Predicting missing lab test values in electronic medical record data using graph neural networks

Abstract

Predicting missing lab test values, which have not been conducted for patients, is crucial in extracting eligible patients for clinical trials, as it may enable the extraction of a larger number of patients. However, since not all lab tests are conducted for every patient, and the lab test values used as input variables may also be missing, some form of imputation is necessary. Embedding Propagation (EP), a type of graph neural network, can effectively impute missing values by aggregating information from neighboring nodes to learn embeddings. In this study, we aimed to predict missing lab test values using patient information such as other lab test results and diagnoses, by conducting experiments with EP and the MIMIC-IV electronic medical record dataset. In the experiments, we incrementally introduced missing values into the input variables and compared the performance of EP with that of a Multi-Layer Perceptron (MLP). The experimental results showed that when 50% of the input variables were missing, the mean squared error (MSE) was EP: 2.65 and MLP: 3.15, demonstrating that EP is effective when there is a high level of missing input data.