Ontology-based feature engineering in machine learning workflows for heterogeneous epilepsy patient records

Area of Science:

• Biomedical informatics
• Machine learning
• Ontology engineering

Background:

• Biomedical ontologies are crucial for harmonizing heterogeneous clinical data.
• Machine learning (ML) is increasingly used for analyzing complex medical records.
• Feature engineering in ML for medical data remains a challenging area.

Purpose of the Study:

• To evaluate the utility of biomedical ontologies in feature engineering for ML workflows.
• To assess the impact of ontology-based feature engineering on the performance of ML models in classifying neuropathology reports.
• To investigate the effect of ontology integration on the computational efficiency of ML models.

Main Methods:

• A retrospective study using 312 neuropathology reports from epilepsy surgery patients.
• Development of the Epilepsy and Seizure Ontology (EpSO) using Web Ontology Language (OWL).
• Application of three tree-based ML models (logistic regression, random forest, gradient tree boosting) with and without ontology-based feature engineering.
• Evaluation using five-fold cross-validation and performance metrics including recall, balanced accuracy, and Hamming loss.

Main Results:

• Ontology-based feature engineering significantly improved the performance of all tested ML models.
• Performance gains ranged from 33.3% to 54.5% across models.
• A substantial reduction in model training and testing time was observed, up to 93.8% with gradient tree boosting.
• Microscopy, imaging, and immunohistochemistry were identified as key features.

Conclusions:

• Biomedical ontologies play a vital role in feature engineering for ML applications.
• Ontology integration enhances the performance and efficiency of ML models in multi-label classification of heterogeneous clinical data.
• EpSO effectively models complex diagnostic information, improving data accessibility for ML.