SO SÁNH HIỆU QUẢ CÁC QUY TRÌNH HỌC MÁY VÀ HỌC SÂU TRONG DỰ ĐOÁN NGUY CƠ ĐỘT QUỴ

Abstract

Objective: This study compares traditional machine learning (ML) pipelines versus modern deep learning (DL) models to identify the optimal approach for stroke risk prediction on tabular data. Subjects and methods: A dual-stream comparative analysis was conducted on 6,387 individuals (including public data and 25% additional private data). The ML stream optimized 5 algorithms (XGBoost, LightGBM, CatBoost, Random Forest) with sampling and feature selection. The DL stream trained specialized models (TabNet, FT-Transformer, ResNet). Performance was evaluated using 5-fold cross-validation. Results: The pipeline combining LightGBM, RandomOverSampler, and Mutual Information feature selection achieved the best performance: Accuracy 95.2%, Macro F1 70.2%. T-test analysis indicated significant superiority over baseline and DL models (p = 0.0403). Conclusion: For medium-sized tabular data, optimized ML models outperform DL approaches. Data augmentation and class imbalance handling are key for prediction improvement.