{"claims": [{"text": "The CNN–GRU–mRMR model achieved a mean accuracy of 96.74%", "quote_or_locator": "Abstract: 'Across 30 independent simulations, the CNN–GRU–mRMR model achieved a mean accuracy of 96.74%' and Section 4.4: 'the primary result of this study is the 5-fold subject-wise cross-validated accuracy of 96.74% ± 1.23%'"}, {"text": "The dataset includes 24 individuals with MDD and 29 healthy controls", "quote_or_locator": "Section 4.1: 'The dataset includes recordings from 24 individuals with MDD and 29 healthy controls, aged 16–52 years'"}, {"text": "Signals were collected using a wearable three-electrode EEG device", "quote_or_locator": "Section 4.1: 'Signals were collected under both resting-state and stimulation conditions using a wearable three-electrode EEG device'"}, {"text": "A bandpass filter from 4.5 to 45 Hz was applied", "quote_or_locator": "Section 3.1: 'A bandpass filter (4.5–45 Hz) was applied to suppress irrelevant frequencies'"}, {"text": "The CNN outputs a 20-dimensional spatial feature vector and the GRU outputs 100 temporal features, concatenated into 120 dimensions", "quote_or_locator": "Section 4.3.1: 'The output of this branch is a 20-dimensional spatial feature vector' and 'The output of the GRU branch is a 100-dimensional temporal feature vector' and 'the outputs of the CNN and GRU branches are concatenated to form a unified 120-dimensional feature representation'"}, {"text": "The 30 highest-ranked features were retained from the initial 120 after mRMR selection", "quote_or_locator": "Section 3.3: 'From the initial set of 120 CNN–GRU features, the 30 highest-ranked features were retained as the optimal subset'"}, {"text": "The dataset was divided into 70% training (371 segments) and 30% testing (159 segments)", "quote_or_locator": "Section 4.3: 'The dataset was randomly divided into training (70%, 371 segments) and testing (30%, 159 segments) sets'"}, {"text": "Total of 530 EEG segments were analyzed: 290 healthy and 240 depressed", "quote_or_locator": "Section 4.3: 'Each recording was segmented into 10 equal-length epochs, yielding 530 EEG segments in total (290 healthy and 240 depressed)'"}, {"text": "Recall for depression class was 97.9% and 100% for healthy class", "quote_or_locator": "Section 4.4: 'the classifier correctly identified 94 of the 96 depression samples, corresponding to a recall of 97.9%' and 'all 63 healthy control samples were correctly recognized, yielding a 100% recall for the normal class'"}, {"text": "Precision was 100% for depression predictions and 96.9% for healthy predictions", "quote_or_locator": "Section 4.4: 'The precision analysis in Figure. 5(C) demonstrates a precision of 100% for depression predictions and 96.9% for healthy predictions'"}, {"text": "F1-score was 98.94%", "quote_or_locator": "Section 4.4: 'the proposed framework achieved an accuracy of 96.74%, precision of 97.91%, recall of 100%, and an F1-score of 98.94%'"}, {"text": "Correlation coefficient between predicted and actual values was 0.97419", "quote_or_locator": "Section 4.5: 'The high correlation coefficient (R = 0.97419) further confirms the strong linear relationship between predicted and experimental values'"}, {"text": "Regression equation shows output approximately equals 0.98 times target plus 0.042", "quote_or_locator": "Section 4.5: 'the regression equation (Output ≈ 0.98 × Target + 0.042) reveals a very small prediction bias, with a mean prediction error of approximately 0.042'"}, {"text": "CNN component comprises three convolutional layers with 32, 64, and 128 filters respectively", "quote_or_locator": "Section 4.3.1: 'The CNN component comprises three convolutional layers with 32, 64, and 128 filters, respectively'"}, {"text": "GRU component consists of two stacked GRU layers, each with 64 hidden units", "quote_or_locator": "Section 4.3.1: 'The GRU component consists of two stacked GRU layers, each with 64 hidden units'"}, {"text": "Model was trained using Adam optimizer with initial learning rate of 0.001", "quote_or_locator": "Section 4.3.1: 'The model was trained using the Adam optimizer with an initial learning rate of 0.001'"}, {"text": "Batch size was 32 and maximum 100 epochs with early stopping at 15 epochs patience", "quote_or_locator": "Section 4.3.1: 'The batch size was set to 32. The model was trained for a maximum of 100 epochs, with early stopping applied to prevent overfitting (patience = 15 epochs)'"}, {"text": "Total trainable parameters in the framework is approximately 0.85 million", "quote_or_locator": "Section 4.3.2: 'The total number of trainable parameters in the proposed CNN–GRU–mRMR–Dense framework is approximately 0.85 million'"}, {"text": "Training was performed on NVIDIA RTX 3060 GPU with 12 GB VRAM and 32 GB RAM", "quote_or_locator": "Section 4.3.2: 'Training was performed on a system equipped with an NVIDIA RTX 3060 GPU (12 GB VRAM) and 32 GB RAM'"}, {"text": "Total training time for 30 independent runs was approximately 4.5 hours", "quote_or_locator": "Section 4.3.2: 'The approximate total training time for 30 independent runs (each with 70% training / 30% testing split, ~ 371 segments per run) was 4.5 hours'"}, {"text": "Inference time per EEG segment was estimated at 12 milliseconds", "quote_or_locator": "Section 4.3.2: 'Inference time per EEG segment was estimated to be 12 milliseconds'"}, {"text": "Five-fold subject-wise cross-validation was performed to provide robust evaluation", "quote_or_locator": "Section 4.3: 'In addition to the single train/test split, we performed a 5-fold subject-wise cross-validation to provide a more robust evaluation of the proposed framework'"}, {"text": "Comparison study showed CNN-GRU achieved 89.63%, CNN 91.01%, ResNet-50+LSTM 90.02%, and CNN Network 95.00%", "quote_or_locator": "Table 1: Lists comparison methods with accuracies of 89.63%, 91.01%, 90.02%, and 95.00% respectively"}], "prompt_version": "p1.0", "verdicts": [{"claim": "The CNN–GRU–mRMR model achieved a mean accuracy of 96.74%", "verdict": "supported", "evidence": "Abstract: 'Across 30 independent simulations, the CNN–GRU–mRMR model achieved a mean accuracy of 96.74%' and Section 4.4: 'the primary result of this study is the 5-fold subject-wise cross-validated accuracy of 96.74% ± 1.23%'", "note": null}, {"claim": "The dataset includes 24 individuals with MDD and 29 healthy controls", "verdict": "supported", "evidence": "Section 4.1: 'The dataset includes recordings from 24 individuals with MDD and 29 healthy controls, aged 16–52 years'", "note": null}, {"claim": "Signals were collected using a wearable three-electrode EEG device", "verdict": "supported", "evidence": "Section 4.1: 'Signals were collected under both resting-state and stimulation conditions using a wearable three-electrode EEG device'", "note": null}, {"claim": "A bandpass filter from 4.5 to 45 Hz was applied", "verdict": "supported", "evidence": "Section 3.1: 'A bandpass filter (4.5–45 Hz) was applied to suppress irrelevant frequencies'", "note": null}, {"claim": "The CNN outputs a 20-dimensional spatial feature vector and the GRU outputs 100 temporal features, concatenated into 120 dimensions", "verdict": "supported", "evidence": "Section 4.3.1: 'The output of this branch is a 20-dimensional spatial feature vector' and 'The output of the GRU branch is a 100-dimensional temporal feature vector' and 'the outputs of the CNN and GRU branches are concatenated to form a unified 120-dimensional feature representation'", "note": null}, {"claim": "The 30 highest-ranked features were retained from the initial 120 after mRMR selection", "verdict": "supported", "evidence": "Section 3.3: 'From the initial set of 120 CNN–GRU features, the 30 highest-ranked features were retained as the optimal subset'", "note": null}, {"claim": "The dataset was divided into 70% training (371 segments) and 30% testing (159 segments)", "verdict": "supported", "evidence": "Section 4.3: 'The dataset was randomly divided into training (70%, 371 segments) and testing (30%, 159 segments) sets'", "note": null}, {"claim": "Total of 530 EEG segments were analyzed: 290 healthy and 240 depressed", "verdict": "supported", "evidence": "Section 4.3: 'Each recording was segmented into 10 equal-length epochs, yielding 530 EEG segments in total (290 healthy and 240 depressed)'", "note": null}, {"claim": "Recall for depression class was 97.9% and 100% for healthy class", "verdict": "supported", "evidence": "Section 4.4: 'the classifier correctly identified 94 of the 96 depression samples, corresponding to a recall of 97.9%' and 'all 63 healthy control samples were correctly recognized, yielding a 100% recall for the normal class'", "note": null}, {"claim": "Precision was 100% for depression predictions and 96.9% for healthy predictions", "verdict": "supported", "evidence": "Section 4.4: 'The precision analysis in Figure. 5(C) demonstrates a precision of 100% for depression predictions and 96.9% for healthy predictions'", "note": null}, {"claim": "F1-score was 98.94%", "verdict": "supported", "evidence": "Section 4.4: 'the proposed framework achieved an accuracy of 96.74%, precision of 97.91%, recall of 100%, and an F1-score of 98.94%'", "note": null}, {"claim": "Correlation coefficient between predicted and actual values was 0.97419", "verdict": "supported", "evidence": "Section 4.5: 'The high correlation coefficient (R = 0.97419) further confirms the strong linear relationship between predicted and experimental values'", "note": null}, {"claim": "Regression equation shows output approximately equals 0.98 times target plus 0.042", "verdict": "supported", "evidence": "Section 4.5: 'the regression equation (Output ≈ 0.98 × Target + 0.042) reveals a very small prediction bias, with a mean prediction error of approximately 0.042'", "note": null}, {"claim": "CNN component comprises three convolutional layers with 32, 64, and 128 filters respectively", "verdict": "supported", "evidence": "Section 4.3.1: 'The CNN component comprises three convolutional layers with 32, 64, and 128 filters, respectively'", "note": null}, {"claim": "GRU component consists of two stacked GRU layers, each with 64 hidden units", "verdict": "supported", "evidence": "Section 4.3.1: 'The GRU component consists of two stacked GRU layers, each with 64 hidden units'", "note": null}, {"claim": "Model was trained using Adam optimizer with initial learning rate of 0.001", "verdict": "supported", "evidence": "Section 4.3.1: 'The model was trained using the Adam optimizer with an initial learning rate of 0.001'", "note": null}, {"claim": "Batch size was 32 and maximum 100 epochs with early stopping at 15 epochs patience", "verdict": "supported", "evidence": "Section 4.3.1: 'The batch size was set to 32. The model was trained for a maximum of 100 epochs, with early stopping applied to prevent overfitting (patience = 15 epochs)'", "note": null}, {"claim": "Total trainable parameters in the framework is approximately 0.85 million", "verdict": "supported", "evidence": "Section 4.3.2: 'The total number of trainable parameters in the proposed CNN–GRU–mRMR–Dense framework is approximately 0.85 million'", "note": null}, {"claim": "Training was performed on NVIDIA RTX 3060 GPU with 12 GB VRAM and 32 GB RAM", "verdict": "supported", "evidence": "Section 4.3.2: 'Training was performed on a system equipped with an NVIDIA RTX 3060 GPU (12 GB VRAM) and 32 GB RAM'", "note": null}, {"claim": "Total training time for 30 independent runs was approximately 4.5 hours", "verdict": "supported", "evidence": "Section 4.3.2: 'The approximate total training time for 30 independent runs (each with 70% training / 30% testing split, ~ 371 segments per run) was 4.5 hours'", "note": null}, {"claim": "Inference time per EEG segment was estimated at 12 milliseconds", "verdict": "supported", "evidence": "Section 4.3.2: 'Inference time per EEG segment was estimated to be 12 milliseconds'", "note": null}, {"claim": "Five-fold subject-wise cross-validation was performed to provide robust evaluation", "verdict": "supported", "evidence": "Section 4.3: 'In addition to the single train/test split, we performed a 5-fold subject-wise cross-validation to provide a more robust evaluation of the proposed framework'", "note": null}, {"claim": "Comparison study showed CNN-GRU achieved 89.63%, CNN 91.01%, ResNet-50+LSTM 90.02%, and CNN Network 95.00%", "verdict": "supported", "evidence": "Table 1: 'CNN + GRU [45] 89.63', 'CNN [46] 91.01', 'ResNet-50 + LSTM [47] 90.02', 'CNN Network [48] 95.00'", "note": null}]}