Temporal tree representation for similarity computation between medical patients

Area of Science:

• Medical Informatics
• Data Science

Background:

• Electronic Health Records (EHRs) contain complex, temporal, and heterogeneous data.
• Computing patient similarity from EHRs is challenging due to data sparsity and irregularity.
• Effective patient similarity methods are crucial for healthcare intelligence, patient stratification, and cohort selection.

Purpose of the Study:

• To propose a novel method for EHR data representation to compute patient similarities.
• To address the challenges of temporal aspects, multivariate, heterogeneous, and irregular data in EHRs.
• To enhance the potential of medical analytics for healthcare intelligence.

Main Methods:

• Developed Temporal Trees, a temporal hierarchical representation for EHR data.
• Utilized temporal co-occurrence to preserve information at different data levels.
• Augmented Temporal Trees with doc2vec embedding for patient similarity computation.
• Empirically evaluated the method on Intensive Care Unit (ICU) EHRs for diagnosis identification.

Main Results:

• The Temporal Trees representation significantly outperformed traditional and state-of-the-art methods.
• Demonstrated superior performance in representing patients and computing their similarities.
• Achieved accurate identification of patients with a specific target diagnosis.

Conclusions:

• Temporal Trees effectively capture temporal relationships in hierarchical medical data.
• This representation enables comprehensive modeling of EHR information for identifying similar patients.
• The proposed method enhances the utility of EHR data for clinical decision-making and research.