{"id":"dc3bf6a2-d8ce-4e09-9c71-e12b5f20ad0b","arxiv_id":"2501.09519","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A YOLO-inspired 1D reformulation jointly detects sleep stages, EEG arousals, and respiratory events in one pass, with strong local but weak external performance.","lead":"The authors reformulate the YOLO object detection algorithm from images to one-dimensional time series, so that a single deep network can simultaneously detect sleep stages, EEG arousals, and respiratory events from polysomnography. They report strong performance on a local test set, but accuracy drops sharply on a small external dataset, showing that cross-hospital generalization remains unresolved.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The S=1, one-bounding-window-per-class encoding in §3.1 means the model cannot detect multiple same-class events within a 30-s epoch; whether this invalidates the central claim depends on whether event metrics are epoch-level rather than event-level.","rationale":"The reader's weakest_assumption already identifies this truncation, and I agree it is the load-bearing issue. The central claim—that one CNN-LSTM can output a 13-component vector containing all three annotation types—is technically true as a vector format, but the clinical claim of simultaneous event detection is only true for at most one event per class per 30-s epoch. The loss-function selection on TS and the 5-recording external set are additional weaknesses, but they do not strike at the architecture's representational capacity as directly. The paper is self-described as preliminary and lists sub-windowing as future work, so the issue is addressable rather than decisive. I recommend leaving the reader's CONDITIONAL verdict unchanged: the condition should explicitly require evidence that event-level recall on un-truncated annotations does not collapse, and that external validity not rest on five recordings whose reference labels are pre-filled baseline annotations from a leg-movement scoring study. If the proposed test shows substantial multi-event windows and poor per-event recall, the claim should be downgraded or substantially qualified.","tokens_in":13251,"tokens_out":4989,"duration_ms":53952,"concrete_test":"Count, on the SHHS TS and HMC-ISA FULL annotations, the fraction of 30-s windows containing at least one event that also contain two or more same-class events (arousals, apneas, hypopneas). If this fraction is non-negligible (>5%), re-evaluate the trained models with per-event matching that counts every annotation (e.g., overlap-based precision/recall) rather than epoch-level kappa/F1. If per-event recall is substantially below the reported epoch-level kappa, the one-per-window encoding materially limits the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §2.2–2.3 the output tensor is S×(B*(3+C)) with S=1 and B=3, and each of the B windows is permanently assigned to one target class (staging, arousal, respiratory). §3.1 then associates every annotation to the single sub-interval containing its centroid, and states that only one annotation per class per sub-interval is considered, with shorter same-class events discarded in favor of the longest. Consequently, the architecture can represent at most one arousal and one respiratory event per 30-s window, and training labels actively delete shorter same-class events. AASM arousals are ≥3 s and apneas/hypopneas ≥10 s, so multiple same-class events can legitimately occur within one 30-s epoch, especially in patients with frequent respiratory events. The reported kappa/F1 values appear to be epoch-level presence/absence scores, not per-event detection rates, and the location regression (x,w) is only evaluated on the single retained event. The claim that the method locates and classifies event occurrences is therefore only demonstrated for a truncated version of the clinical annotation task. The Discussion acknowledges sub-windowing as future work, but the central claim as stated in the abstract is broader than what the encoding can represent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-task deep-learning method, inspired by YOLO object detection, for simultaneous detection of sleep stages, EEG arousals, and respiratory events (apneas and hypopneas) from polysomnographic recordings in a single pass. The method reformulates 2D bounding boxes as 1D bounding windows on a multi-channel time series, with an output tensor per 30-s window containing stage probabilities, an arousal presence/width/center, and respiratory event presence/class/width/center. Experiments use SHHS for training/local evaluation and HMC-ISA for external evaluation, across different channel montages (D=4,6,8) and task assemblies. Local results show average kappa of 0.82 for staging, 0.61 for arousals, and 0.58 for respiratory events, with substantial degradation on the external set (0.45, 0.23, 0.21). The authors conclude that the method supports flexible input configurations and multi-event detection in one pass, with future work on sub-windowing and additional events.","tokens_in":13570,"tokens_out":2885,"duration_ms":29413,"significance":"If the claims are accepted, the paper offers a practical contribution to automated PSG analysis by combining several annotation tasks in one architecture, avoiding the need to concatenate separate algorithms. The release of source code and the use of two independent datasets (local and external) are strengths that support reproducibility and initial generalization assessment. However, the central claim about 'location and classification of event occurrences' is constrained by the output encoding, which limits each 30-s window to at most one arousal and one respiratory event per class. The reported metrics appear to be epoch-level presence/absence scores rather than per-event detection rates, so the true event-level performance remains unclear. The small external test set (5 recordings) also limits the generality of the external validation results.","major_comments":[{"comment":"The output encoding with S=1 and B=3, where each bounding window is permanently assigned to one event class, can represent at most one EEG arousal and one respiratory event per 30-s window. This is confirmed in §3.1, which states that 'we only consider the occurrence of one annotation per class on each sub-interval' and that if two same-class annotations fall in the same sub-interval, the shorter one is discarded. Because AASM arousals are at least 3 s and apneas/hypopneas at least 10 s, multiple same-class events can legitimately occur within one 30-s epoch (especially in patients with frequent respiratory events), so the training labels actively delete clinically relevant events. The reported kappa and F1 values are therefore epoch-level presence/absence scores, not per-event detection rates, and the location regression (x,w) is evaluated only on the single retained event. The abstract's claim of 'location and classification of event occurrences in one pass' is broader than what this encoding can represent. The authors should either re-scope the central claim to epoch-level multi-event classification plus localization of at most one event per class per window, or add sub-windowing experiments (e.g., S>1) to support event-level detection.","section":"§2.3, §3.1"},{"comment":"The numerical comparisons between configurations (e.g., arousal kappa dropping from 0.70 at D=4 to 0.47 at D=8, or respiratory kappa rising from 0.48 to 0.65 with added channels) are presented as performance trends without any confidence intervals or statistical significance tests. Given that the local evaluation uses a single 80/20 split and the external set has only 5 recordings, these differences may not be reliable. The paper should include at least bootstrap confidence intervals across recordings (or per-recording scores) for the key metrics, or the claims of channel-montage effects should be softened accordingly.","section":"Tables 2 and 3"},{"comment":"The external test set HMC-ISA consists of only 5 PSG recordings that were originally selected for a leg-movement scoring study, not for evaluating arousal or respiratory event detection. Using this set as the sole evidence of 'true-generalization effects' is problematic: the small sample size and the task mismatch make it difficult to separate domain shift from sampling variability. The conclusion that external performance degrades significantly (average kappa 0.45/0.23/0.21) is qualitatively important, but the authors should explicitly acknowledge that a 5-recording set is not a representative generalization test and avoid strong quantitative claims about external robustness.","section":"§2.5 and Table 3"}],"minor_comments":[{"comment":"The notation 'M = N ± δ' is unclear because the input tensor size is later given as D × (2δ + N); the manuscript should state that the context margin adds δ samples on each side, so the input length is N + 2δ.","section":"§2.4"},{"comment":"The output vector notation mixes subscripts inconsistently (e.g., c_s, c_a, c_r, p_r, x_r, w_r); a table summarizing the 13 components and their meaning would improve readability.","section":"§2.3"},{"comment":"Decimal commas are used in tables (e.g., '0,78') while the text uses decimal points; please harmonize the format for consistency.","section":"Table 1"},{"comment":"The figures are referenced but not included in the submitted text; the paper should ensure the figures display clearly, with axis labels and legends for the different event-specific MAE curves.","section":"Figures 2 and 3"},{"comment":"The discussion mentions 'significant improvement' and 'significant degradation' in several places without statistical testing; please replace such language with descriptive statements such as 'large observed difference' unless significance tests are provided.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a worthwhile idea and provides open code, but the S=1 encoding is a core limitation that materially narrows the claimed capability. The authors should be asked to either (a) add sub-windowing experiments to support per-event detection, or (b) substantially re-scope the title, abstract, and conclusions to epoch-level multi-label classification with single-event localization. The external validation is too small to support robust generalization claims. No concerns about duplicate publication or citation ethics beyond the normal expectation that the HMC-ISA dataset's intended use be described accurately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible proof-of-concept, and the genuinely new piece is small but real — a 1D reformulation of YOLO-style bounding windows that outputs sleep staging, EEG-arousal, and respiratory-event predictions from one forward pass. Prior multi-task sleep work topped out at two event types; the 13-component output vector here goes a step beyond. The YOLO-to-time-series translation is explained cleanly, the montage/assembly ablations are systematic, and the code ships on GitHub and Zenodo.\n\nThe main thing to flag is the encoding limit, and the stress-test note holds up on reading: §3.1 keeps only the longest annotation per class per 30-s sub-interval, and with S=1 each bounding window is hard-assigned to one target class. The model therefore cannot represent two apneas or two arousals in the same epoch. The reported kappa and F1 values are epoch-level presence/absence scores, and the location regression (x, w) is scored only on the single surviving event. The abstract's phrase 'location and classification of event occurrences' is demonstrated for a truncated version of the clinical task. The Discussion calls this preliminary and lists sub-windowing as future work, so the gap is acknowledged — but the abstract should match the encoding.\n\nOther soft spots, in proportion. Tables 2 and 3 carry no confidence intervals or significance tests; the headline trend that extra channels hurt arousal detection (κ 0.70 → 0.47) rests on roughly 20 test recordings and could be noise. The external set is five HMC-ISA recordings from a leg-movement study — weak, though the drop is honestly reported and informative. Table 1's loss comparison runs on the local test partition, so the three-component loss is effectively chosen on test data — a mild selection issue.\n\nNone of this is fatal. The train/test split is clean, hyperparameters are fixed before testing, there is no circularity, and the method does what it claims at epoch level. Who this is for: people building multi-event PSG pipelines, who will find a useful baseline and a clear statement of the encoding trade-off. It deserves a serious referee. A revision should state the epoch-level framing honestly, add error bars, and either show S>1 or explicitly scope out multiple same-class events per epoch.","headline":"Solid proof-of-concept for single-pass staging plus arousal and respiratory detection, but the one-event-per-class encoding means the 'detection' claim is narrower than the abstract suggests.","tokens_in":14039,"tokens_out":4500,"would_cite":true,"duration_ms":43451,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One CNN-LSTM model can mark sleep stages, arousals, and respiratory events in a single forward pass.","keywords":["sleep staging","EEG arousal detection","respiratory event detection","multi-task learning","single-shot detection","polysomnography","CNN-LSTM","time-series event localization"],"falsifier":"Run the trained model on recordings where expert scorers place two or more events of the same class inside one 30-second window and count how many of the events are recovered; systematic loss of the shorter events would show that the centroid-based single-annotation data-generation rule, not the network, is what limits dense-event detection.","tokens_in":13077,"feed_emoji":"💤","tokens_out":8115,"duration_ms":74504,"temperature":0.7,"pith_summary":"The paper sets out to show that one neural network can carry the full annotation workload of an overnight sleep study in a single pass: assigning a sleep stage to every 30-second epoch and, in the same forward pass, locating and classifying EEG arousals and apneas or hypopneas. The authors reformulate single-shot object detection from computer vision into one-dimensional time series so that each window emits 'bounding windows' with event confidence, position, width, and class. On a held-out part of the training database the method reaches kappa values of 0.82 for sleep staging, 0.61 for arousal detection, and 0.58 for respiratory events; on an unseen external database the same models drop to average kappas of 0.45, 0.23, and 0.21. The contribution is the demonstration that joint detection in one pass is feasible, and that multi-task training, channel montage, and loss design each measurably change what the network learns.","feed_headline":"One pass over sleep signals finds stages, arousals, and apnea","feed_subtitle":"A single neural network outputs sleep stages plus arousal and respiratory-event locations in one forward pass.","key_machinery":"The carrying mechanism is the transfer of grid-cell object detection from two dimensions to one dimension: each time slice plays the role of a grid cell, and 'bounding windows' play the role of bounding boxes, each carrying a confidence score, a temporal center, a width, and a class vector. With one 30-second slice per input and three dedicated bounding windows, the network assigns one window to sleep stages (five class probabilities, no location), one to EEG arousals (presence probability merged with confidence), and one to respiratory events (full confidence-center-width-class vector). The CNN-LSTM structure supplies the features, and a three-component loss lets classification and localization be learned together.","core_discovery":"The central claim is that the one-dimensional reformulation of single-shot object detection works for sleep analysis: given a 30-second window of several physiological channels padded with 60 seconds of context on each side, a CNN-LSTM returns a 13-component vector encoding the five sleep-stage probabilities, an EEG-arousal presence with its temporal center and width, and a respiratory-event presence with its class, center, and width. The network is trained with a three-part loss so that classification and localization errors are minimized together, and at inference overlapping predictions are merged by non-maximal suppression along time while clinical post-processing rules are applied. The local-test numbers are offered as the evidence for feasibility, and the external-test drop is interpreted as dataset shift rather than as a failure of the one-pass formulation.","pith_inferences":["A direct extension of the paper's own parameter space is to raise the number of bounding windows per class so that two or more same-class events inside one 30-second window can be reported; the current one-annotation-per-window rule prevents that.","The same bounding-window encoding could be pointed at other episodic time-series annotations, such as limb movements, seizures, or coughs, provided each event type gets a dedicated bounding window; nothing in the reformulation is sleep-specific.","Because the external evaluation was performed on only five recordings originally selected for a leg-movement scoring study, the reported external kappas should be read as a pilot estimate of generalization rather than a stable population-level number; a larger multi-center external set would tighten that estimate."],"forward_implications":["A full sleep-study annotation could be reduced to one automated forward pass per epoch, producing the hypnogram and event annotations together instead of through separate specialist algorithms.","Joint training helps: adding arousal or respiratory targets to the staging task improves or maintains sleep-stage kappa compared with staging alone, consistent with multi-task learning sharing features across tasks.","Input montage matters asymmetrically: adding saturation and airflow channels raises respiratory-event kappa on local data, while EEG arousal kappa falls as channels are added, suggesting a capacity trade-off inside the fixed model.","The substantial external-data drop means that local test results overstate what a sleep laboratory should expect; models trained on one scoring population will need retraining, more heterogeneous data, or decentralized learning before they generalize."],"supporting_citations":[{"why":"Supplies the single-shot object-detection formulation that the paper reworks from images to time series.","marker":"[26]"},{"why":"Gives the CNN-LSTM architecture that the paper adapts from a sleep-staging-only model to multi-event output.","marker":"[30]"},{"why":"Defines the clinical scoring rules that shape the 13-component output vector and the post-processing filters.","marker":"[3]"},{"why":"Provides the external database of expert-scored recordings on which the generalization test is run.","marker":"[7]"},{"why":"Documents the cohort used to build the training and local-test partitions.","marker":"[32]"}],"fun_headline_variants":["One pass deep learning decodes sleep stages and events","Single network outputs sleep stages, arousals, and apnea locations","Multi-task model detects sleep events and stages simultaneously","One-shot sleep analysis: stages, arousals, apneas from raw signals","Deep learning in one pass: sleep stages plus event locations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach assumes that one annotation per event class per 30-second window is enough: when two arousals or two respiratory events share a window, the shorter is silently discarded, so the model is never trained or tested on the dense-event patterns a sleep detector actually needs to report.","fun_headline_variants_meta":{"raw":{"variants":["One pass deep learning decodes sleep stages and events","Single network outputs sleep stages, arousals, and apnea locations","Multi-task model detects sleep events and stages simultaneously","One-shot sleep analysis: stages, arousals, apneas from raw signals","Deep learning in one pass: sleep stages plus event locations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1350,"prompt_tokens":902,"completion_tokens":448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":518,"tokens_out":448,"duration_ms":4885,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:55:27.939861+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on recordings where expert scorers place two or more events of the same class inside one 30-second window and count how many of the events are recovered; systematic loss of the shorter events would show that the centroid-based single-annotation data-generation rule, not the network, is what limits dense-event detection.","supporting_citations":[{"cited_title":"Inter-database validation of a deep learning approach for automatic sleep scoring,","cited_arxiv_id":null,"evidence_quote":"Gives the CNN-LSTM architecture that the paper adapts from a sleep-staging-only model to multi-event output."},{"cited_title":"The aasm manual for the scoring of sleep and associated events: rules, terminology and technical specifications (version 3),","cited_arxiv_id":null,"evidence_quote":"Defines the clinical scoring rules that shape the 13-component output vector and the post-processing filters."},{"cited_title":"Computer-assisted analysis of polysomnographic recordings improves inter-scorer associated agreement and scoring times,","cited_arxiv_id":null,"evidence_quote":"Provides the external database of expert-scored recordings on which the generalization test is run."},{"cited_title":"The sleep heart health study: Design, rationale, and methods,","cited_arxiv_id":null,"evidence_quote":"Documents the cohort used to build the training and local-test partitions."}],"review_version":1}