Ordinal evolutionary artificial neural networks for predicting diabetic nephropathy progression

Diabetic Nephropathy (DN) is a complex, multi-factorial condition that often coexists with other diabetes-related comorbidities. Although DN progresses through a series of ordered stages, i.e., from mild to advanced, studies have primarily focused on classifying its presence or absence, i.e., a static risk prediction, rather than its progression. This gap underscores the need for advanced methodologies to predict DN progression, which could improve patient outcomes and optimise healthcare interventions. This study proposes a novel ordinal perspective to predicting DN progression using clinical Electronic Health Records (EHR) data. This approach is based on Ordinal Evolutionary Artificial Neural Networks (OEANNs), which integrate a Cumulative Link Model to perform ordinal predictions, and leverage Evolutionary Algorithms to optimise OEANNs architecture and weights by dynamically adapting to the sparsity of EHR data. The proposed ordinal perspective involves discretising the disease risk into four ordinal severity classes, each representing a different stage of disease progression. In addition, the temporal variability of DN progression is modelled by considering a feature engineering stage that constructs variables capturing early indicators of DN. Experimental results on a clinical EHR dataset demonstrate that OEANNs outperform state-of-the-art nominal and ordinal models, achieving significant improvements in ordinal metrics such as Mean Absolute Error and Quadratic Weighted Kappa. Furthermore, unlike traditional static risk prediction, OEANNs minimise misclassification errors between distant severity stages, enabling more accurate predictions of disease severity. Therefore, the proposed innovative ordinal approach bridges a critical gap in DN management, offering robust, clinically relevant predictions to inform personalised treatment planning.