{"id":"8fb40e35-0643-4f48-9019-b4347501b5fd","arxiv_id":"2608.12229","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A PPG-plus-SpO2 deep learning framework that predicts apnea burden and converts it to AHI improved out-of-distribution severity classification versus SpO2 alone in a 30-subject external cohort.","lead":"This paper tests whether adding pulse-wave shape features from finger PPG to blood oxygen measurements improves sleep apnea severity scoring when the model is tested on a different patient group. The proposed apnea burden framework showed consistent gains on a 30-subject external stroke cohort.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The OOD gain may be an artifact of the fixed AB-to-AHI conversion: D_mean in Eq. (4) is a single training-set constant, and the nonlinear severity thresholds make the PPG-vs-SpO2 comparison sensitive to it.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point: the fixed D_mean in Eq. (4) sits between the model output and the paper's primary endpoint, severity classification. I agree that this is the most serious threat to the central claim. The concern is not that the conversion is approximate in general; it is that severity boundaries are absolute thresholds on AHI, so a multiplicative rescaling of AB can systematically move subjects across boundaries in a way that differs between predictor combinations. Since all compared models share the same D_mean, one might think the conversion cancels, but it does not: the models produce different AB estimates, and the threshold-crossing pattern depends jointly on those estimates and the conversion factor. The paper's Figure 8 and the reported MAE for AHI_AB do not resolve this, because they are computed with the same D_mean and therefore cannot reveal sensitivity to it. A sensitivity sweep or a conversion using OSASUD-specific durations is a direct, feasible check. If the PPG advantage persists across realistic D_mean values, the central claim survives; if it collapses or reverses, the conclusion should be downgraded. The reader's CONDITIONAL verdict already captures this; my stress test adds specificity but does not move the verdict. I also note the absence of significance tests on the 30-subject OOD cohort, but that is secondary to the conversion issue because even a significant difference could be an artifact of D_mean.","tokens_in":20322,"tokens_out":4454,"duration_ms":45193,"concrete_test":"Recompute OOD macro-F1 for 2CNN-2LSTM with PPGf+SpO2+E(Tp) versus SpO2-only while replacing D_mean in Eq. (4) with (a) the OSASUD mean event duration from its annotations, (b) per-subject mean event durations, and (c) a sweep over 15-35 s. If the PPG advantage shrinks below the seed-level noise (e.g., macro-F1 gap < 5 points) or reverses in any scenario, the headline OOD claim is not established without per-domain conversion calibration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that fusing PPG features with SpO2 improves OOD sleep-apnea severity classification (Section V). The endpoint is not AB regression but AHI-derived severity classes, obtained via Eq. (4): AHI_AB = AB * 3600 / D_mean, with D_mean = 21.39 s estimated from MESA training data. Severity thresholds (5, 15, 30) are applied to AHI_AB, so the mapping from predicted AB to class is nonlinear and depends on D_mean. Event durations vary from roughly 10 s to 2 min, and OSASUD is a stroke cohort with different physiology and annotation conditions; the mean event duration on OSASUD need not equal 21.39 s. Because D_mean multiplies every subject's AB, a mismatch rescales all predicted AHIs and moves subjects across severity boundaries. This does not cancel when comparing feature sets: the comparison is between two imperfect AB predictors, and threshold crossings depend on the scale factor. If, for example, OSASUD's true mean event duration is 30 s, using 21.39 s overestimates AHI by about 40%, systematically shifting moderate/severe boundary crossings. The paper acknowledges this in Section IV.C but only as a limitation, without quantifying sensitivity. With 30 OOD subjects and no significance tests, the reported +16.01 macro-F1 gain is not robustly separated from conversion artifacts.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an apnea burden (AB) guided framework for sleep apnea severity classification from PPG and SpO2 signals. Segment-level AB is predicted by three neural network architectures (2CNN-GRU, 2CNN-2LSTM, and parallel CNNs), aggregated to subject level, and converted to AHI via a fixed mean event duration D_mean = 21.39 s estimated from the MESA training data. The framework is evaluated on MESA as in-distribution data and on OSASUD (30 stroke patients) as out-of-distribution data. The headline claim is that adding PPG-derived features (including a new mean-envelope feature E(Tp)) to SpO2 improves OOD macro-F1 from 28.82% to 44.83% for the 2CNN-2LSTM architecture.","tokens_in":20631,"tokens_out":3644,"duration_ms":33380,"significance":"If the results are robust, the paper offers a practical, wearable-compatible pipeline for sleep apnea severity assessment that explicitly addresses out-of-distribution generalization. The study is reasonably designed, with two external cohorts, multiple architectures, and five repeated runs per configuration. However, the central OOD improvement rests on a small 30-subject cohort and on an AB-to-AHI conversion that may introduce unquantified bias, making the headline gains less certain than the text suggests.","major_comments":[{"comment":"The validation of the AB-to-AHI conversion is circular. Because AB is computed from the same annotated events and D_mean is derived from the training data, the near-identity relationship in Figure 8 is an algebraic consequence of the definitions, not independent empirical validation. The reported MAE between AHI_AB and reference AHI therefore does not validate the conversion for out-of-distribution cohorts. The authors should either present a non-circular justification for the fixed D_mean or quantify how classification outcomes change under plausible D_mean values for OSASUD.","section":"II-E, Equation (4) and Figure 8"},{"comment":"The OOD evaluation uses only 30 subjects and no statistical inference. The standard deviations reported in Table 4 reflect five training runs, not subject-level sampling variability. For example, the 2CNN-2LSTM OOD macro-F1 differences (28.82±6.18 vs. 44.83±6.94) could be compatible with considerable overlap under subject-level bootstrap resampling. I request subject-level bootstrap confidence intervals or a paired significance test (e.g., McNemar or bootstrap for macro-F1) for the key SpO2-only vs. PPG+SpO2 comparisons.","section":"III.C and Table 4"},{"comment":"The sensitivity of the severity classification to the choice of D_mean is not analyzed, and this is load-bearing because the severity boundaries (5, 15, 30) are applied to AHI_AB. The paper acknowledges this in Section IV.C but does not quantify it. I ask the authors to report the mean event duration in the OSASUD test set from the annotated labels, compare it to D_mean = 21.39 s, and re-evaluate the OOD headline metrics with D_mean varied across the plausible range (e.g., ±30%). This will show whether the reported PPG-vs-SpO2 improvements are robust to the conversion constant.","section":"III.A and IV.C"}],"minor_comments":[{"comment":"The description of E(Tp) in Equation (1) is brief; please specify the window length and the exact interpolation procedure more precisely, as the implementation is not fully reproducible from the text.","section":"II-C"},{"comment":"D_mean is first used in Section II-E but its definition (mean duration of all annotated events in the training set) appears only in Section III.A. Please define it earlier and state why a fixed value is clinically acceptable.","section":"III.A and Equation (4)"},{"comment":"The severity labels 'normal' and 'no apnea' are used interchangeably. Please unify the terminology to avoid confusion.","section":"Figure 3 and Table 4"},{"comment":"The sentence 'a zero-phase, fourth-order Butterworth band-pass filter (0.4–9 Hz)' should clarify whether the filter is applied before or after resampling; as written it is ambiguous.","section":"II-B"},{"comment":"The phrase 'We assume that the maximum slope of PPG, Smax, may provide complementary information' is speculative; consider rephrasing as a hypothesis or expectation.","section":"II-C"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is plausible and the experimental setup is reasonable, but the conversion-step sensitivity and the lack of statistical inference on a 30-subject OOD set are load-bearing. The proposed sensitivity analysis and significance testing should be added before publication. I would not reject the paper because the direction of the effect is consistent across two hybrid architectures and the limitations are acknowledged, but the missing quantifications prevent the current version from supporting the strength of the conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a quick look if you work on PPG-based sleep apnea. The core idea—regress apnea burden, then convert to AHI for severity classification—is useful and a bit different from the usual per-event detection or direct classification. The paper compares three architectures and three feature sets on MESA (ID) and OSASUD (OOD), averages five seeds, reports stds, and makes its data-prep scripts available. That is more than many papers in this space do.\n\nThe finding that PPG features help OOD is directionally consistent across the two hybrid recurrent models, and the external cohort is a real domain shift (stroke patients, different hardware). Credit where due: the experimental design is systematic, and the limitations section names the D_mean issue explicitly.\n\nNow the soft spots. The severity endpoint is computed via Eq. (4): AHI_AB = AB*3600/D_mean with D_mean=21.39 s from training data. The thresholds (5/15/30) are applied to that, so a mismatch between D_mean and the true mean event duration in OSASUD rescales every subject's AHI and moves borderline cases across classes. The paper acknowledges this but does not quantify the sensitivity. That matters: with only 30 OOD subjects and no significance tests, the reported +16 F1 gain could be smaller—or larger—if the correct D_mean for OSASUD were used. Figure 8 does not rescue this; it is ground-truth AB converted with the same formula, so it only measures the fixed-duration approximation error, not independent validation. The stress-test note frames this as a potential artifact; I would soften it: since D_mean is shared across all feature sets, the relative comparison is not automatically nullified, but the magnitude of the effect is uncertain until a sensitivity analysis is run.\n\nSecondary: the OOD set is tiny, standard deviations overlap for some key comparisons, and there is no direct AHI-regression baseline to compare against. Training code would help.\n\nBottom line: the direction is plausible and the framework is worth engaging with, but the headline numbers need CI/significance testing and a D_mean sweep before I'd trust the magnitude. Send it to peer review with a request for those additions; it is not a desk reject.","headline":"Solid comparison study with a plausible direction, but the headline OOD gain is not yet separated from a shaky AB-to-AHI conversion and 30-subject statistics.","tokens_in":21151,"tokens_out":4222,"would_cite":true,"duration_ms":40855,"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":"This paper claims that adding PPG pulse-wave morphology features to SpO2, inside a hybrid convolutional-recurrent network that first predicts apnea burden, improves sleep apnea severity classification and markedly improves…","keywords":["apnea burden","out-of-distribution generalization","photoplethysmography","PPG pulse wave features","sleep apnea severity","SpO2","convolutional-recurrent networks","apnea-hypopnea index"],"falsifier":"Re-run the trained models on the OSASUD cohort and recompute AHI from the predicted apnea burden using the cohort's own mean event duration instead of 21.39 s; if several subjects move across the 15 or 30 counts/h severity thresholds, the fixed-duration conversion is carrying the result. As a second check, test the same PPG-plus-SpO2 fusion on another independent external dataset: if the macro-F1 gain relative to SpO2 alone disappears, the claimed out-of-distribution benefit is specific to the 30-subject stroke cohort studied here.","tokens_in":20135,"feed_emoji":"😴","tokens_out":9331,"duration_ms":75361,"temperature":0.7,"pith_summary":"This paper tries to establish that the shape of the PPG pulse wave carries information about sleep apnea that oxygen saturation alone misses, and that adding this information to a neural network makes severity grading more robust when the model meets patients from a different population. The proposed pipeline first predicts apnea burden, the fraction of time spent in respiratory events, as a continuous number, then converts it into the clinical apnea-hypopnea index using a fixed mean event duration from training data, and assigns one of four severity classes. On an external cohort of 30 stroke patients, the best hybrid model's macro-F1 rose from 28.82% with SpO2 alone to 44.83% with six PPG features plus SpO2, while macro-sensitivity rose from 31.90% to 47.62%. This matters because wearable home monitoring already records both PPG and SpO2, and robustness to a new device or population is the main barrier to practical, long-term screening.","feed_headline":"PPG waveform features lift OOD sleep apnea F1 to 44.8","feed_subtitle":"On a 30-patient stroke cohort, six pulse-wave features plus SpO2 nearly double macro-F1 versus SpO2 alone.","key_machinery":"The load-bearing objects are: the apnea burden target, defined as the fraction of a 5-minute segment occupied by apnea or hypopnea events, which turns the problem into regression rather than classification; the conversion $AHI_{AB} = AB \\cdot 3600 / D_{mean}$ with the single training-set constant $D_{mean} = 21.39$ s; and the predictor set, five standard pulse wave features (pulse wave interval $T_p$, peak-to-peak amplitude $A_{pp}$, systolic time $ST$, diastolic time $DT$, and maximum slope $S_{max}$) plus the proposed mean envelope $E(T_p)$, the average of the upper and lower envelopes of the pulse-to-pulse interval time series within a 1-minute epoch. The argument's force is that predicting burden first and thresholding the converted AHI, with hybrid convolutional-recurrent models, transfers better to an unseen stroke-patient cohort than classifying directly from SpO2 alone.","core_discovery":"The paper's central claim, stated in its own conclusion, is that fusing PPG-derived pulse wave features with SpO2 in artificial neural networks improves sleep apnea severity classification and enhances out-of-distribution generalization. The authors argue that a regression-based pipeline that predicts continuous apnea burden per 5-minute segment, aggregates the predictions per subject, and converts them into AHI via $AB \\cdot 3600 / D_{mean}$, then thresholds into normal, mild, moderate, and severe, is a clinically aligned alternative to direct classification. Averaged over five seeds, the 2CNN-2LSTM model using PPG features, SpO2, and the proposed mean-envelope feature $E(T_p)$ achieved the highest overall performance, with OOD macro-accuracy 65.36%, macro-F1 44.83%, and macro-sensitivity 47.62%, versus 56.14%, 28.82%, and 31.90% for SpO2 alone. The authors further claim that the newly proposed PPG features, particularly the mean envelope of the pulse wave interval, add a modest but consistent benefit under distribution shift.","pith_inferences":["The paper leaves untested whether the same PPG-plus-SpO2 gain appears on a second independent dataset with different devices and demographics; that experiment would separate a general out-of-distribution benefit from one specific to the 30-subject stroke cohort.","The paper does not quantify how sensitive its results are to the fixed $D_{mean}$ conversion, although it flags the risk; a sensitivity analysis that perturbs $D_{mean}$ across the observed 10-second to 2-minute duration range would show how many borderline patients change severity class.","If the mean-envelope feature really captures autonomic arousal during apnea, it may transfer to other autonomic-modulation screening tasks, such as nocturnal arrhythmia or stress monitoring; that is an untested implication of the proposed physiology."],"forward_implications":["A wearable that records PPG and SpO2 could assign a patient to one of the four clinical severity categories without full polysomnography, using only the burden-to-AHI conversion.","PPG pulse wave morphology contributes the largest performance gains precisely when the test population differs from training, which is the situation home monitoring devices actually face.","Low-complexity hybrid convolutional-recurrent models (2CNN-GRU and 2CNN-2LSTM) generalize better than a larger multi-scale convolutional model, suggesting that model simplicity is not a barrier to out-of-distribution robustness.","The proposed mean-envelope feature $E(T_p)$ adds a small but consistent benefit for hybrid architectures, supporting the view that minute-scale autonomic modulation is informative for apnea severity.","Regression toward apnea burden, followed by conversion to AHI, aligns model output with clinical severity thresholds while avoiding direct classification decisions at the segment level."],"supporting_citations":[{"why":"Supplies the four-category AHI severity thresholds (normal, mild, moderate, severe) that define the classification endpoint.","marker":"[8]"},{"why":"Prior demonstration that combined PPG and SpO2 signals can distinguish apnea types, motivating the fusion hypothesis.","marker":"[13]"},{"why":"PPG-only apnea sensing study whose high-quality subject selection and multi-dataset validation rationale are adopted.","marker":"[32]"},{"why":"Provides the training, validation, and in-distribution test recordings of PPG and SpO2.","marker":"[34]"},{"why":"Supplies the 30-subject after-stroke cohort used as the out-of-distribution test set.","marker":"[38]"},{"why":"Beat detection algorithm used to extract all five pulse wave feature sequences from PPG.","marker":"[43]"},{"why":"Hybrid convolutional-recurrent architecture for PPG-based blood pressure and heart rate estimation, basis of the 2CNN-2LSTM model.","marker":"[44]"},{"why":"Multi-scale parallel convolution design used for the parallel CNNs architecture.","marker":"[45]"}],"fun_headline_variants":["PPG features + SpO2 hike OOD apnea F1 to 44.8","Apnea-burden-guided framework boosts PPG OOD performance","Low-complexity models with PPG features beat SpO2-only on OOD","New PPG pulse features improve severe apnea grading under shift","Regression-to-AHI approach enhances PPG-based apnea characterization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire AHI estimate and severity label rest on a single constant $D_{mean} = 21.39$ s, the mean apnea/hypopnea event duration computed from the training set and applied to every subject, although event durations vary from about 10 seconds to 2 minutes; if that constant is unrepresentative for a population, every AHI value and every severity boundary shifts.","fun_headline_variants_meta":{"raw":{"variants":["PPG features + SpO2 hike OOD apnea F1 to 44.8","Apnea-burden-guided framework boosts PPG OOD performance","Low-complexity models with PPG features beat SpO2-only on OOD","New PPG pulse features improve severe apnea grading under shift","Regression-to-AHI approach enhances PPG-based apnea characterization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000584,"raw_usage":{"total_tokens":2826,"prompt_tokens":1101,"completion_tokens":1725,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":717,"completion_tokens_details":{"reasoning_tokens":1633}},"tokens_in":717,"tokens_out":1725,"duration_ms":13943,"temperature":1.0,"reasoning_tokens":1633,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:12:23.081379+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the trained models on the OSASUD cohort and recompute AHI from the predicted apnea burden using the cohort's own mean event duration instead of 21.39 s; if several subjects move across the 15 or 30 counts/h severity thresholds, the fixed-duration conversion is carrying the result. As a second check, test the same PPG-plus-SpO2 fusion on another independent external dataset: if the macro-F1 gain relative to SpO2 alone disappears, the claimed out-of-distribution benefit is specific to the 30-subject stroke cohort studied here.","supporting_citations":[{"cited_title":"Sleep-related breathing disorders in adults: recommendations for syn- drome definition and measurement techniques in clinical research,","cited_arxiv_id":null,"evidence_quote":"Supplies the four-category AHI severity thresholds (normal, mild, moderate, severe) that define the classification endpoint."},{"cited_title":"OSASUD: A dataset of stroke unit recordings for the detection of ob- structive sleep apnea syndrome,","cited_arxiv_id":null,"evidence_quote":"Supplies the 30-subject after-stroke cohort used as the out-of-distribution test set."},{"cited_title":"An automatic beat detection algorithm for pressure signals,","cited_arxiv_id":null,"evidence_quote":"Beat detection algorithm used to extract all five pulse wave feature sequences from PPG."}],"review_version":1}