{"id":"e8237c65-edc7-4812-9b35-6cc1852ed384","arxiv_id":"2502.01224","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new public dataset of 11,640 hours of wearable multimodal recordings from 125 patients with focal epilepsy, including 886 annotated seizures, is released with two seizure detection baselines.","lead":"The authors release SeizeIT2, a new open dataset of 11,640 hours of wearable recordings from 125 patients with focal epilepsy, including behind-the-ear EEG, ECG, EMG, and movement data with 886 annotated seizures. It is intended as a shared testbed for automated seizure detection algorithms, with code and two baseline models provided.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Seizure labels are asserted to be 'carefully aligned' but no alignment protocol or inter-rater validation is reported; a data-side check of the event TSVs is needed before the 886-seizure benchmark claim is fully trustworthy.","rationale":"The reader's weakest assumption (annotations are ground truth) is the same load-bearing concern. The paper is a dataset descriptor; the existence and scale of the dataset are checkable and likely true, but the scientific utility rests on label correctness and temporal alignment. The paper's own categorical totals disagree in a way that makes the label bookkeeping appear less reliable than claimed. This is not an external disagreement with consensus; it is an internal inconsistency and an unverified measurement assumption. The proposed data-side checks are implementable from the public files and would settle the mechanical part of the alignment concern. Since the reader already issued a CONDITIONAL verdict, this stress-test does not move the verdict; it reinforces the same condition with a concrete validation protocol.","tokens_in":8961,"tokens_out":4344,"duration_ms":41011,"concrete_test":"Download ds005873 from OpenNeuro and run a script over all subjects: parse each _events.tsv and the paired EDF headers; assert (1) exactly 886 seizure events across the dataset, (2) every seizure has onset >= 0 and onset + duration <= the EDF duration in the eeg/ecg/emg/mov files, (3) onset/duration values are representable at the native sample rate (250 Hz for eeg/ecg/emg, 25 Hz for mov) rather than truncated to integer seconds, and (4) no overlapping events in the same file. If any check fails on more than 1% of events, the 'carefully aligned' statement is falsified; if all pass, the residual concern is clinical event truth, which would require a small re-annotation sample by a second epileptologist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SeizeIT2's central value is as a public ground-truth resource for focal seizure detection. For that value to be real, each of the 886 seizure events must be a true clinical event and must be correctly time-aligned to the wearable streams. The paper states this alignment was done 'carefully' and that seizures not visible in SD data are still true, but it gives no alignment protocol, no inter-rater agreement, no estimate of alignment error, and no independent check. This matters because Table 2 benchmark metrics are computed on these labels: a systematic offset of even 2 seconds would put the 2-second bte-EEG training windows mostly on non-seizure data. The text also contains an unresolved count inconsistency: after 317 FA and 393 FIA, the remaining 176 seizures are described as 55 FBTC + 12 unclear awareness + 2 subclinical + 93 unknown = 162, whereas Table 1 lists 55 + 98 + 17 + 6 = 176 for the same categories. That inconsistency does not disprove the 886 total, but it weakens the claim that the annotation bookkeeping is 'careful' enough to support a ground-truth benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SeizeIT2, described as the first open dataset of wearable data recorded in patients with focal epilepsy. The dataset comprises more than 11,000 hours of multimodal wearable recordings (bte-EEG, ECG, EMG, accelerometer and gyroscope) from 125 patients across five European Epilepsy Monitoring Centers, with 886 focal seizures annotated by clinicians on the basis of simultaneous full-scalp vEEG and video. The manuscript also proposes a training/validation split and presents two benchmark seizure detectors, a feature-based SVM and a deep-learning ChronoNet model, with epoch-based and any-overlap evaluation metrics. The data are publicly available on OpenNeuro in BIDS format, and code is provided via a GitHub repository.","tokens_in":9189,"tokens_out":3974,"duration_ms":618643,"significance":"If the dataset and its annotations are trustworthy, SeizeIT2 would fill a clear gap in the epilepsy monitoring literature: there are currently no publicly available wearable multimodal datasets from patients with focal epilepsy, and this release could enable reproducible development of wearable seizure detectors. The strengths are the large size (11,640 hours, 886 seizures, 125 patients), the multicenter design, the BIDS-compliant format, the public repository, and the open code for loading and benchmarking. These features make the resource potentially very valuable to the seizure detection community. However, the current manuscript gives limited evidence for the reliability of the seizure labels, which are the foundation of both the dataset and the benchmark evaluations. The unresolved counting inconsistencies and the missing annotation-alignment protocol are therefore central concerns that need to be addressed before the resource can be used with confidence.","major_comments":[{"comment":"The seizure count bookkeeping is internally inconsistent. Table 1 sums to 886 seizures, and 317 FA + 393 FIA = 710, leaving 176 remaining seizures. However, the text in the Dataset Content section states that the remaining seizures comprise 55 FBTC, 12 with unclear awareness status, 2 subclinical focal seizures, and 93 with unknown or unreported onset, which sums to 162. The table lists instead 55 FBTC, 98 Focal, 17 Subclinical, and 6 Unknown, summing to 176. The discrepancy of 14 and the mismatched category sizes (e.g., 12 versus 98 for unclear awareness) need to be resolved. Since the dataset's value depends on accurate annotation metadata, this inconsistency weakens the claim that the data bookkeeping is reliable.","section":"Dataset Content and Table 1"},{"comment":"The annotation alignment protocol is not described. The paper states that 'the annotations and the wearable data of this dataset were carefully aligned with the full-scalp EEG data' and that seizures not visible in the wearable data are still true seizures, but it does not specify how onsets and offsets were transferred from vEEG/video to the wearable streams, how alignment errors were measured or bounded, whether any inter-rater agreement was assessed, or how seizures with no wearable correlate were handled. Because every event in the _events.tsv files and every benchmark metric in Table 2 depends on these labels, a systematic alignment offset of even a few seconds would degrade the bte-EEG training and evaluation windows. The paper should include a detailed annotation and alignment protocol, along with inter-rater statistics or an equivalent validation of label reliability.","section":"Technical Validation"},{"comment":"The benchmark narrative contradicts the reported numbers. The text states that 'the maximum sensitivity is lower' for ChronoNet and that the SVM has a more suitable sensitivity/false-alarm trade-off, yet Table 2 reports ChronoNet sensitivity 84.2 versus SVM 71.1 at the decision threshold of 0.5. If 'maximum sensitivity' refers to the extreme of the sensitivity/false-alarm curve rather than the operating point in Table 2, this should be stated explicitly with the corresponding values from Figure 5. As written, the comparison is internally inconsistent and could mislead readers who use Table 2 as the reference benchmark.","section":"Seizure detection, Table 2 and following paragraph"}],"minor_comments":[{"comment":"The phrase 'focal-to-bilateral tonic clinic' contains a typo; 'clinic' should be 'clonic' (focal-to-bilateral tonic-clonic).","section":"Dataset Content"},{"comment":"The column header 'Focal' is ambiguous alongside FA, FIA, FBTC, Subclinical, and Unknown; consider renaming it to something like 'Focal, unknown awareness' to align with standard ILAE terminology and with the text's description of 'unclear awareness status'.","section":"Table 1"},{"comment":"The phrase 'first and largest phase 3 clinical study' is unclear: is 'phase 3' a reference to the third iteration of the SeizeIT project, or to a clinical trial phase? Please clarify the terminology.","section":"Background & Summary"},{"comment":"Figure 5 is described as a sensitivity/false-alarm-rate curve, but the axes and the operating points corresponding to Table 2 (threshold 0.5) are not described in the text; adding these annotations would improve interpretability.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The dataset itself appears to be a potentially valuable resource, and the public release in BIDS/OpenNeuro is a genuine contribution. However, the annotation reliability claims are under-supported and the counting inconsistencies in the seizure metadata are concerning for a data descriptor. The paper should be revised to provide a concrete alignment protocol, inter-rater validation, and a corrected and clarified seizure count breakdown. Also, the benchmark description needs to be made consistent with Table 2. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SeizeIT2 is worth knowing about. It's the first open multimodal wearable dataset for focal epilepsy that I'm aware of, and if you work on seizure detection from wearable biosignals, this is the resource the field has been missing. 125 patients, 886 focal seizures, 11,640 hours of bte-EEG, ECG, EMG, ACC, and GYR, stored in BIDS on OpenNeuro, with a fixed train/validation split and two baselines. That's a real contribution, and the public code and data loader make it usable.\n\nThe baseline algorithms themselves aren't new—SVM and ChronoNet are adapted from earlier work by the same group—but that's fine for a dataset paper. The split and the any-overlap evaluation metrics are useful.\n\nThe soft spots are real but fixable. First, the seizure count bookkeeping is sloppy. Table 1 totals 886, and the column sums check out, but the prose breakdown of the non-FA/FIA seizures gives 55 + 12 + 2 + 93 = 162, while the table lists 55 + 98 + 17 + 6 = 176 for the same categories. Either the 'Focal' column means something different from 'unclear awareness status' or the categories were mislabeled; the reader can't tell. The total remains intact, but this inconsistency needs correcting.\n\nSecond, the benchmark text contradicts its own table. The paper says 'Despite the ChronoNet method surpassing the SVM in all metrics associated to the area under the curves, the maximum sensitivity is lower,' but Table 2 shows ChronoNet at 84.2% sensitivity vs SVM at 71.1%. That's not a lower maximum sensitivity. Probably a wording slip, but it shouldn't be in print.\n\nThird, no uncertainty estimates on the metrics. Single-point numbers with no CIs or cross-validation variability, which is common in this literature but unfortunate given the benchmark claim.\n\nThe biggest concern, and the one the stress-test flagged, is the annotation alignment. The paper says the wearable data were 'carefully aligned' with the full-scalp EEG, and that seizures not visible in the wearable data are still true seizures. I don't doubt that. But there is no protocol, no inter-rater agreement, and no estimate of alignment error. Since the benchmarks use 2-second windows, a systematic offset of a few seconds would change the labels substantially. This isn't a reason to reject the dataset, but it is a reason to ask the authors to document the alignment pipeline and, ideally, provide the event files in a way that users can inspect per-seizure. To be clear, the stress-test's arithmetic is right: Table 1's 55+98+17+6 sums to 176, while the prose sums to 162, so this is an internal inconsistency, not a wrong total.\n\nBottom line: this paper deserves a serious referee and will be cited. The resource is solid, the issues are mostly cosmetic plus one documentation gap. Send it to review with a request for a data-side check of the event timestamps and a corrected, clarified seizure count section.","headline":"SeizeIT2 is a genuinely useful open wearable focal epilepsy dataset, but the paper needs to fix count inconsistencies and document the vEEG-to-wearable alignment before the ground truth is fully bankable.","tokens_in":9759,"tokens_out":4541,"would_cite":true,"duration_ms":31249,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper presents SeizeIT2, a public multimodal wearable dataset from 125 patients with focal epilepsy, containing over 11,000 hours of synchronized behind-the-ear EEG, ECG, EMG, and movement data with 886 expert-annotated seizures, plus…","keywords":["focal epilepsy","wearable EEG","seizure detection","multimodal monitoring","behind-the-ear EEG","public benchmark dataset","BIDS format","wearable seizure annotation"],"falsifier":"Independently re-annotate a random sample of 50 of the 886 seizures from the synchronized vEEG and video, recording each seizure's onset time without knowledge of the wearable labels, and compare onset times to the provided annotations; systematic offsets above a few seconds or low inter-rater agreement would demonstrate that the wearable ground truth is not trustworthy.","tokens_in":8778,"feed_emoji":"🧠","tokens_out":7121,"duration_ms":56237,"temperature":0.7,"pith_summary":"SeizeIT2 is presented as the first open dataset of wearable physiological recordings from patients with focal epilepsy, created to give the seizure-detection community a standardized public resource for developing automated detection. It contains more than 11,000 hours of synchronized behind-the-ear EEG, ECG, EMG, and movement data from 125 patients across five European epilepsy monitoring centers, along with 886 expert-annotated focal seizures. The paper argues that no such public multimodal wearable dataset existed before, so algorithms have had to train on hospital-scale EEG or small private recordings, limiting generalization. It provides a fixed training/validation split and two baseline detectors, so future methods can be compared on the same data and metrics. If the dataset is sound, it lowers the barrier to building practical outpatient seizure monitoring and makes results reproducible.","feed_headline":"First open wearable dataset tracks 886 focal seizures in 11,000 hours","feed_subtitle":"Multimodal wearable recordings from 125 patients give seizure-detection AI a standardized public benchmark.","key_machinery":"The carrying mechanism is the curated multimodal dataset itself, organized as synchronized recordings from two small wearable modules: one behind-the-ear module records two channels of bte-EEG at 250 Hz, and a chest module records ECG, EMG, and movement at 25 Hz. Expert annotations from full-scalp vEEG and video are aligned to the wearable traces, so each labeled seizure window carries type, lateralization, localization, and whether it is visible in the wearable channels. The fixed 80/20 patient split, the file layout in the Brain Imaging Data Structure (BIDS) format, and the two baseline pipelines convert the raw recordings into a common evaluation platform.","core_discovery":"The central claim is that a large, multicenter, public dataset of wearable data from patients with focal epilepsy now exists and is usable as a benchmark. The dataset pairs a lightweight behind-the-ear EEG device with ECG, EMG, accelerometer, and gyroscope channels, and lines up every wearable recording with expert vEEG/video-based seizure annotations, including seizure type, lateralization, localization, and visibility. The authors release it in a standard neuroimaging data organization, with a prespecified split into 96 training patients and 29 validation patients, plus feature-based (SVM) and deep-learning (ChronoNet) baselines and standard sensitivity/false-alarm metrics. Their stated purpose is to enable the development and fair comparison of automated focal seizure detection methods for continuous monitoring outside the hospital.","pith_inferences":["The two baselines are generic detectors trained across patients; the paper does not explore per-patient adaptation, so testing whether personalization reduces the reported false-alarm rates is a natural extension.","Because 43% of seizures have unknown lateralization and many have unknown onset lobe, the dataset makes it possible to ask how much of the detection difficulty comes from annotation uncertainty rather than sensor noise, a question the paper leaves open.","The inclusion of subclinical and unknown-onset seizure labels invites exploration of detectors for seizures that lack clear behavioral or scalp correlates, though the paper does not validate that use."],"forward_implications":["Seizure-detection models can be trained and validated on standardized multimodal wearable data instead of hospital-scale EEG, which should make results more reproducible and comparable.","New algorithms can be measured against the reported SVM and ChronoNet baselines, giving the field reference points for sensitivity and false-alarm rate.","The public release makes it possible to study how much ECG, EMG, and movement information adds to bte-EEG alone for focal seizure detection.","Because the split is prespecified, later papers can report directly comparable numbers on the same validation set.","A successful detection system built on this data could support outpatient monitoring, where seizure diaries currently miss more than half of seizures."],"supporting_citations":[{"why":"shows behind-the-ear EEG performs similarly to scalp EEG for focal seizure detection, motivating the wearable modality.","marker":"[16]"},{"why":"provides the feature-based SVM pipeline that the paper adapts as its first benchmark.","marker":"[17]"},{"why":"reports clinical validation, diagnostic yield, and user experience of the wearable in focal epilepsy, setting the performance context.","marker":"[27]"},{"why":"adapts the ChronoNet deep architecture to behind-the-ear EEG, supplying the second benchmark.","marker":"[30]"},{"why":"is the released SeizeIT2 dataset record that the paper describes and evaluates.","marker":"[31]"},{"why":"defines the BIDS organization used to structure and share the dataset.","marker":"[32]"},{"why":"introduces the ChronoNet architecture on which the deep-learning baseline is based.","marker":"[33]"},{"why":"defines the epoch-based and any-overlap evaluation metrics used for the benchmarks.","marker":"[34]"}],"fun_headline_variants":["Open wearable dataset: 886 seizures, 11k hours","First open focal epilepsy wearable dataset with 886 seizures","Focal epilepsy: open wearable dataset, 125 patients, 886 seizures","Public multimodal wearable dataset for focal seizure detection","11k hours of wearable focal epilepsy data now open"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The expert annotations and their alignment to the wearable traces are correct, meaning every labeled seizure is truly a seizure and occurs at the labeled time in the wearable data.","fun_headline_variants_meta":{"raw":{"variants":["Open wearable dataset: 886 seizures, 11k hours","First open focal epilepsy wearable dataset with 886 seizures","Focal epilepsy: open wearable dataset, 125 patients, 886 seizures","Public multimodal wearable dataset for focal seizure detection","11k hours of wearable focal epilepsy data now open"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001013,"raw_usage":{"total_tokens":4235,"prompt_tokens":858,"completion_tokens":3377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":3296}},"tokens_in":474,"tokens_out":3377,"duration_ms":19927,"temperature":1.0,"reasoning_tokens":3296,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:59:16.984390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Independently re-annotate a random sample of 50 of the 886 seizures from the synchronized vEEG and video, recording each seizure's onset time without knowledge of the wearable labels, and compare onset times to the provided annotations; systematic offsets above a few seconds or low inter-rater agreement would demonstrate that the wearable ground truth is not trustworthy.","supporting_citations":[{"cited_title":"Comparison between Scalp EEG and Behind-the-Ear EEG for Development of a Wearable Seizure Detection System for Patients with Focal Epilepsy","cited_arxiv_id":null,"evidence_quote":"shows behind-the-ear EEG performs similarly to scalp EEG for focal seizure detection, motivating the wearable modality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the feature-based SVM pipeline that the paper adapts as its first benchmark."},{"cited_title":"& Van Paesschen, W","cited_arxiv_id":null,"evidence_quote":"reports clinical validation, diagnostic yield, and user experience of the wearable in focal epilepsy, setting the performance context."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"adapts the ChronoNet deep architecture to behind-the-ear EEG, supplying the second benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the released SeizeIT2 dataset record that the paper describes and evaluates."},{"cited_title":"ChronoNet: A Deep Recurrent Neural Network for Abnormal EEG Identification","cited_arxiv_id":"1802.00308","evidence_quote":"introduces the ChronoNet architecture on which the deep-learning baseline is based."},{"cited_title":"Prevalence of Epilepsy and Sleep Disturbances in Alzheimer Disease","cited_arxiv_id":null,"evidence_quote":"defines the epoch-based and any-overlap evaluation metrics used for the benchmarks."}],"review_version":1}