JMIR Publications

ABSTRACT

Background:

Early and accurate identification of diabetic complications is crucial for reducing the associated high rates of morbidity and mortality. In particular, microvascular and macrovascular complications represent the most prevalent and critical conditions that determine patient prognosis. Although machine learning methods show promise, several limitations hinder their successful translation into clinical practice. Many methods rely on hard to acquire clinical data and overlook the pathological relationships among complications. Additionally, the propensity of purely data driven models to capture spurious correlations hinders application in clinical practice.

Objective:

This study aims to address these limitations. The goal is to develop a low-cost, clinically applicable framework the screening of major diabetic complications including microvascular and macrovascular conditions. In this way, the research seeks to improve clinical plausibility and facilitate the practical application of classification models in resource limited settings.

Methods:

This study proposes a novel diagnostic framework: the Attention-based Neuro-Symbolic Regularization Framework (ANSR). First, the framework constructs an effective feature set by integrating accessible immune-inflammatory biomarkers. Subsequently, an Attention-based Neural Perception Module (ABNP) is employed to extract non-linear risk features from these data. To bridge the gap between data patterns and clinical logic, a Bi-level Symbolic Logic Regularization Mechanism (BSLR) is introduced. This mechanism enforces constraints through two components: a pairwise co-occurrence regularizer that encodes explicit correlation patterns, and a high-order global pattern regularizer that mines global association rules.

Results:

Experimental results demonstrate that ANSR outperforms all mainstream baseline models in key metrics such as accuracy and Precision-Recall curve. Moreover, the McNemar test confirms that ANSR achieves a statistically significant advantage over all baseline models across all five complications .

Conclusions:

These findings suggest that the proposed ANSR framework has substantial potential for early and accurate screening of major diabetic complications and for supporting reliable clinical decision-making.