Order-Aware Metric Learning for Electronic Health Records Applications

Electronic Health Records, widely used as sources of patient information, include structured data types: nominal, ordinal, and numerical. Effectively utilising these heterogeneous data types remains a fundamental challenge in medical data mining due to their intrinsic differences. While nominal and numerical data are typically handled using conventional encoding and arithmetic methods, ordinal data occupies a unique intermediary position. Ordinal variables embody an inherent ordered structure but often assume uniform intervals between categories, and the subjective nature of patient-reported ordinal values further complicates their utilisation. These challenges necessitate specialised frameworks to handle ordinal features accurately. In this study, we introduce and evaluate our proposed framework, applying it to MIMIC-IV-ED as a publicly available dataset and real-world data from the Emergency Department (ED) of Royal Melbourne Hospital (RMH). Our evaluation comprises two scenarios: one focusing exclusively on ordinal predictors and another involving a combination of nominal, ordinal, and numerical features. In the ordinal-only scenario of RMH, our approach enhances Logistic Regression, improving AUROC from 0.733 to 0.746 and AUPRC from 0.748 to 0.753. Our findings highlight the critical role of updating ordinal values while preserving order, thereby improving predictive accuracy and mitigating subjectivity in ED disposition decisions.