Machine Learning Based Semi-Automatic Iterative Annotation Method of Similar Laboratory Test Item Combination Between...

Data-driven research has emerged as a prominent trend in clinical studies particularly following discussions on the secondary use of clinical datasets. The construction of big data from multiple facilities using Electronic Health Records (EHRs) poses significant challenges. Creation and management of a common laboratory test master across multiple facilities impose a considerable burden in building such databases. Therefore, this study proposes a sustainable method for generating a common master that evolves through the accumulation of the test result data. A classification machine learning model based on XGBoost is proposed to reduce the number of the manually mapped items. The model employs test result statistics and similarity scores of metadata as the classification parameters. Data from 21 facilities were used for the training dataset, while data from 10 other facilities were used for the test dataset. 11 laboratory test items were selected for the initial trial to create a seed model. As a result, the model demonstrated high performance across the proposed evaluation metrics. The Mean Success Rate (MSR) showed consistently high mapping capabilities, while the Productivity Improvement Ratio (PIR) revealed substantial efficiency gains in the mapping process in the first iteration. In conclusion, the proposed model is expected to serve as an iteratively enhancing classification model, improving further as more facility data is incorporated.