{"claims": [{"text": "Epilepsy affects over 50 million people worldwide", "quote_or_locator": "Introduction, Section 1: 'Epilepsy is a chronic neurological disorder characterized by recurrent unprovoked seizures and remains a major global public health challenge, affecting over 50 million people worldwide [1].'"}, {"text": "The study combined data from Tamale Teaching Hospital and OpenNeuro dataset ds005873, initially considering 200 hospital patients and 125 public patients", "quote_or_locator": "Section 3.2: 'The study initially considered multimodal biosignal recordings from 200 eligible patients at Tamale Teaching Hospital and 125 patients from the publicly available OpenNeuro dataset (ds005873).'"}, {"text": "After quality screening, 120 patients were retained", "quote_or_locator": "Section 3.2: 'Following predefined inclusion and exclusion criteria, including modality completeness, signal quality, recording duration, and annotation consistency, 120 patients were selected for model development and evaluation.'"}, {"text": "Data was partitioned at patient level: 70% training, 15% validation, 15% testing, before window segmentation", "quote_or_locator": "Section 3.2: 'After patient selection, the combined multimodal dataset obtained from Tamale Teaching Hospital and the OpenNeuro repository was partitioned at the patient level into 70% training, 15% validation, and 15% testing subsets. The partitioning was performed before window segmentation to ensure that recordings from the same patient could not appear in more than one subset.'"}, {"text": "2-second windows with 50% overlap were used", "quote_or_locator": "Section 3.2: 'Window selection was performed using fixed-duration 2-second windows with 50% overlap, after patient-wise partitioning, to prevent information leakage between the training, validation and unseen datasets.'"}, {"text": "Twelve classifiers were trained: six traditional (logistic regression, SVM, random forest, XGBoost, decision tree, KNN) and six deep (CNN, LSTM, GRU, TCN, Transformer, CNN-LSTM)", "quote_or_locator": "Section 3.10: 'Twelve machine learning and deep learning models were used for a unified comparison. Logistic Regression (LR), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), and XGBoost, Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), Transformer, and a hybrid CNN–LSTM.'"}, {"text": "On the 15% unseen patient-wise test set, the fusion-aware ensemble achieved 97.25% accuracy and 99.60% AUC", "quote_or_locator": "Abstract and Section 4.5, Table 7: 'Experimental results indicate that the proposed method achieved 97.25% accuracy and 99.60% AUC on unseen data' and Table 7 shows 'Fused' model with 97.25% Accuracy and 99.60% AUC on 15% Unseen Dataset."}, {"text": "On the Patient-wise Unseen Hold-Out Test Dataset, accuracy was 95.25% with 99.00% AUC", "quote_or_locator": "Abstract and Section 4.4, Table 6: 'and 95.25% accuracy with 99.00% AUC on the Patient-wise Unseen Test Dataset' and Table 6 shows 'Fused' model with 95.25% Accuracy and 99.00% AUC."}, {"text": "Sensitivity was 96.00%, specificity 94.50% on the Patient-wise Unseen Test Dataset", "quote_or_locator": "Section 4.6, Table 8: 'Sensitivity 96.00%, Specificity 94.50%' and 'The proposed ensemble achieved an AUC of 99.00%, sensitivity of 96.00%, and specificity of 94.50%'"}, {"text": "Stacking ensemble achieved 98.45% accuracy and 97.90% F1-score on the 15% unseen dataset", "quote_or_locator": "Section 4.5, Table 7: 'Stacking 98.45 97.90 99.10' showing Accuracy, F1-Score, and AUC for Stacking on 15% Unseen Dataset."}, {"text": "The proposed fusion-aware ensemble had the smallest generalization gap (ΔAUC = 0.60) compared to other methods", "quote_or_locator": "Section 4.7, Table 10: 'Fused 0.60' and 'The fused model achieved the lowest ΔAUC (0.60), indicating the most stable generalization performance across unseen and unseen hospital data.'"}, {"text": "SVM performance improved from 65% to 92.8% accuracy after fusion reformulation", "quote_or_locator": "Section 4.2, Table 4: 'SVM 65.0 92.8' showing Before and After Accuracy."}, {"text": "Transformer performance improved from 82.0% to 97.4% accuracy after fusion", "quote_or_locator": "Section 4.3, Table 5: 'Transformer 82.0 97.4' showing Before and After Accuracy."}, {"text": "Complete quasi-quadratic fusion achieved 98.10% accuracy, 97.40% F1-score, and 98.90% AUC on internal validation", "quote_or_locator": "Section 4.1, Table 3: 'Complete Quasi-Quadratic Fusion (All Four Parameters) 98.10 97.40 98.90' showing Accuracy, F1-Score, and AUC."}, {"text": "Baseline global-mean fusion achieved 88.60% accuracy on internal validation", "quote_or_locator": "Section 4.1, Table 3: 'Global Mean Fusion Only 88.60 87.30 90.10' showing Accuracy, F1-Score, and AUC."}, {"text": "Particle swarm optimization used 30 particles with 100 maximum iterations", "quote_or_locator": "Section 3.17: 'The PSO optimization employed a swarm size of 30 particles with 100 maximum iterations.'"}, {"text": "Wilcoxon signed-rank test showed statistically significant differences (p < 0.05) between proposed and conventional ensemble methods", "quote_or_locator": "Section 4 and Table 2: 'Table 2 presents the results of the Wilcoxon signed-rank test for paired AUC values from repeated patient-wise validation runs. The statistically significant p-values indicate differences' with p-values of 0.021, 0.008, 0.004 for comparisons versus Stacking, Voting, and Bagging respectively."}], "prompt_version": "p1.0", "verdicts": [{"claim": "Epilepsy affects over 50 million people worldwide", "verdict": "supported", "evidence": "Epilepsy is a chronic neurological disorder characterized by recurrent unprovoked seizures and remains a major global public health challenge, affecting over 50 million people worldwide [1].", "note": null}, {"claim": "The study combined data from Tamale Teaching Hospital and OpenNeuro dataset ds005873, initially considering 200 hospital patients and 125 public patients", "verdict": "supported", "evidence": "The study initially considered multimodal biosignal recordings from 200 eligible patients at Tamale Teaching Hospital and 125 patients from the publicly available OpenNeuro dataset (ds005873).", "note": null}, {"claim": "After quality screening, 120 patients were retained", "verdict": "supported", "evidence": "Following predefined inclusion and exclusion criteria, including modality completeness, signal quality, recording duration, and annotation consistency, 120 patients were selected for model development and evaluation.", "note": null}, {"claim": "Data was partitioned at patient level: 70% training, 15% validation, 15% testing, before window segmentation", "verdict": "supported", "evidence": "After patient selection, the combined multimodal dataset obtained from Tamale Teaching Hospital and the OpenNeuro repository was partitioned at the patient level into 70% training, 15% validation, and 15% testing subsets. The partitioning was performed before window segmentation to ensure that recordings from the same patient could not appear in more than one subset.", "note": null}, {"claim": "2-second windows with 50% overlap were used", "verdict": "supported", "evidence": "Window selection was performed using fixed-duration 2-second windows with 50% overlap, after patient-wise partitioning, to prevent information leakage between the training, validation and unseen datasets.", "note": null}, {"claim": "Twelve classifiers were trained: six traditional (logistic regression, SVM, random forest, XGBoost, decision tree, KNN) and six deep (CNN, LSTM, GRU, TCN, Transformer, CNN-LSTM)", "verdict": "supported", "evidence": "Twelve machine learning and deep learning models were used for a unified comparison. Logistic Regression (LR), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), and XGBoost, Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), Transformer, and a hybrid CNN–LSTM.", "note": null}, {"claim": "On the 15% unseen patient-wise test set, the fusion-aware ensemble achieved 97.25% accuracy and 99.60% AUC", "verdict": "supported", "evidence": "Experimental results indicate that the proposed method achieved 97.25% accuracy and 99.60% AUC on unseen data. Table 7 shows 'Fused' model with 97.25% Accuracy and 99.60% AUC on 15% Unseen Dataset.", "note": null}, {"claim": "On the Patient-wise Unseen Hold-Out Test Dataset, accuracy was 95.25% with 99.00% AUC", "verdict": "supported", "evidence": "Experimental results indicate that the proposed method achieved 97.25% accuracy and 99.60% AUC on unseen data, and 95.25% accuracy with 99.00% AUC on the Patient-wise Unseen Test Dataset. Table 6 shows 'Fused' model with 95.25% Accuracy and 99.00% AUC.", "note": null}, {"claim": "Sensitivity was 96.00%, specificity 94.50% on the Patient-wise Unseen Test Dataset", "verdict": "supported", "evidence": "The proposed ensemble achieved an AUC of 99.00%, sensitivity of 96.00%, and specificity of 94.50%. Table 8 reports Sensitivity 96.00%, Specificity 94.50%.", "note": null}, {"claim": "Stacking ensemble achieved 98.45% accuracy and 97.90% F1-score on the 15% unseen dataset", "verdict": "supported", "evidence": "Table 7 shows Stacking 98.45 97.90 99.10 showing Accuracy, F1-Score, and AUC for Stacking on 15% Unseen Dataset.", "note": null}, {"claim": "The proposed fusion-aware ensemble had the smallest generalization gap (ΔAUC = 0.60) compared to other methods", "verdict": "supported", "evidence": "The fused model achieved the lowest ΔAUC (0.60), indicating the most stable generalization performance across unseen and unseen hospital data. Table 10 shows 'Fused 0.60'.", "note": null}, {"claim": "SVM performance improved from 65% to 92.8% accuracy after fusion reformulation", "verdict": "supported", "evidence": "Table 4 shows 'SVM 65.0 92.8' showing Before and After Accuracy.", "note": null}, {"claim": "Transformer performance improved from 82.0% to 97.4% accuracy after fusion", "verdict": "supported", "evidence": "Table 5 shows 'Transformer 82.0 97.4' showing Before and After Accuracy.", "note": null}, {"claim": "Complete quasi-quadratic fusion achieved 98.10% accuracy, 97.40% F1-score, and 98.90% AUC on internal validation", "verdict": "supported", "evidence": "Table 3 shows 'Complete Quasi-Quadratic Fusion (All Four Parameters) 98.10 97.40 98.90' showing Accuracy, F1-Score, and AUC.", "note": null}, {"claim": "Baseline global-mean fusion achieved 88.60% accuracy on internal validation", "verdict": "supported", "evidence": "Table 3 shows 'Global Mean Fusion Only 88.60 87.30 90.10' showing Accuracy, F1-Score, and AUC.", "note": null}, {"claim": "Particle swarm optimization used 30 particles with 100 maximum iterations", "verdict": "supported", "evidence": "The PSO optimization employed a swarm size of 30 particles with 100 maximum iterations.", "note": null}, {"claim": "Wilcoxon signed-rank test showed statistically significant differences (p < 0.05) between proposed and conventional ensemble methods", "verdict": "supported", "evidence": "Table 2 presents the results of the Wilcoxon signed-rank test for paired AUC values from repeated patient-wise validation runs. The statistically significant p-values indicate differences with p-values of 0.021, 0.008, 0.004 for comparisons versus Stacking, Voting, and Bagging respectively.", "note": null}]}