Long-term, flexible therapy strategies are needed for chronic diseases like cardiovascular ailments, diabetes, and chronic kidney disease and they have to deal with patient heterogeneity, changing physiological states and privacy constraints. This research suggests a framework for privacy-preserving and personalized artificial intelligence that combines Federated Learning (FL), a Res-HyperTransformerNet deep prediction model, Personalized Multi-Agent Reinforcement Learning (PMARL) and Explainable Artificial Intelligence (XAI) for adaptive chronic disease management. The framework is structured with two main aims: (i) to create an accurate and privacy-aware predictive model for chronic disease risk using diverse data sources, and (ii) to permit adaptive, individual therapy optimization through multi-agent reinforcement learning. Federated learning is utilized to conduct the training of Res-HyperTransformerNet over distributed Internet of Medical Things (IoMT) nodes without the requirement of transferring raw patient data. The embeddings that are created from patient data are then passed on to a PMARL module, where several agents optimize therapy dimensions like medications, diet, physical activity, and mental health interventions independently. To make the clinical support more understandable, the SHAP-based explainability method is used for both predictive and decision-making parts. The framework is tested using two public datasets—the CDC Chronic Disease dataset and the UCI Chronic Kidney Disease Risk Factor dataset. The performance is measured by employing classification metrics (accuracy, precision, recall, F1-score, MCC) as well as reinforcement learning metrics (reward score, convergence steps, episode return) and federated system metrics (communication overhead, convergence rounds, and training time). The experiments indicate that the proposed framework demonstrated improved performance in terms of predictive accuracy and policy convergence speed compared to the baseline deep learning and reinforcement learning models, and at the same time it is more cost-effective when it comes to communication in a federated setting. This means that the proposed method suggests the potential of being a reproducible and open AI framework for adaptive chronic disease therapy management. The combination of ResNet and Transformer blocks, i.e., Federated Res-HyperTransformerNet, achieved strong performance in both datasets, namely CDC Chronic Disease (Dataset 1) and UCI CKD Risk Factor (Dataset 2). The obtained accuracy for Dataset 1 was 98.61% and for Dataset 2, it was 97.75%, under the conducted experimental settings.
The molecular axis underlying the epidemiologically established association between periodontitis and coronary heart disease (CHD) remains undefined at mechanistic, systems-level resolution. Classical comparative transcriptomics captures correlational gene lists but cannot reconstruct directional disease-state dynamics or locate the point along a disease-state ordering at which inflammatory resolution fails. ReNODE-Ferro integrates three public bulk transcriptomic datasets (GSE16134, periodontitis, n = 310; GSE28829, CHD staged, n = 29; GSE100927, CHD plaque, n = 104) through a nine-module biologically curated feature space (123 genes), a fully connected latent encoder, a Neural Ordinary Differential Equation (latent dim = 12) learning continuous disease-state dynamics across six biologically defined states (S0–S5), and a model-based reinforcement learning (RL) controller identifying optimal biological intervention policies in latent space. Twenty-four genes were concordantly dysregulated (FDR < 0.05) across all three datasets. Cross-disease fold-change concordance was significant (Spearman r = 0.473, p = 3.29 × 10 − 8, periodontitis vs. CHD staged). Ferroptotic vulnerability and coronary coupling co-varied with r = 0.830 ( p = 3.89 × 10 − 114). Neural ODE tipping-point analysis identified the S2→S3 metabolic strain-to-ferroptosis-primed transition as the transition carrying the highest modelled velocity, although under a held-out validation scheme the velocity profile proved comparatively flat and the location of the maximum unstable across bootstrap refits. S2 was the least responsive state to single-step intervention (3.7%), but it contains only six specimens and was the least stable cluster on resampling. RL policy optimization identified antioxidant restoration, mitochondrial quality enhancement, and pro-resolution amplification as the universal intervention triad (mean reward improvement = 0.583). Robustness analyses added at revision support and delimit these findings. Tissue of origin explained no measurable variance in the nine-module space (PERMANOVA R² < 0.001, p = 1.00) whereas disease status explained 14.0% ( p < 1 × 10⁻⁴), and the ferroptotic vulnerability–coronary coupling association persisted after adjustment for estimated cellular composition (ρ = 0.764) and within every cohort analysed separately (ρ = 0.77–0.89). An independent diffusion-pseudotime ordering reproduced the state sequence (ρ = 0.872) and placed S2 before S3 ( p = 2.4 × 10⁻¹⁰), and leave-one-cohort-out re-derivation recovered the same structure in all three folds. Against 5,000 size-matched random gene panels the curated feature space was strongly informative (empirical p = 2.0 × 10⁻⁴), although an uncurated transcriptome-wide analysis showed the ferroptotic axis to be one of several concordant programmes rather than a uniquely dominant one. Periodontitis and coronary atherogenesis share a conserved ferroptotic-metabolic resolution failure trajectory. ReNODE-Ferro provides a mechanistically grounded, interventionally actionable map of this continuum, establishing ferroptotic vulnerability as a principal and unusually tightly coupled molecular axis of this shared programme. Because all three cohorts are cross-sectional, the reconstructed trajectory is a statistical ordering rather than a demonstrated temporal or causal sequence.
Edge AI-based reinforcement learning is less reliant on mathematical models and relies on experience to help with the design and optimization of Consumer Technology models, making it ideal for learning dynamic treatment techniques. However, the current research still has the following problems: 1. The optimality of the learned strategy is considered without taking risks into account, leading to strategies with certain risks. 2. The issue of distribution shift is ignored, resulting in learned strategies that are completely different from those of doctors. 3. The patient’s historical observational data and treatments are ignored, making it difficult to accurately assess the patient’s state, which in turn leads to the inability to learn the optimal strategy. Based on this, we propose the DOSAC-DTR dynamic treatment strategy model, which integrates Dead-ends and offline monitoring into the Actor-Critic framework. First, we consider the risk of treatment actions recommended by the learned strategy and incorporate the concept of Dead-ends into the Actor-Critic framework. Second, to mitigate the distribution shift problem, we integrate physician monitoring into the Actor-Critic framework. This approach achieves better performance, resulting in lower estimated mortality rates and higher Jaccard index scores.