{"id":"a3809fcd-b914-477c-b5a4-b6c1d8096f40","arxiv_id":"2508.18782","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The mapping from physiological signals to subjective arousal shifts within individuals over months, and emotion estimation models must be periodically updated to stay accurate.","lead":"Researchers measured heart rate, skin conductance, and other signals from 24 office workers over two three-month periods and found that the link between these signals and self-reported arousal drifts over time. Models trained on data from one period lost about 5% accuracy when tested months later, suggesting emotion-sensing systems need periodic updates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"5% accuracy drop lacks between-participant statistical support; with n=6 and near-chance AUC, the central temporal-shift claim is not established.","rationale":"The central claim of long-term variability in physiological–arousal relationships rests on the 5% accuracy drop when a model trained on period 1 is tested on period 2. The reader's weakest assumption was the season confound, which is a valid interpretive concern but would still support the broader conclusion that the mapping is non-stationary, since seasonal change is a form of temporal variability. A more fundamental problem is that the drop itself may not be statistically reliable. The paper's reported standard errors are within-participant, averaged over many random splits; they do not quantify the uncertainty across the six participants. With n=6, the effective sample size for generalization is six, not the thousands of repeated-split observations. The near-chance AUC values further weaken the interpretation: a model with AUC around 0.55 has a weak signal, and its accuracy can fluctuate for reasons unrelated to a genuine change in the physiology–arousal relationship, such as minor distribution shifts in feature values or sampling noise. The paper is honest about the small sample and other limitations, and the longitudinal dataset and EBM-based interpretability are valuable. However, the strongest quantitative claim needs between-participant statistical evidence before it can be accepted. The proper verdict remains CONDITIONAL: the paper's direction is plausible, but the headline result must be reanalyzed with participant-level inference. This is why I keep the reader's verdict unchanged.","tokens_in":12105,"tokens_out":4632,"duration_ms":48287,"concrete_test":"For each of the six participants, compute the per-participant accuracy difference between case (a) (train 1st/test 1st) and case (b) (train 1st/test 2nd), using the same 100 splits. Perform a one-sample Wilcoxon signed-rank test on the six differences and construct a bootstrap 95% confidence interval for the mean difference. If the CI includes 0 or p > 0.05, the 5% drop is not statistically significant. Repeat the same procedure for AUC differences. This directly checks the central claim when participant is the unit of analysis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table II reports an accuracy drop from 0.771 (case a) to 0.716 (case b), but the standard errors shown are computed from 100 repeated train/test splits within each participant and then averaged across participants, not across participants as the unit of analysis. The paper never reports a between-participant variance estimate, a paired test, or a confidence interval for the drop. With only six participants, the 5% decrease could be driven by one or two individuals, and the tiny SEs (0.004–0.006) are misleading because they reflect split-sampling noise, not participant-level variability. Moreover, the AUC values (0.548–0.586) are near chance, so the model is only weakly related to arousal; a small drop in a barely-better-than-chance classifier is not robust evidence that the physiology–arousal relationship changed. The authors acknowledge the small sample in Section V, but the Abstract's central claim—'indicating long-term variability'—is supported primarily by this drop. Without between-participant inference, the headline quantitative result is statistically unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether the mapping between wearable physiological features and self-reported arousal is stable within individuals over several months. Data were collected from 24 office workers over two three-month periods (November 2023–February 2024 and April–July 2024); a subset of six participants with sufficient high/low arousal labels was used for within-subject EBM modeling. The main claim is that a model trained on the first period loses about 5% accuracy (0.771 to 0.716; AUC 0.586 to 0.548, Table II) when tested on the second period, indicating long-term temporal variability in physiological–arousal associations. Using the EBM decomposition in Eq. (1), the authors further report that HR is the most stable predictor while EDA_min is the most variable, and they recommend periodic model updating approximately every five months. The paper explicitly lists several limitations: small participant count, two periods only, potential seasonal confounds, and a label distribution biased toward low-arousal states.","tokens_in":12285,"tokens_out":3542,"duration_ms":39559,"significance":"If the central claim is statistically supported, the paper would provide useful empirical evidence that within-individual physiological–arousal relationships are not stationary over months, with practical implications for adaptive emotion-estimation systems. The study's strengths include a longitudinal naturalistic dataset, a transparent interpretable model (EBM), and explicit acknowledgment of many limitations. However, the headline result rests on only six participants and on comparisons that lack between-participant statistical inference, and the design cannot separate temporal drift from seasonal/environmental changes. The feature-stability analysis also has a self-referential validation component. These issues are fixable in revision, but they currently limit the strength of the abstract and conclusion claims.","major_comments":[{"comment":"The principal evidence for temporal shift is the 0.055 accuracy drop and 0.038 AUC drop, but the reported standard errors are computed from 100 repeated train/test splits within each participant and then averaged across the six participants. This reflects split-sampling noise, not participant-level variability. No paired test, confidence interval, or participant-level effect size is reported. With n=6, the average drop could be driven by one or two participants. In addition, the AUC values (0.548–0.586) are close to chance, so the classifier is only weakly related to arousal. Please report per-participant accuracies/AUCs and conduct a between-participant statistical test (e.g., paired bootstrap or mixed-effects model) before claiming a reliable cross-period drop.","section":"Section IV-B, Table II"},{"comment":"The two collection periods are winter (Nov–Feb) and early summer (Apr–Jul), so period is completely confounded with season, temperature, humidity, daylight, and clothing. The authors acknowledge in Section V that 'it remains unclear whether observed patterns follow circadian, seasonal, or irregular trends.' Nevertheless, the Abstract and Conclusion attribute the drop to 'long-term variability' and recommend a five-month update cycle. This is stronger than the evidence supports. To make the temporal-drift claim, the authors should either include environmental covariates, compare same-season periods, or explicitly reframe the conclusion as evidence of cross-period instability without attributing it to a specific temporal mechanism.","section":"Section III-A and Section V"},{"comment":"The temporal-shift analysis using Pearson's r is partly circular. In Eq. (1), fcom and fint are fitted on the same data, and r is computed between fcom and fcom+fint. The paper then validates r 'aligning well with qualitative observations' by visually inspecting those same fitted functions. This does not provide independent evidence that the physiological–arousal relationship has changed; it mostly reflects the model's own parameterization and estimation noise. Please provide a null distribution (e.g., permutation or block bootstrap) for r, or evaluate period-specific models on held-out data, before ranking features by temporal stability.","section":"Section IV-C, Figures 3–5"},{"comment":"Feature selection is performed on the combined dataset across all participants and both data collection periods, before the train/test splits used in Table II. This means that 2nd-period data (including the participants later used as test subjects) is used to select the five features that are then evaluated in the cross-period test. This can leak information and biases the reported performance. Please use nested feature selection inside the training periods, or at least demonstrate that the selected features are stable across periods when selection is performed on the 1st period only.","section":"Section III-C and Section IV-B"}],"minor_comments":[{"comment":"The caption says 'fcom reflects the common contribution of HR' but the figure is about EDA_min. This appears to be a copy-paste error and should be corrected.","section":"Section IV-C, Fig. 4 caption"},{"comment":"The description 'randomly sampled 90 data from both the 1st and 2nd periods' is ambiguous. Was the sampling stratified by arousal class? Given that the dataset is imbalanced toward low-arousal states, the class distribution of the 90 samples per period affects accuracy and AUC. Please report the class balance and consider class-balanced evaluation.","section":"Section III-D / IV-B"},{"comment":"The phrase 'for the six participants per each' is grammatically unclear. Also, the caption should make explicit that the standard errors are within-participant split-sampling errors averaged over participants, not between-participant errors.","section":"Table II caption"},{"comment":"The recommendation to update 'every five months' is presented in the Abstract and Conclusion, but Section V itself states that 'rather than adhering to a fixed update cycle, a more practical approach may involve monitoring model performance.' Please align the wording so the five-month figure is not presented as a validated constant.","section":"Section V"},{"comment":"The x- and y-axis labels are missing or too small. The y-axis of the additive function plots should indicate log-odds contribution, and the x-axis should indicate feature values. This is needed to interpret the shape and shift of the functions.","section":"Figures 3–5"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on Sakimura et al. The genuinely useful thing is the data: 24 office workers, two three-month blocks (winter and spring/early summer), with wrist-worn physiology and in-situ arousal self-reports. That alone is a contribution, and the EBM-based feature stability analysis is a reasonable way to ask which signals shift. The result that HR stays a fairly stable predictor while EDA_min wobbles is a concrete, reportable finding.\n\nThe 5% accuracy drop from training on period 1 to testing on period 2 (Table II, a vs b) is in the right direction, and the contrast with cases (c) and (d) bolsters the temporal-shift interpretation rather than just 'period 2 is harder.' But the stress-test note is right: the standard errors are from repeated splits within participants, not between participants. With n=6, a 5% drop could easily be driven by one or two individuals, and the paper never gives a paired test or CI for the drop. The AUC values (0.548–0.586) are close to chance, so the model is a weak classifier; the drop, while consistent with drift, is not strong evidence on its own. The season confound is real and the authors acknowledge it, but they still frame the result as 'long-term variability' and suggest a five-month retraining cycle, which is not supported by two time points.\n\nThe r-metric analysis is descriptive and the validation is a bit circular—they check r against visual inspection of the same fitted functions that produced r. That's not fatal, but it's weaker than the text implies. I also agree with the reader that the feature-level stability ranking rests on six participants, five of them female, so generalizing is risky.\n\nWhat's actually new: as far as I know, no one has quantified intra-individual physiological-arousal drift over two multi-month naturalistic periods with interpretable models. The paper is honest about its limitations and doesn't oversell the mechanism. The central direction is probably right, but the quantitative claims are provisional.\n\nWho should read it: anyone building deployed affect estimators, especially with wearables; also people doing longitudinal self-report + physiology studies. It deserves a serious referee, but I'd want a between-participant analysis, a paired test, some sensitivity analysis, and a toned-down update recommendation.","headline":"A genuinely new longitudinal dataset and a plausible drift signal, but the headline 5% drop is statistically under-supported; worth refereeing.","tokens_in":12862,"tokens_out":2251,"would_cite":true,"duration_ms":23235,"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":"A model trained on one three-month period loses about 5% accuracy when applied to a later period, showing that the physiology-to-arousal relationship shifts over time within individuals.","keywords":["emotion estimation","physiological signals","temporal variability","arousal","wearable sensors","explainable boosting machines","electrodermal activity","longitudinal study"],"falsifier":"Measure the same individuals in a third period that falls in the same season as the first period (e.g., the following winter). If a model trained on the first period performs as well on this third period as on first-period test data, then the observed drop is seasonal, not long-term drift.","tokens_in":11927,"feed_emoji":"🫀","tokens_out":5192,"duration_ms":53761,"temperature":0.7,"pith_summary":"This paper tries to establish that the relationship between physiological signals and self-reported arousal is not stable over months-long periods within the same person. From 24 office workers (six analyzed in depth) wearing wrist sensors for two three-month periods, the authors find that a model trained on the first period drops about 5% in accuracy when tested on the second. The interpretable model they use shows the shift is not uniform across features: heart rate stays a reliable predictor, while a minimum-skin-conductance feature shifts the most. If correct, this undermines a common static assumption in affective computing and argues for periodic retraining or drift-triggered updates in real-world emotion-estimation systems.","feed_headline":"Physiology-to-arousal mapping drifts within months","feed_subtitle":"A trained model lost 5% accuracy across two periods; heart rate stayed stable, skin conductance did not.","key_machinery":"The central object is the Explainable Boosting Machine (EBM) with a period-interaction decomposition: the log-odds of high arousal are a sum of period-invariant feature functions plus a second-period-specific shift function (scaled by a period dummy). This lets the authors separate common physiology-arousal curves from time-specific deviations per feature. A second quantitative tool is Pearson's correlation coefficient r between the common and total second-period feature functions, used as a per-feature, per-participant stability score.","core_discovery":"Using Explainable Boosting Machines with a period-interaction term, the authors estimate each physiological feature's contribution to binary high/low arousal separately for a common effect and a second-period-specific shift. Across six participants with enough data in both periods, training on the first period and testing on the second lowered accuracy from 0.771 to 0.716 (AUC from 0.586 to 0.548), while training and testing within the second period scored 0.731, showing the drop is not simply because the second period is harder. The additive-function analysis singles out minimum EDA as the least temporally stable feature and heart rate as the most stable. The paper interprets these results","pith_inferences":["Because the two collection windows are in different seasons (winter and early summer), the 5% drop and EDA_min instability could be explained by ambient temperature, humidity, or daylight rather than by a slow within-person drift; a third same-season measurement would separate these.","If environmental drivers dominate, the 'five-month update' rule is a proxy; a more direct implication is that models should carry season or environmental context as features, which the paper notes but does not test.","A stronger test of the drift claim would be to fit the same EBM on monthly or biweekly windows and see whether feature functions change continuously or jump at a seasonal boundary; the paper's two-point design cannot distinguish these."],"forward_implications":["Emotion-estimation systems deployed for months should expect accuracy decay; retraining on recent data (the paper suggests about every five months) can recover part of the lost performance.","Heart rate is a comparatively dependable predictor across time, so models can keep relying on it; the electrodermal minimum is a drift-prone feature that may need recalibration.","Monitoring performance or feature-contribution shifts can serve as an adaptive trigger for updating, instead of a fixed schedule.","Combining first- and second-period training data improved second-period prediction over first-period-only training, indicating that including recent data mitigates drift.","The r-based stability score offers a simple way to flag which physiological features have shifted for which individuals."],"supporting_citations":[{"why":"Supplies the Core Affect Model used to define the binary arousal labels from self-reported emotions.","marker":"[26]"},{"why":"Provides the Explainable Boosting Machine method that the paper uses to estimate period-invariant and period-specific feature contributions.","marker":"[40]"},{"why":"Prior long-term longitudinal multimodal sensing study that this work extends by quantitatively assessing intra-individual temporal variability.","marker":"[23]"},{"why":"Supports the premise that autonomic-emotion associations may not be consistent, motivating the stability analysis.","marker":"[22]"},{"why":"Characterizes real-world affect detection from wearables and the typical short duration of field studies, against which this two-period design is contrasted.","marker":"[19]"}],"fun_headline_variants":["Physiological arousal mapping drifts over months, study finds","Emotion models lose accuracy as physiology-arousal links shift","Heart rate stays stable, skin conductance drifts in emotion prediction","Arousal prediction models need updates every five months","Min EDA shifts, heart rate stable: arousal model drift over months"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The only time comparison is between two three-month blocks in different seasons, so the 5% drop could come from seasonal or environmental changes rather than a genuine long-term shift in how physiology maps to arousal.","fun_headline_variants_meta":{"raw":{"variants":["Physiological arousal mapping drifts over months, study finds","Emotion models lose accuracy as physiology-arousal links shift","Heart rate stays stable, skin conductance drifts in emotion prediction","Arousal prediction models need updates every five months","Min EDA shifts, heart rate stable: arousal model drift over months"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":1955,"prompt_tokens":778,"completion_tokens":1177,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":1092}},"tokens_in":522,"tokens_out":1177,"duration_ms":10718,"temperature":1.0,"reasoning_tokens":1092,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:11:59.822294+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same individuals in a third period that falls in the same season as the first period (e.g., the following winter). If a model trained on the first period performs as well on this third period as on first-period test data, then the observed drop is seasonal, not long-term drift.","supporting_citations":[{"cited_title":"Core affect, prototypical emo- tional episodes, and other things called emotion: dissecting the elephant","cited_arxiv_id":null,"evidence_quote":"Supplies the Core Affect Model used to define the binary arousal labels from self-reported emotions."},{"cited_title":"Measuring affective state: Subject-dependent and -independent prediction based on lon- gitudinal multimodal sensing,","cited_arxiv_id":null,"evidence_quote":"Prior long-term longitudinal multimodal sensing study that this work extends by quantitatively assessing intra-individual temporal variability."},{"cited_title":"Is there consistency and specificity of autonomic changes during emotional episodes? guidancefromtheconceptualacttheoryandpsychophysiology,","cited_arxiv_id":null,"evidence_quote":"Supports the premise that autonomic-emotion associations may not be consistent, motivating the stability analysis."},{"cited_title":"Affect detection from wear- ables in the “real","cited_arxiv_id":null,"evidence_quote":"Characterizes real-world affect detection from wearables and the typical short duration of field studies, against which this two-period design is contrasted."}],"review_version":1}