Answers are grounded in this paper and cite their sources.
Objectives: Complications from type 2 diabetes mellitus (T2DM) pose a significant clinical and economical burdens on patients, and early prediction of comorbidity onset remains difficult. This study proposes a probabilistic machine learning approach using Bayesian Networks (BNs) to predict both the onset and progression of T2DM-related complications, targeting enhanced risk identification and support clinical decisions. Methods: A retrospective cohort of anonymized patient data was obtained from the Dasman Diabetes Institute in Kuwait, comprising clinical variables, International Classification of Diseases Tenth Revision (ICD-10) codes, and laboratory results. Complication severity was encoded via a three-tier system (mild, moderate, severe) derived from ICD-10 classifications. Two predictive tasks were formulated: identifying new complications, and monitoring the progression of existing ones. Two BN formulations were trained: a Features-Only (FO) BN that uses clinical and laboratory variables as predictors, and a Complications-as-Features (CF) BN, which additionally incorporated the status of other complications as predictors. Results: The FO model achieved 72% PR-AUC for ophthalmic complications and 73% for kidney complications. Adding more complications improved performance: CF achieved 74% PR-AUC for ophthalmic complications and 83% for kidney complications. For comparative analysis, we selected the best-performing BN model, CF (kidney), and evaluated it against logistic regression, random forest, naïve Bayes, and support vector machines. In terms of PR-AUC performance, logistic regression reached 40%; random forest 92%, naïve Bayes 39%, XGBoost 82%, and support vector machine 35%. Conclusions: Bayesian methods provide transparent, interpretable modeling of T2DM complication trajectories and offer a promising tool for risk stratification and early intervention at competitive accuracy rates while mapping complex clinical dependencies.
No takes yet. Share an insight, caveat, or question.
Aljenfawi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: