2026 International Conference on Networks Computers and Communications (ICNCCom 2026)

Individualized dietary management of cardiac surgery patients in the post-operative period is a critical clinical challenge affecting wound healing, haemodynamic stability, and long-term outcomes, yet current guidelines offer only population level recommendations. We propose a novel Explainable Machine Learning and Deep Learning framework for post-operative dietary classification. The framework is developed and evaluated on a rule-grounded, algorithmically constructed cohort of 305 cardiac patients—parameterised to published cardiac cohort statistics and augmented with two domain-engineered features (surgery type and recovery day)—yielding a 17-feature input space that enables controlled, fully reproducible evaluation. We benchmark five classifiers (Logistic Regression, Decision Tree, Random Forest, XGBoost, TabNet) that assign patients to three evidence-based regimens: Strict, Moderate, and Standard diets. SHAP [8] and TabNet [12] attention provide dual explainability. Logistic Regression achieved the highest accuracy (77.06%; AUC ROC 0.9153; macro-F1 0.77) and significantly outperformed Random Forest (p = 0.031); SHAP identified recovery day and surgery type as top predictors, consistent with their weighting in the transparent label-generation rule. The framework is deployed as a full-stack clinical web application with role-stratified access, automated PDF reporting, a phase-aware 7-day meal planner aligned to AHA guidelines [3], and a Groq-powered Llama 3.3 70B dietary chatbot grounded by a verbatim classifier-injected system prompt. This work is positioned as an end-to-end proof of-concept evaluation of the proposed clinical decision-support framework, establishing a concrete roadmap for prospective val idation on dietitian-labelled real-world data. To our knowledge, it is the first study to formulate post-operative dietary classification as a supervised ML task, demonstrating the technical feasibility of integrating explainable ML, deep learning, and grounded LLMs into a unified, clinician-interpretable nutritional decision support platform.