ScholarWorks@중앙대학교

초록

While deep learning has achieved revolutionary success in image and natural language processing, traditional gradient boosting-based machine learning (ML) models still dominate in the biomedical domain for tabular data. This study systematically evaluates the performance and efficiency of three ML models (XGBoost, LightGBM, CatBoost) and four deep learning (DL) models on five public biomedical datasets, applying identical preprocessing and hyperparameter tuning. Experimental results show that for small to medium-sized datasets (under 10,000 samples), ML models consistently demonstrated superior performance and speed. On large-scale datasets (over 200,000 samples), DL models showed comparable performance but with significantly decreased efficiency as the number of features increased. In conclusion, gradient boosting-based ML models remain a robust choice for most biomedical tabular problems, while Transformer-based DL models may offer limited benefits only when applied to very large datasets with sufficient computational resources.

키워드