles, which were narrowed down to 38 during the title abstract screening. The content of 28 article were extracted during the full text review. Most studies used machine learning (24) models, including deep learning techniques or support vector machines. For the unstructured data usually natural language processing techniques were used, while data pipelines (4) were created to overcome the unique structure of EMR/EHR databases. Concerning the purpose of data processing, articles were classified as follows: determining parameters to describe health states and assess transition probability (15), providing clinical decision support (10), predicting mortality or risk (5) and focusing exclusively on feature selection (4) standardization (4) and temporal sequence alignment (1).

CONCLUSIONS

This review suggests that advanced analytical methods improve the usability of EMRs and EHRs for real-world data analysis and for model input generation in decision analytic modelling. However, further research is needed about the applicability and generalizability of these methods with unstandardized EHRs or EMRs.