Bidirectional Prediction of Diabetes-Hypertension Comorbidity Transitions from Multimodal Outpatient EMRs: A Reference-Deviation-Aware Multimodal Stacking Framework (RD-MMS)

Abstract Background The highly specialized management of chronic diseases creates sys-tematic blind spots in cross-disease risk identification. Existing predictive modelspredominantly rely on inpatient data and focus on unidirectional disease pre-diction, lacking systematic solutions tailored to the characteristics of real-worldoutpatient data. Methods A large-scale real-world outpatient multimodal dataset spanning2023 to 2024 is constructed, and the Reference-Deviation-Aware MultimodalStacking (RD-MMS) framework is proposed to perform bidirectional predictionof transitions between diabetes mellitus (DM) and hypertension (HTN) as arepresentative comorbidity prediction task. Results RD-MMS achieves an area under the receiver operating characteristiccurve (AUROC) of 0.767 for DM-to-HTN and 0.735 for HTN-to-DM in 5-foldpatient-level cross-validation, outperforming existing benchmarks in an outpa-tient setting where laboratory test coverage is only 13–18%. External validationon a 2025 holdout cohort confirms temporal stability, with AUROC values of0.748 and 0.715 and temporal drift below 1.5%. Reference-Deviation (RD) seman-tic perturbation causes AUROC decreases of 0.028–0.030, confirming that themodel depends on clinical semantics. The top 10% high-risk group achieves2.9-fold enrichment in the event rate, and decision curve analysis (DCA) demon-strates consistently superior net benefit over extreme strategies across thresholdprobabilities of 2%–50%. Conclusions Through reference-deviation-aware encoding, four-dimensionalleakage control, and multimodal stacking, the RD-MMS framework achievesrobust performance on low-quality outpatient data, with temporal stability con-firmed by external validation. The core design is decoupled from specific diseasetypes and supports task switching in a configuration-driven manner, providinga complete methodological paradigm for constructing reliable and deployablechronic comorbidity prediction models from real-world outpatient electronicmedical records (EMR).