Abstract:
Electronic Health Records (EHR) are a very valuable data mine for clinical data analytics. The
prevalence of missing data in EHR poses a significant challenge to healthcare analytics and patient
care. This affects the reliability of data-driven decisions. It is evident that healthcare professionals
can obtain an immense advantage and make educated and efficient decisions in clinical context.
To deal with missing data in EHR, apart from other complexities such as sparsity etc. prevalent in
EHR, imputation methodologies are introduced. Most traditional imputation methods are not
suitable for EHR data imputation due to the suboptimal accuracy and generalizability. This study
intends to address the need for a solid and robust EHR imputation methodology, proposing a
transformer-based neural network model to effectively handle various missing data mechanisms,
such as Missing Completely At Random (MCAR), Missing At Random (MAR), Missing Not At
Random (MNAR).
The research adapts the self-attention mechanism in transformers, which enables the model to
dynamically weigh features based on relevance, enhancing adaptability to diverse data patterns. A
structured approach was applied which included data preprocessing, model training and iterative
validation with custom training loops.
The implementation of the transformer neural network was able to outperform the existing EHR
data imputation methodologies by a close yet considerable margin and a demonstration platform
was built to upload, impute and download datasets.
