Comorbidity structure as an inductive bias: Comparing output-head designs for multi-label prediction of diabetes and myocardial infarction complications

Technical Summary

Problem Statement
In clinical machine learning, multi-label prediction tasks for disease complications are frequently modeled using separate sigmoid outputs, treating each target label as an independent event. This approach often overlooks the underlying biological reality that complications frequently arise from related pathophysiological processes. The paper challenges the convention of treating label-dependence mechanisms as generic modeling additions, proposing instead that the choice of output-layer architecture should serve as an explicit inductive bias reflecting specific hypotheses about comorbidity structures.

Methodology
The study evaluates the impact of different output-head designs across two distinct clinical domains: Type 2 Diabetes (T2D) and Myocardial Infarction (MI).

• T2D Task: Predicting three microvascular complications (nephropathy, neuropathy, and retinopathy). Six output-head architectures were compared: an independent baseline, linear additive, multiplicative, symmetric conditional random field (CRF), residual multilayer perceptron (MLP), and a combined additive-multiplicative head.
• MI Task: Predicting three electrophysiological complications (ventricular tachycardia, ventricular fibrillation, and atrioventricular block). Four heads were evaluated: independent baseline, linear additive, multiplicative, and symmetric CRF.
• Experimental Protocol: All experiments utilized a shared-backbone protocol within each disease setting. Performance was assessed across five different training data fractions and validated using seven independent random seeds to ensure robustness.

Key Results
The performance of output-head designs varied significantly based on the alignment between the architectural assumption and the biological structure of the disease:

• Type 2 Diabetes: The symmetric CRF demonstrated the most consistent improvement. It ranked highest at full data capacity and at the two lowest data fractions (including 20% training data), where it was the only interaction-based head to exceed the independent baseline's aggregate mean. Notably, this performance was achieved with only three interaction parameters. In contrast, the residual MLP, despite utilizing 123 interaction parameters, consistently underperformed the baseline across all data fractions. The combined additive-multiplicative head failed to improve robustness and showed the largest negative deviations from the baseline at lower data fractions.
• Myocardial Infarction: No single architecture emerged as a stable winner. Rankings fluctuated across data fractions: the multiplicative head led at 80% and 60% data, the CRF led at 100% and 20%, and the independent baseline led at 40%. This instability suggests that the heterogeneous electrophysiology of MI complications does not align well with a single, fixed interaction mechanism.

Significance and Claims
The paper argues for a shift in how output layers are conceptualized and selected. The findings support a "biology-guided" view of output-layer design, where the choice of architecture is a direct reflection of assumed disease structure rather than a routine hyperparameter tuning exercise.

• Structural Alignment: A small, constrained mechanism (the symmetric CRF) proved most effective when its symmetry matched the shared microvascular structure of T2D complications.
• Heterogeneity: In cases of heterogeneous pathophysiology (MI), no single interaction mechanism provided a stable advantage, highlighting that complex biological structures may not always benefit from forced interaction modeling.
• Reporting Standard: The authors conclude that output-layer choices should be explicitly reported and defended as assumptions about disease structure, moving away from the treatment of these components as generic modeling defaults.