aiqbench analysis
This paper compares standard ML and deep learning classifiers for atrial fibrillation detection on PhysioNet ECG data, with the main variable being the hyperparameter optimizer. Hyperband-tuned XGBoost reached a 96.95% F1 score under inter- and intra-patient splits, which matter more for clinical credibility than the headline number, since class imbalance and cross-patient generalization are where such models often fail. Readers should weigh the result against benchmark-dataset limits and the absence of prospective clinical validation.
Enhancing Atrial Fibrillation Detection Through Hyperparameter Optimization
arXiv:2610.40001v1 Announce Type: new Abstract: This study explores the role of hyperparameter optimization (HPO) in enhancing the detection of Atrial Fibrillation (AF) using both machine learning and deep learning models. We evaluated a range of standard classifiers - Decision Tree, Random Forest, and XGBoost classifiers, alongside deep learning architectures including Recurrent Neural Networks, Deep Neural Networks, and Siamese Neural Networks. To fine-tune these models, we applied four HPO methods: Genetic Algorithm, Hyperband, Optuna, and SMAC3. The experiments were conducted using inter-patient and intra- patient data splits to assess model generalization and robustness in the presence of class imbalance, which is reflective of real-world AF occurrence. Our results, based on PhysioNet electrocardiogram datasets, showed that the Hyperband-optimized XGBoost model outperformed all others, achieving a peak F1 score of 96.95%. These findings highlight the critical role of suitable HPO strategies in enhancing the accuracy and robustness of AF classification outcomes in clinical settings.
Abstract:This study explores the role of hyperparameter optimization (HPO) in enhancing the detection of Atrial Fibrillation (AF) using both machine learning and deep learning models. We evaluated a range of standard classifiers - Decision Tree, Random Forest, and XGBoost classifiers, alongside deep learning architectures including Recurrent Neural Networks, Deep Neural Networks, and Siamese Neural Networks. To fine-tune these models, we applied four HPO methods: Genetic Algorithm, Hyperband, Optuna, and SMAC3. The experiments were conducted using inter-patient and intra- patient data splits to assess model generalization and robustness in the presence of class imbalance, which is reflective of real-world AF occurrence. Our results, based on PhysioNet electrocardiogram datasets, showed that the Hyperband-optimized XGBoost model outperformed all others, achieving a peak F1 score of 96.95%. These findings highlight the critical role of suitable HPO strategies in enhancing the accuracy and robustness of AF classification outcomes in clinical settings.
11 pages, 3 figures, 6 tables. Presented at the ICT Innovation 2025 Conference, Ohrid, North Macedonia. Accepted for publish by Springer in the Communications in Computer and Information Science Series (CCIS). In press
Machine Learning (cs.LG)
arXiv:2610.40001 [cs.LG]
(or arXiv:2610.40001v1 [cs.LG] for this version)
Full text under the publisher's open licence · 190 words