A novel framework for detection of noncommunicable diseases via prompt engineering and domain knowledge integration

Noncommunicable diseases (NCDs) are the leading cause of global morbidity, yet existing AI models rely predominantly on clinical data, neglecting rich information available in population surveys. This study introduces a novel framework integrating prompt-based domain knowledge extraction with traditional machine learning for NCD detection from survey data. Unlike knowledge graph construction or retrieval-augmented generation approaches requiring extensive infrastructure, our method employs prompt engineering to systematically elicit epidemiological expertise from large language models for feature selection, validated against independent medical expert rankings with strong concordance (Kendall’s tau of 0.71). The dual-track approach fuses prompt-guided feature screening with Recursive Feature Elimination and Cross-Validation, reconciling domain priors with statistical rigor. Applied to Danish population survey data with 18,957 respondents across 243 variables, the framework achieves F1-score of 91.35%, precision of 98.59%, and accuracy of 85.25%, outperforming existing methods by 2%–3% in F1-score and 3%–10% in accuracy. Synthetic data validation confirms robustness across diverse data distributions. Local Interpretable Model-agnostic Explanations reveal that medical expenditure, physical mobility, and self-rated health constitute primary predictors, enabling clinically actionable risk stratification. A deployed web interface demonstrates real-world applicability with sub-100ms inference. The work establishes prompt-mediated knowledge integration as a scalable pathway for healthcare prediction from population-level determinants.