A biologically informed hierarchical feature engineering and learning framework for genetic disorder prediction.

Genetic disorder prediction from diverse clinical and hereditary sources is very difficult due to complex inheritance patterns, high interdependencies among symptoms and high-dimensional mixed-type feature spaces. Traditional machine learning methods usually involve generic preprocessing pipelines and flat single-stage classification which do not effectively utilize the biological structure and hierarchical relationships inherent in genetic disorders. To overcome these issues, this work presents a hierarchical, biology-informed genetic disorder prediction framework consisting of three closely coupled modules. First, a Genetically-Aware Feature Engineering Framework (GAFE-FE) is proposed to convert raw clinical, symptomatic and hereditary data into biologically relevant representations by encoding inheritance logic, symptom co-occurrence embeddings, physiological ratio constraints and missingness indicators. Second, a Consensus Attention-SHAP-Frequency Feature Selection (CASF-FS) approach is proposed to select robust and interpretable predictors by combining the strengths of explainability-driven attention, permutation-based relevance and cross-validation stability. Third, a Hierarchical Two-Stage Classification strategy is adopted where Stage-1 predicts clinically relevant disorder subclasses and Stage-2 makes final genetic disorder predictions by incorporating subclass probability distributions as contextual features with leakage-free out-of-fold training. Experiments were performed on a multi-class genetic disorder dataset with both subclass-level and final disorder-level annotations for direct assessment of both stages. Using GAFE-FE with a baseline XGBoost classifier, the proposed framework attained 89.36% accuracy in Stage-1 subclass prediction and 86.92% accuracy in Stage-2 disorder prediction. By incorporating CASF-FS, it achieves 95.57% accuracy. The overall hierarchical two-stage framework showed the best performance with 95.1% accuracy in Stage-1 and 98.7% accuracy in Stage-2.