Personalized disease prediction framework based on genomic variants and disease histories using deep embeddings and alignment-based process conformance checking
This study proposes a novel personalized disease prediction framework that integrates heterogeneous biomedical information, including structured genomic variant annotations, deep semantic embeddings of longitudinal disease histories, and process conformance metrics derived from historical disease pathways. Disease history embeddings were generated using three domain-specialized BERT-based models including BioBERT, BioClinicalBERT and BiomedBERT, to comparatively evaluate the impact of different pretraining strategies on disease prediction performance. Because the concatenation of high-dimensional contextual embeddings and sparse multi-hot variant annotations produces a large feature space compared to the number of available samples, principal component analysis and truncated singular value decomposition were used for BERT embeddings and genomic variant annotations, respectively. Alignment-based process conformance checking was applied to quantify how closely an individual’s disease trajectory conforms to typical progression patterns observed in the population. Seven feature configurations were evaluated, and performance was reported with 95% bootstrapped confidence intervals to quantify estimation uncertainty. The results demonstrate that incorporating conformance-based fitness features improves prediction performance across all disease categories and classifiers, yielding consistently higher AUROC values and lower Brier scores, while embedding-only and genomic variant annotation-only configurations consistently ranked among the lowest-performing models. These findings indicate that process-level disease pathway conformity captures critical temporal and behavioral information not fully represented by genomic or deep semantic features alone, highlighting the importance of integrating genetic, semantic, and process-based signals for personalized disease prediction in precision medicine. As the target diseases were broadly defined in this study, however, future work will target more narrowly defined diseases.