{"id":"967155c8-d773-4d45-828b-262c1937a3a7","arxiv_id":"1908.05133","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Wristband electrodermal signals classified video-labeled high- versus low-risk construction activities with 76.9% accuracy in eight workers, without measuring workers' own perceived risk.","lead":"This paper tests whether wristband sweat and electrical skin signals from eight construction workers can reveal how risky they feel during work. A classifier reached 76.9% accuracy in telling researcher-labeled high-risk from low-risk activities, but the labels came from video review, not from the workers themselves.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Labels are researcher-assigned activity risk, not workers' perceived risk; the 76.9% accuracy supports activity classification, not the stated claim.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the video-based activity labels are not a valid proxy for workers' perceived risk. This concern is truly load-bearing because it attacks the construct validity of the entire experiment. Even if the overlapping-window leakage were fixed and the accuracy estimate were honest, the model would still be predicting observer-assigned activity risk, not perceived risk. The paper's own discussion admits the label limitation, which strengthens the concern. The proposed concrete test would directly validate whether EDA can predict self-reported perceived risk, which is the only way to support the central claim. The reader's verdict of CONDITIONAL is appropriate: the data-processing pipeline is credible and reproducible in principle, but major revision is required to either collect self-report ground truth or substantially weaken the claim. Therefore, no change to the reader's verdict is needed.","tokens_in":5835,"tokens_out":3826,"duration_ms":38483,"concrete_test":"Run a follow-up field study with the same wristband and feature pipeline, but collect concurrent self-reports of perceived risk (e.g., experience-sampling prompts every 1–2 minutes or at task boundaries) as ground truth. Train the same KNN classifier on these self-report labels using subject-independent splits with non-overlapping windows, and compare accuracy against chance (50%). If the classifier fails to significantly exceed chance, it would show that the original activity-based labels, not perceived risk, drove the 76.9% accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that EDA features recognize workers' perceived risk, but the ground truth used to train and test the model is not perceived risk: it is a binary activity risk label assigned by two research team members from video (Section 3). No worker self-reports were collected; the paper even cites self-report bias as motivation and then uses observer labels without validation. The 76.9% KNN accuracy therefore only demonstrates that EDA features can distinguish activity categories as perceived by the research team, not the worker's internal risk state. The paper acknowledges this in Section 4: if the subjects perceived high risk by other stimuli that were hard to be captured by video, actually high risk samples can be mislabelled as low risk. Moreover, the label is tied to the activity, not the individual; the same activity may invoke different perceived risk across workers, and EDA can also reflect physical exertion or motion artifacts. Unless the labels are validated against self-reports, the reported accuracy cannot be interpreted as evidence for monitoring perceived risk. The overlapping-window leakage is a separate evaluation flaw, but even fixing it would not resolve the construct validity problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a wristband-based EDA sensing and supervised-learning framework intended to recognize construction workers' perceived risk during ongoing work. EDA signals were collected from eight construction workers on real sites, and two research team members labeled video-recorded activities as high-risk or low-risk. After artifact removal and EDA decomposition into tonic and phasic components, the authors extracted 11 time- and frequency-domain features from 10-second windows with 9-second overlap. Six classifiers were compared; KNN achieved 76.9% accuracy on a randomly undersampled, randomly split 80/20 training/test protocol repeated 20 times. The authors conclude that EDA features have the capability to recognize workers' perceived risk under different activity risk levels.","tokens_in":6060,"tokens_out":4435,"duration_ms":45867,"significance":"If the central claim were supported, this would be a meaningful step toward continuous, objective, noninvasive monitoring of construction workers' perceived risk, addressing a real limitation of survey-based methods. The use of field data from actual construction sites, the focus on EDA as a sympathetically innervated signal, and the systematic comparison of several classifiers are strengths. The paper also clearly describes the data collection and preprocessing pipeline. However, the significance hinges entirely on whether the ground-truth labels actually measure the construct the paper claims to predict, and on whether the reported accuracy reflects genuine generalization. As it stands, the evidence supports only activity-risk classification based on expert video annotation, not the recognition of workers' subjective perceived risk.","major_comments":[{"comment":"The ground-truth labels are not workers' perceived risk; they are risk levels assigned by two research team members from video observation. The abstract and conclusion repeatedly claim that the model recognizes workers' perceived risk, but no self-reports, physiological validation, or any other independent measure of the workers' subjective risk state were collected. The paper even criticizes self-report bias in the introduction, yet uses researcher-assigned activity labels without any validation. Section 4 explicitly acknowledges that subjects may have perceived risk from stimuli not visible in the video, which could mislabel true high-risk samples as low-risk. This is a load-bearing construct-validity problem: the 76.9% accuracy demonstrates that EDA features can distinguish activities categorized as high- or low-risk by the research team, not that they can recognize the workers' internal perceived risk. Without an independent validation of the labels against workers' self-reports or a controlled experimental manipulation, the central claim of the paper is unsupported.","section":"Section 3 (Field Data Collection) and Section 4 (Results and Discussion)"},{"comment":"The 10-second windows with 9-second overlap (i.e., a 1-second moving step) create highly autocorrelated, non-independent samples. The random 80/20 split of these overlapping windows into training and test sets means that windows from the same worker and the same continuous time period can appear in both partitions, which leaks temporally adjacent information and inflates the reported accuracy. The evaluation also does not block by subject, so all eight subjects' windows can appear in both training and test, further overestimating performance. This issue is fixable by using non-overlapping windows and subject-level cross-validation, but even after such a fix, the construct-validity concern in the first major comment remains.","section":"Section 2.2 (Feature Extraction) and Section 2.3 (Classification)"},{"comment":"The classification protocol randomly undersamples the majority class to balance the two classes, then reports accuracy on this artificially balanced test set. Under such a protocol, a trivial classifier that always predicts the minority class would achieve 50% accuracy, and the reported numbers are not comparable to accuracy on the natural class distribution. Moreover, the paper averages accuracy over 20 random repetitions but does not report the variance, standard deviation, or confidence interval across these repetitions, so the stability of the 76.9% figure cannot be assessed. This statistical reporting should be corrected or at least supplemented with per-split variability.","section":"Section 2.3 (Classification) and Table 3"}],"minor_comments":[{"comment":"The keywords line still contains the template placeholder 'Type your keywords here, separated by semicolons' followed by the actual keywords; this placeholder should be removed.","section":"Keywords line"},{"comment":"The algorithm name 'Banging Tree' appears to be a typo for 'Bagging Tree'; please correct it.","section":"Section 2.3 (Classification)"},{"comment":"The phrase 'the prevision and recall' contains a typo: 'prevision' should be 'precision'.","section":"Section 4 (Results and Discussion)"},{"comment":"The figure numbering is inconsistent: Section 2.1 refers to 'Figure 2. Data Pre-processing', while Section 3 refers to a wristband and labeling examples as 'Figure 2' and later as 'Figure 3'; the text and figure captions should be aligned.","section":"Figure numbering"},{"comment":"The paper would benefit from reporting the standard deviation or 95% confidence interval of the accuracy across the 20 random splits, as well as results from subject-independent cross-validation, so readers can gauge the robustness of the comparisons.","section":"General"}],"recommendation":"reject","confidential_remarks":"The central problem is not a statistical detail but the mismatch between the measured construct (researcher-assigned activity risk from video) and the claimed construct (workers' perceived risk). Fixing this would require new data collection, such as collecting workers' self-reports or using a validated physiological or experimental protocol, which is beyond the scope of a revision. The overlapping-window and evaluation issues could be corrected, but they would not salvage the fundamental claim. I would be happy to reconsider a substantially revised manuscript that either collects an appropriate ground-truth measure or carefully reframes the contribution as classification of expert-labeled activity risk, with the analysis redone using non-overlapping windows and subject-level blocking."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Real field EDA data from eight construction workers, six classifiers, and a best accuracy of 76.9%. That benchmark is new, though incremental, and the data processing pipeline (filtering, cvxEDA decomposition, standard features) is credible. The authors also do real on-site data collection and clearly acknowledge in Section 4 that their video-based labeling can miss risk factors outside the camera angle. That honesty deserves credit.\n\nThe problem is load-bearing. The ground truth is not workers' perceived risk; it is a binary activity risk label assigned by two research team members from video. No worker self-reports were collected, and the paper criticizes self-report bias before using observer labels without validation. Therefore the 76.9% accuracy only shows that EDA features can distinguish activity categories as perceived by the research team, not the worker's internal risk state. The paper's own caveat about mislabeling is real but does not resolve the construct validity issue.\n\nThe evaluation also leaks information. Windows of 10 seconds overlap by 9 seconds, and the random 80/20 split does not block by subject, so adjacent windows from the same worker can appear in both training and testing. This inflates accuracy. Fixing it with a subject-independent split would be straightforward. The random undersampling to balance classes is acceptable, though the ratio and its effect are not discussed.\n\nThese issues are not fatal to the idea, but they make the paper a proof of concept for activity-risk classification, not perceived risk monitoring. The fix is conceptually simple: collect self-reported ground truth, or at least validate the observer labels against some independent measure, and report subject-independent performance. Construction safety researchers working on wearable sensing will find the dataset and classifier comparison useful, but mainly as a warning about label validity. I would not cite it in my own work, but I would discuss it in a reading group on physiological sensing because the construct validity issue is instructive. It deserves a serious referee, with the expectation of major revision toward honest claims and cleaner evaluation.","headline":"Real field EDA data and a six-classifier benchmark, but the stated construct (perceived risk) is not what the labels measure; the 76.9% accuracy supports activity classification, not risk perception.","tokens_in":686,"tokens_out":1356,"would_cite":false,"duration_ms":26998,"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":"EDA features from wristband biosensors can recognize construction workers' perceived risk as low or high, with a K-nearest-neighbor classifier reaching 76.9% accuracy.","keywords":["Construction Safety","Perceived Risk","Wearable Sensor","Electrodermal Activity","Supervised Learning","K-Nearest Neighbor","Physiological Monitoring"],"falsifier":"Collect real-time self-reports (for example, randomly timed brief prompts on a worn device) from the same workers, then test whether the classifier's high-risk predictions agree with those self-reports; if agreement is no better than chance once activity labels are held fixed, the EDA features are tracking activity category rather than perceived risk. A cheaper check is to re-run the classification under block-wise validation that keeps all windows from the same worker in one fold, which removes the leakage from overlapping windows.","tokens_in":5667,"feed_emoji":"🏗️","tokens_out":5633,"duration_ms":53648,"temperature":0.7,"pith_summary":"This paper tries to establish that a wristband biosensor measuring electrodermal activity (EDA) can tell, from the signal alone, whether a construction worker is currently facing high or low perceived risk. Eight workers wore wristbands during their normal field tasks; their activities were video-recorded and labeled as high- or low-risk by two researchers, and the EDA was filtered, decomposed, and turned into 11 features before six classifiers were compared. The best classifier, K-nearest neighbor, reached 76.9 percent accuracy. If the claim holds, it would provide a continuous, objective, noninvasive alternative to survey-based risk-perception assessment, which the paper argues is post-hoc, biased, and intrusive.","feed_headline":"Wristband skin signals flag high-risk construction work at 76.9%","feed_subtitle":"Continuous wristband readings could replace surveys with objective, real-time risk-perception monitoring at work.","key_machinery":"The load-bearing object is the electrodermal activity signal, the skin's electrical conductance driven by the sympathetic nervous system, and its processing pipeline: a high-pass filter and moving average remove artifacts, a convex-optimization decomposition splits the signal into tonic electrodermal level and phasic electrodermal response, and 10-second windows with 9-second overlap yield 11 time- and frequency-domain features. Six supervised classifiers are compared on an undersampled, balanced dataset, and K-nearest neighbor gives the highest accuracy, which the paper takes as evidence that nonparametric classification fits the EDA feature space better than parametric models.","core_discovery":"The paper's central claim is that features computed from wristband-measured electrodermal activity can separate episodes of high perceived risk from low perceived risk during ongoing construction work. Decomposing EDA into tonic (EDL) and phasic (EDR) components, segmenting into 10-second windows, and training a K-nearest-neighbor classifier on eight workers' field data produced 76.9 percent test accuracy, with recall generally exceeding precision. The authors conclude this demonstrates a feasible, noninvasive route to continuous, objective monitoring of perceived risk, replacing survey- and interview-based assessment.","pith_inferences":["Editorial inference: The 76.9 percent figure may partly reflect classification of activity types rather than of perceived states, since the ground truth is a video-based activity label; direct validation against workers' self-reports is needed before treating the model as a perception monitor.","Editorial inference: The nine-second overlap between consecutive 10-second windows means adjacent training and test samples share almost all their signal; a block-wise or per-worker cross-validation would likely give a lower, more honest accuracy estimate.","Editorial inference: Perceived risk is personal and threshold-dependent, so a model calibrated per worker, or conditioned on individual baselines, might outperform the current population-level KNN classifier.","Editorial inference: The framework could be tested in a controlled lab setting where risk level is manipulated and self-reported, which would separate the signal's response to actual hazard from its response to activity-related movement artifacts."],"forward_implications":["If the central claim is correct, perceived risk could be monitored continuously during work, without stopping workers for surveys or interviews.","A wristband-based system could issue real-time alerts when a worker's EDA indicates a high-risk episode, enabling immediate intervention.","The framework could extend beyond EDA by adding other physiological streams such as PPG and skin temperature, which the paper names as future work.","Because EDA is the only sympathetic signal not contaminated by the parasympathetic system, it offers a cleaner physiological channel for arousal-based risk perception than other wearable signals.","The accuracy and recall figures imply that false negatives on high-risk episodes are relatively rare, which matters if the goal is safety alerting."],"supporting_citations":[{"why":"Prior feasibility study that this paper extends, establishing EDA from wearable sensors as a candidate signal for construction risk assessment.","marker":"[11]"},{"why":"Supplies the convex-optimization decomposition of EDA into tonic and phasic components used in preprocessing.","marker":"[15]"},{"why":"Defines artifacts in EDA recording and motivates the filtering steps that clean the field-collected signal.","marker":"[13]"},{"why":"Links physiological signals to sympathetic nervous system arousal, the pathway by which perceived risk is expected to alter EDA.","marker":"[12]"},{"why":"Provides the risk-homeostasis theory connecting perceived risk to safety behavior, the conceptual basis for why monitoring perception matters.","marker":"[8]"},{"why":"Documents the experience-sampling alternative and supports the paper's critique that retrospective surveys suffer from memory-based bias.","marker":"[10]"}],"fun_headline_variants":["Wristband sweat signals predict workers' risk perception at 76.9%","Biosensor wristband reads risk perception in real time, hits 76.9%","Wristband EDA predicts perceived risk on site with 76.9% accuracy","Real-time risk perception from wristband biosignals: 76.9% accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the video-based risk labels assigned by the research team are a valid proxy for what each worker actually perceived, because if that premise fails the accuracy only shows EDA distinguishes two activity categories, not perceived risk; a secondary premise treats overlapping 10-second windows as independent samples in the random split.","fun_headline_variants_meta":{"raw":{"variants":["Wristband sweat signals predict workers' risk perception at 76.9%","Biosensor wristband reads risk perception in real time, hits 76.9%","Wristband EDA predicts perceived risk on site with 76.9% accuracy","Real-time risk perception from wristband biosignals: 76.9% accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000631,"raw_usage":{"total_tokens":2910,"prompt_tokens":938,"completion_tokens":1972,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":1881}},"tokens_in":554,"tokens_out":1972,"duration_ms":13059,"temperature":1.0,"reasoning_tokens":1881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:22:14.681831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect real-time self-reports (for example, randomly timed brief prompts on a worn device) from the same workers, then test whether the classifier's high-risk predictions agree with those self-reports; if agreement is no better than chance once activity labels are held fixed, the EDA features are tracking activity category rather than perceived risk. A cheaper check is to re-run the classification under block-wise validation that keeps all windows from the same worker in one fold, which removes the leakage from overlapping windows.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior feasibility study that this paper extends, establishing EDA from wearable sensors as a candidate signal for construction risk assessment."},{"cited_title":"Greco, G","cited_arxiv_id":null,"evidence_quote":"Supplies the convex-optimization decomposition of EDA into tonic and phasic components used in preprocessing."},{"cited_title":"Boucsein, Electrodermal activity, Springer Science & Business Media2012","cited_arxiv_id":null,"evidence_quote":"Defines artifacts in EDA recording and motivates the filtering steps that clean the field-collected signal."},{"cited_title":"Habibnezhad, J","cited_arxiv_id":null,"evidence_quote":"Links physiological signals to sympathetic nervous system arousal, the pathway by which perceived risk is expected to alter EDA."},{"cited_title":"Wilde, The Theory of Risk Hom eostasis: Implications for Safety and Health, Risk Analysis, 2 (1982) 209 - 225","cited_arxiv_id":null,"evidence_quote":"Provides the risk-homeostasis theory connecting perceived risk to safety behavior, the conceptual basis for why monitoring perception matters."},{"cited_title":"Larson, M","cited_arxiv_id":null,"evidence_quote":"Documents the experience-sampling alternative and supports the paper's critique that retrospective surveys suffer from memory-based bias."}],"review_version":1}