Chronic disease management reveals a fundamental gap in clinical AI: most deployed prediction systems evaluate a single disease in isolation, despite overwhelming clinical evidence that patients develop and manage multiple interrelated conditions simultaneously. This paper presents an architecture for concurrent multi-disease prediction that addresses three core limitations of prior work. First, the patient health state is modeled as a temporal sequence using an LSTM-based encoder that captures disease trajectory alongside current biomarker values. Second, a hybrid knowledge graph is constructed from SNOMED-CT and DisGeNET ontology priors overlaid with data-driven co-occurrence weights. Third, the tabular ensemble and Graph Neural Network are fused through a joint co-training loss enabling shared gradient flow. Experimental results on three benchmark datasets demonstrate strong performance: the Heart Disease model achieves F1=0.8923 and ROC-AUC=0.9310; Parkinson's attains accuracy=0.9231 with perfect recall (1.000); and the Diabetes model achieves ROC-AUC=0.8388. Comorbidity analysis further confirms a +5.17% average heart disease risk elevation in the diabetic cohort, validating inter-disease interaction modeling.
Sawant et al. (2026) studied this question.