{"claims": [{"text": "The dataset consists of 34 clients, 14 chosen EEG channels, 27.15 hours of recordings at 128 Hz and 12,512,552 EEG samples.", "quote_or_locator": "Abstract: 'The data consists of 34 clients, 14 chosen EEG channels, 27.15 hours of recordings at 128 Hz and 12,512,552 EEG samples.'"}, {"text": "Isolated local learning achieved accuracy of 0.8840 and macro F1 score of 0.8610.", "quote_or_locator": "Abstract: 'the performance of isolated local learning is weak, with accuracy of 0.8840 and macro F1 score of 0.8610.'"}, {"text": "The global Federated Averaging model achieved accuracy of 0.9214 and macro F1 score of 0.9107.", "quote_or_locator": "Abstract: 'The global Federated Averaging model further improves these results to an accuracy of 0.9214 and macro F1 score of 0.9107'"}, {"text": "The personalized model achieved accuracy of 0.9476, precision of 0.9526, recall of 0.9164 and macro F1 score of 0.9260.", "quote_or_locator": "Abstract: 'The customized model had an accuracy of 0.9476, precision of 0.9526, recall of 0.9164 and macro F1 score of 0.9260.'"}, {"text": "The dataset contains 34 clients with sample sizes ranging from 214,540 to 483,732, with a mean of 368,016 samples per client.", "quote_or_locator": "Table 2: 'Minimum samples/client 214,540; Maximum samples/client 483,732; Average samples/client 368,016'"}, {"text": "Personalization produced a mean accuracy improvement of 0.0262, t(33)=3.12, p=0.0037, with a 95% confidence interval of [0.0091, 0.0433] and Cohen's dz=0.535.", "quote_or_locator": "Table 9: 'personalization produced a mean accuracy improvement of 0.0262, t(33)=3.12, p=0.0037, with a 95% confidence interval of [0.0091, 0.0433] and a moderate paired effect size of dz =0.535.'"}, {"text": "Precision increased by 0.0298, t(33)=3.45, p=0.0016, with a 95% confidence interval of [0.0122, 0.0474] and dz=0.592.", "quote_or_locator": "Table 9: 'Precision increased by 0.0298, t(33)=3.45, p=0.0016, with a 95% confidence interval of [0.0122, 0.0474] and dz=0.592.'"}, {"text": "Macro F1 improved by 0.0153, t(33)=2.78, p=0.0089, with a 95% confidence interval of [0.0041, 0.0265] and dz=0.477.", "quote_or_locator": "Table 9: 'Macro F1 improved by 0.0153, t(33)=2.78, p=0.0089, with a 95% confidence interval of [0.0041, 0.0265] and dz=0.477.'"}, {"text": "Recall decreased slightly by 0.0029, but the difference was not statistically significant, t(33)=-0.41, p=0.6845.", "quote_or_locator": "Table 9: 'Recall decreased slightly by 0.0029, but the difference was not statistically significant, t(33)=-0.41, p=0.6845, with a 95% confidence interval of [-0.0173, 0.0115] and dz=-0.070.'"}, {"text": "Mean temporal consistency was 0.9140, meaning consecutive test windows tended to have the same predicted attention state.", "quote_or_locator": "Section 4.8: 'The mean temporal consistency was 0.9140, which means that consecutive test windows tended to have the same predicted attention state.'"}, {"text": "Mean state-transition rate was 0.0860, indicating a low level of switching between low-attention and high-attention states.", "quote_or_locator": "Section 4.8: 'The corresponding mean state-transition rate was 0.0860, indicating a low level of switching between low-attention and high-attention states.'"}, {"text": "Mean Expected Calibration Error was 0.0370, suggesting small difference between predicted confidence and observed classification performance.", "quote_or_locator": "Section 4.8: 'The mean ECE was 0.0370, suggesting that the difference between the confidence predicted and the classification performance observed was small.'"}, {"text": "Mean repeated-window stability was 0.9020, indicating CDT outputs were stable under repeated-window evaluation.", "quote_or_locator": "Section 4.8: 'The mean repeated-window stability was 0.9020, indicating that the CDT outputs were stable under the repeated-window evaluation procedure defined in Section 3.5.'"}, {"text": "47.06% of clients demonstrated strong dominance with one attention state predicted at least 75% of the time.", "quote_or_locator": "Table 13: 'Sixteen clients, 47.06% of all clients, demonstrated strong dominance, with one attention state being predicted at least 75% of the time.'"}, {"text": "41.18% of clients were in moderate dominance category with prevailing attention state between 60% and 74.99%.", "quote_or_locator": "Table 13: 'Fourteen clients (41.18%) were in the moderate dominance category (prevailing attention state between 60% and 74.99%).'"}, {"text": "11.76% of clients showed balanced distribution with predominant state less than 60%.", "quote_or_locator": "Table 13: 'The remaining 4 clients (11.76%) were more evenly distributed with the predominant state being less than 60%.'"}, {"text": "The window size was 256 samples with a window step of 128 samples, providing 50% overlap between windows.", "quote_or_locator": "Section 3.1: 'A window size of 256 samples is used with a window step of 128 samples, which means that there is 50% overlap between windows in the same partition.'"}, {"text": "At 128 Hz, a window of 256 samples covers exactly 2 seconds of brain activity.", "quote_or_locator": "Section 3.1: 'At 128 Hz, a window of 256 samples covers exactly 2 seconds of brain activity, which is a reasonable temporal resolution for attention classification.'"}, {"text": "The model uses three convolutional layers with filter counts of 32, 64, and 64 and kernel sizes of 7, 5, and 3 respectively.", "quote_or_locator": "Section 3.2.1: 'Three such layers are stacked, with filter counts of 32, 64, and 64 and kernel sizes of 7, 5, and 3 respectively.'"}, {"text": "Federated learning was conducted over 10 rounds with 5 local epochs per round and full client participation.", "quote_or_locator": "Section 3.3: 'There are 10 federated rounds of training. The 34 clients in every round take part and locally train 5 epochs and submit their parameters to the aggregator.'"}, {"text": "Personalization involved 3 epochs of fine-tuning on each client's local training data starting from global parameters.", "quote_or_locator": "Section 3.4: 'Each client k starts from the global parameters θglobal and minimizes its local training objective for an additional 3 epochs.'"}, {"text": "The authors acknowledge that federated learning does not constitute a formal privacy guarantee, as shared parameters could be subject to gradient leakage, membership inference, or reconstruction attacks.", "quote_or_locator": "Section 5 (Discussion): 'This does not constitute a formal privacy guarantee, as shared parameters or updates could still be subject to gradient leakage, membership inference, reconstruction attacks or client-level profiling.'"}], "prompt_version": "p1.0", "verdicts": [{"claim": "The dataset consists of 34 clients, 14 chosen EEG channels, 27.15 hours of recordings at 128 Hz and 12,512,552 EEG samples.", "verdict": "supported", "evidence": "Abstract: 'The data consists of 34 clients, 14 chosen EEG channels, 27.15 hours of recordings at 128 Hz and 12,512,552 EEG samples.'", "note": null}, {"claim": "Isolated local learning achieved accuracy of 0.8840 and macro F1 score of 0.8610.", "verdict": "supported", "evidence": "Abstract: 'the performance of isolated local learning is weak, with accuracy of 0.8840 and macro F1 score of 0.8610.'", "note": null}, {"claim": "The global Federated Averaging model achieved accuracy of 0.9214 and macro F1 score of 0.9107.", "verdict": "supported", "evidence": "Abstract: 'The global Federated Averaging model further improves these results to an accuracy of 0.9214 and macro F1 score of 0.9107'", "note": null}, {"claim": "The personalized model achieved accuracy of 0.9476, precision of 0.9526, recall of 0.9164 and macro F1 score of 0.9260.", "verdict": "supported", "evidence": "Abstract: 'The customized model had an accuracy of 0.9476, precision of 0.9526, recall of 0.9164 and macro F1 score of 0.9260.'", "note": null}, {"claim": "The dataset contains 34 clients with sample sizes ranging from 214,540 to 483,732, with a mean of 368,016 samples per client.", "verdict": "supported", "evidence": "Table 2: 'Minimum samples/client 214,540; Maximum samples/client 483,732; Average samples/client 368,016'", "note": null}, {"claim": "Personalization produced a mean accuracy improvement of 0.0262, t(33)=3.12, p=0.0037, with a 95% confidence interval of [0.0091, 0.0433] and Cohen's dz=0.535.", "verdict": "supported", "evidence": "Table 9: 'personalization produced a mean accuracy improvement of 0.0262, t(33)=3.12, p=0.0037, with a 95% confidence interval of [0.0091, 0.0433] and a moderate paired effect size of dz =0.535.'", "note": null}, {"claim": "Precision increased by 0.0298, t(33)=3.45, p=0.0016, with a 95% confidence interval of [0.0122, 0.0474] and dz=0.592.", "verdict": "supported", "evidence": "Table 9: 'Precision increased by 0.0298, t(33)=3.45, p=0.0016, with a 95% confidence interval of [0.0122, 0.0474] and dz=0.592.'", "note": null}, {"claim": "Macro F1 improved by 0.0153, t(33)=2.78, p=0.0089, with a 95% confidence interval of [0.0041, 0.0265] and dz=0.477.", "verdict": "supported", "evidence": "Table 9: 'Macro F1 improved by 0.0153, t(33)=2.78, p=0.0089, with a 95% confidence interval of [0.0041, 0.0265] and dz=0.477.'", "note": null}, {"claim": "Recall decreased slightly by 0.0029, but the difference was not statistically significant, t(33)=-0.41, p=0.6845.", "verdict": "supported", "evidence": "Table 9: 'Recall decreased slightly by 0.0029, but the difference was not statistically significant, t(33)=-0.41, p=0.6845, with a 95% confidence interval of [-0.0173, 0.0115] and dz=-0.070.'", "note": null}, {"claim": "Mean temporal consistency was 0.9140, meaning consecutive test windows tended to have the same predicted attention state.", "verdict": "supported", "evidence": "Section 4.8: 'The mean temporal consistency was 0.9140, which means that consecutive test windows tended to have the same predicted attention state.'", "note": null}, {"claim": "Mean state-transition rate was 0.0860, indicating a low level of switching between low-attention and high-attention states.", "verdict": "supported", "evidence": "Section 4.8: 'The corresponding mean state-transition rate was 0.0860, indicating a low level of switching between low-attention and high-attention states.'", "note": null}, {"claim": "Mean Expected Calibration Error was 0.0370, suggesting small difference between predicted confidence and observed classification performance.", "verdict": "supported", "evidence": "Section 4.8: 'The mean ECE was 0.0370, suggesting that the difference between the confidence predicted and the classification performance observed was small.'", "note": null}, {"claim": "Mean repeated-window stability was 0.9020, indicating CDT outputs were stable under repeated-window evaluation.", "verdict": "supported", "evidence": "Section 4.8: 'The mean repeated-window stability was 0.9020, indicating that the CDT outputs were stable under the repeated-window evaluation procedure defined in Section 3.5.'", "note": null}, {"claim": "47.06% of clients demonstrated strong dominance with one attention state predicted at least 75% of the time.", "verdict": "supported", "evidence": "Table 13: 'Sixteen clients, 47.06% of all clients, demonstrated strong dominance, with one attention state being predicted at least 75% of the time.'", "note": null}, {"claim": "41.18% of clients were in moderate dominance category with prevailing attention state between 60% and 74.99%.", "verdict": "supported", "evidence": "Table 13: 'Fourteen clients (41.18%) were in the moderate dominance category (prevailing attention state between 60% and 74.99%).'", "note": null}, {"claim": "11.76% of clients showed balanced distribution with predominant state less than 60%.", "verdict": "supported", "evidence": "Table 13: 'The remaining 4 clients (11.76%) were more evenly distributed with the predominant state being less than 60%.'", "note": null}, {"claim": "The window size was 256 samples with a window step of 128 samples, providing 50% overlap between windows.", "verdict": "supported", "evidence": "Section 3.1: 'A window size of 256 samples is used with a window step of 128 samples, which means that there is 50% overlap between windows in the same partition.'", "note": null}, {"claim": "At 128 Hz, a window of 256 samples covers exactly 2 seconds of brain activity.", "verdict": "supported", "evidence": "Section 3.1: 'At 128 Hz, a window of 256 samples covers exactly 2 seconds of brain activity, which is a reasonable temporal resolution for attention classification.'", "note": null}, {"claim": "The model uses three convolutional layers with filter counts of 32, 64, and 64 and kernel sizes of 7, 5, and 3 respectively.", "verdict": "supported", "evidence": "Section 3.2.1: 'Three such layers are stacked, with filter counts of 32, 64, and 64 and kernel sizes of 7, 5, and 3 respectively.'", "note": null}, {"claim": "Federated learning was conducted over 10 rounds with 5 local epochs per round and full client participation.", "verdict": "supported", "evidence": "Section 3.3: 'There are 10 federated rounds of training. The 34 clients in every round take part and locally train 5 epochs and submit their parameters to the aggregator.'", "note": null}, {"claim": "Personalization involved 3 epochs of fine-tuning on each client's local training data starting from global parameters.", "verdict": "supported", "evidence": "Section 3.4: 'Each client k starts from the global parameters θglobal and minimizes its local training objective for an additional 3 epochs.'", "note": null}, {"claim": "The authors acknowledge that federated learning does not constitute a formal privacy guarantee, as shared parameters could be subject to gradient leakage, membership inference, or reconstruction attacks.", "verdict": "supported", "evidence": "Section 5 (Discussion): 'This does not constitute a formal privacy guarantee, as shared parameters or updates could still be subject to gradient leakage, membership inference, reconstruction attacks or client-level profiling.'", "note": null}]}