{"id":"2399a2ac-26d9-471a-9fdc-53ab37c8f16e","arxiv_id":"2505.09508","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Wearable-derived physiological features predict low-level blast exposure on held-out subjects, with blast events at 0.25 psi linked to detectable changes in eye and body movement signals.","lead":"This study used wearable sensors to track eye movements, gait, and balance in Special Forces instructors during breaching training, and applied machine learning to relate those signals to measured blast exposure. A fused model predicted blast exposure level on held-out subjects, and the authors report physiological changes at blast pressures as low as 0.25 psi.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.25-psi threshold is the maximum over 4 blast metrics × 7 thresholds; without correction for multiple comparisons the headline claim may be a selection artifact.","rationale":"The reader's weakest assumption focuses on unmeasured confounders (fatigue, heat, exertion) explaining the residual blast correlation. That is a real concern, but the more immediately testable and load-bearing problem is that the threshold claim is selected post hoc over 28 correlated metric/threshold combinations: the reported maximum correlation and the R=0.60 held-out value are both products of this selection, so they are inflated relative to any single pre-specified comparison. A permutation test can settle whether the observed maximum is above chance without new data. If it is not, the central 'as low as 0.25 psi' claim in the abstract and conclusion is unsupported, though the wearable-monitoring contribution and the case study remain informative. Because the paper already frames the threshold as candidate in places and the reader's conditional verdict already requires external validation, the appropriate verdict remains CONDITIONAL rather than REJECT; I do not see an internal inconsistency that invalidates the whole study. Hence UNCHANGED from the reader's CONDITIONAL verdict, with the condition sharpened to include a multiple-comparisons-corrected threshold analysis.","tokens_in":13378,"tokens_out":5529,"duration_ms":57464,"concrete_test":"Permutation test of the selection procedure: shuffle the four cumulative blast metric vectors across the 91 sessions (or bootstrap sessions with replacement) while holding the fused risk-score maxima and exposure-duration values fixed; for each resample recompute the maximum Spearman correlation over all 4 metrics × 7 thresholds, exactly as in Figure 4. Repeat at least 1,000 times. If the observed maximum (~0.60) falls below the 95th percentile of the null maximum distribution, the 0.25-psi threshold is indistinguishable from a multiple-comparisons artifact and should not be cited as a dose-response finding without external validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 and Figure 4 derive the central threshold (160 dB SPL = 0.25 PSI) by evaluating four cumulative blast metrics at seven peak-pressure thresholds (28 conditions) and selecting the largest Spearman correlation with the fused risk score. No multiple-comparison correction, confidence interval, or independent replication is reported. With 91 sessions and highly correlated thresholds, the maximum-over-threshold correlation can exceed the single duration baseline even under a true null, so the statement that blast count 'showed the highest maximum correlation overall' does not establish a physiological threshold. The same selection affects the headline held-out R=0.60, since Table 2 reports the 160 dB configuration chosen from this search. The comparison to exposure duration (correlations 0.24–0.36) is not an adequate control: duration is only one confounder, and blast count / cumulative peak pressure are also training-load variables (more explosive events, more exertion, more noise, more heat) that can drive the same physiological changes. The authors acknowledge in Section 4.2 that physiology changes are non-specific and that acute-change-to-injury links are unproven, which weakens but does not resolve this. Thus the 0.25-psi claim is the least secure load-bearing element of the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports a wearable system that records blast overpressure together with eye (EOG), gait, and balance (accelerometry) signals during Special Forces breaching training, and uses these signals to build per-feature risk scores via a GMM-staircase regression trained to discriminate exposure levels. The authors evaluate the fused risk score with leave-one-subject-out cross-validation, report a held-out Spearman correlation of R=0.60 between the fused score and cumulative blast measures, perform a sensor-ablation study, and describe a single case-study subject with deteriorating ANAM reaction-time scores who showed rapid changes in blink-based risk scores after low-level blast events. The central claims are that blast events as low as 0.25 psi (160 dB SPL) are associated with acute physiological changes, that blast exposure is a direct predictor of these changes, and that the wearable approach is a viable complement to sparse neurocognitive assessments in austere environments.","tokens_in":13657,"tokens_out":4659,"duration_ms":51668,"significance":"If the main claims hold, the contribution is practically significant: it would provide a fieldable, individualized, real-time method for monitoring acute physiological responses to low-level blast, addressing a real gap in occupational blast-exposure management. The study has notable strengths: leave-one-subject-out evaluation, a sensor-ablation analysis, a larger dataset than the authors' prior work, and evaluation of multiple cumulative dose metrics. The risk-score construction is transparently a supervised predictive model, and the held-out correlation is a legitimate predictive-validity result rather than an independent discovery. However, the headline 0.25 psi threshold is selected as the maximum over a grid of four dose metrics and seven thresholds, the causal interpretation is under-controlled for time-varying confounders, and the case study is a single post-hoc observation. These issues are load-bearing for the paper's strong threshold and causal claims, although they are fixable by reframing and additional analysis.","major_comments":[{"comment":"The 160 dB SPL / 0.25 psi threshold claim is identified as the maximum Spearman correlation over four cumulative blast metrics and seven peak-pressure thresholds (28 conditions). Because the thresholds are highly correlated and no multiple-comparison correction, confidence interval, or independent replication is reported, the maximum-over-grid correlation can exceed the exposure-duration baseline even under a null model. The paper should either provide corrected inference (e.g., permutation-based threshold-selection p-values), validate the threshold on a held-out set or a second cohort, or explicitly reframe 0.25 psi as a hypothesis-generating observation. As written, the abstract and conclusion do not support the strength of the threshold claim.","section":"Section 3.2, Figure 4"},{"comment":"The argument that blast exposure is a 'direct predictor' because blast-metric correlations exceed exposure-duration correlations is not sufficient. Exposure duration is only one of several confounds that increase with training load; blast count, cumulative peak pressure, cumulative impulse, and LZeq8hr are also proxies for fatigue, heat exposure, physical exertion, noise, and other stressors that can drive the same physiological changes. Section 4.2's acknowledgement that physiology changes are non-specific is in tension with this causal claim. To support the conclusion, the analysis should adjust for session length and other available load markers, or demonstrate that the physiological response tracks blast events in a temporally specific way within sessions.","section":"Section 3.2"},{"comment":"The case-study subject was selected post hoc: the authors first fit ANAM reaction-time trends across 29 subjects, identified four with significant positive slopes, and then searched their physiology data for anomalous changes. One subject is then presented as evidence of individual susceptibility and is featured in the abstract and conclusion. This is a single-subject, post-hoc observation with no pre-specified criterion or correction for multiple comparisons. It should be clearly labeled as an illustrative, hypothesis-generating case and removed from the abstract unless it is independently validated.","section":"Section 3.1"},{"comment":"The headline held-out R=0.60 also inherits the threshold-selection issue: the ablation table reports the 160 dB configuration chosen from the 28-condition search. The reported improvement from adding one EOG channel is therefore assessed at a threshold selected using the full data. The ablation comparison should be reported at a fixed threshold, an average over thresholds, or a threshold selected within the training folds only, so that the sensor-fusion conclusion is not confounded by the same selection effect.","section":"Table 2 and Section 3.3"}],"minor_comments":[{"comment":"The abstract contains a typo: 'electrooculuography' should be 'electrooculography'.","section":"Abstract"},{"comment":"The text states 'As can be see in Figure 3 (d)' and the figure caption contains an incomplete phrase 'indicated by in order in the bottom row'; both need correction.","section":"Section 3.1 and Figure 3"},{"comment":"Equation (1) and the gait-segmentation notation are poorly typeset ('nX', 'σm(t)2'); the path-length summation and variance thresholds should be rendered unambiguously.","section":"Section 2.5"},{"comment":"Equation (2) has unclear notation: the symbol 'q' appears to denote a square root, and the denominator's parentheses should be checked so that 1/w(n) is unambiguously inside the square-root term.","section":"Section 2.7"},{"comment":"The case-study text gives a threshold of 'greater than 155 dB SPL' for counting low-level blast events, whereas the main dose-response analysis identifies 160 dB SPL; the inconsistency should be explained or reconciled.","section":"Section 3.1"},{"comment":"The session counts differ across modalities (91 sessions for balance and blink, 79 for gait, 36 for saccades); the caption and ablation table should state clearly that correlations are computed on different session subsets. Clarify whether the 91-session rows in Table 2 use only sessions with both modalities or all available sessions for that configuration.","section":"Section 2.9 and Table 2"}],"recommendation":"major_revision","confidential_remarks":"The data collection and leave-one-subject-out evaluation are genuine strengths, and the paper's core methodological architecture is defensible. My recommendation is driven by the 0.25 psi threshold being selected from a 28-condition grid without correction, the causal interpretation being under-controlled for confounders, and a single-subject case study appearing in the abstract. These are fixable with reframing and additional analyses, so I do not recommend rejection. I would also note that the manuscript's audience may be more specialized (physiological measurement, military medicine, human performance) than the general eess.SP readership, though the signal-processing contributions are within scope. I have no concerns about the novelty disclosure or citation practices beyond those already stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid engineering report from MIT Lincoln Lab: 91 sessions, 28 instructors, wearable EOG and accelerometry during breaching training, with leave-one-subject-out cross-validation and an informative ablation study. The multimodal fusion result (R = 0.60 for accelerometry + one EOG channel) is credible, and the move to continuous-time change tracking is a real improvement over their earlier session-level analyses. The honest limitations section is a plus. This is genuinely useful for people building real-time blast-exposure monitors and for range commanders who need something better than pre/post cognitive tests.\n\nThe soft spots are concentrated in the headline claims. The 0.25-psi threshold comes from scanning four cumulative blast metrics across seven peak-pressure thresholds (28 conditions) and reporting the maximum Spearman correlation. No multiple-comparison correction or confidence intervals. With 91 sessions and highly correlated thresholds, selecting the max can easily produce an inflated correlation. The held-out R = 0.60 is also reported for the threshold chosen from that same search, so the validation is not fully independent. The comparison to exposure duration (r = 0.24–0.36) is a useful control but not sufficient: blast count and cumulative peak pressure are also training-load variables, so fatigue, heat, and exertion remain plausible confounders. The case study is a single subject selected after seeing the data; it is illustrative, not evidence.\n\nThe authors do not overclaim in the discussion—they explicitly note non-specificity and the unproven acute-to-chronic link—but the abstract and conclusion phrase the threshold as a finding. The stress-test note is right: this claim is the least secure load-bearing element.\n\nWho should read this? People working on wearable blast dosimetry and military health monitoring. It is a credible proof-of-concept for real-time physiology-based risk scoring. It deserves peer review, but a serious referee should ask for a pre-specified or cross-validated threshold selection, error bars on the correlations, and a more careful treatment of confounders. I would not cite the 0.25-psi value in my own work without independent replication.","headline":"A useful field study with a real validation design, but the headline 0.25-psi threshold is a post-hoc selection artifact and should not be cited as a physiological limit.","tokens_in":14184,"tokens_out":923,"would_cite":false,"duration_ms":10833,"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":"Blast exposure as low as 0.25 psi leaves a measurable trace in blinks, gait, and balance.","keywords":["low-level blast","blast overpressure","wearable sensors","electrooculography","gait analysis","balance","dose-response model","brain injury monitoring"],"falsifier":"Collect matched sessions where total time, physical workload, and heat are held constant while the number of blast events above 0.25 psi varies, for example by changing standoff distance or charge size within the same breaching drill. The paper predicts the fused risk score should track blast count even when session duration and exertion are matched; if the score instead tracks duration or workload, the claim that 0.25 psi events directly drive physiology collapses.","tokens_in":13215,"feed_emoji":"🧠","tokens_out":6697,"duration_ms":72004,"temperature":0.7,"pith_summary":"This paper claims that a body-worn sensor kit can continuously track both blast exposure and the body's physiological reactions during military breaching training, and that a machine-learning model fusing eye-blink, gait, and balance signals turns those reactions into a single risk score. When tested on instructors whose data were held out during model training, the fused score predicted each session's cumulative blast exposure with Spearman R = 0.60. The authors further claim that blast events as weak as 0.25 psi peak pressure are associated with measurable physiological changes, well below the current 4 psi single-event guideline. A case-study subject with years-long declining reaction times consistently showed a sharp rise in risk score after roughly a dozen low-level blasts per session. If right, this gives a way to warn individuals mid-training rather than relying only on sparse before-and-after cognitive tests.","feed_headline":"Blinks, gait, and balance reveal low-level blast exposure","feed_subtitle":"A fused wearable risk score predicted cumulative blast dose in held-out trainees (R=0.60), pointing to real-time alerts.","key_machinery":"The central object is the GMM staircase regression model: an ensemble of Gaussian mixture models, each trained to separate higher-exposure sessions from lower-exposure sessions at a different percentile threshold, whose combined log-likelihood ratio becomes the risk score. Feeding it are feature change scores produced by an online z-scoring method with recursive first and second moment estimates (Equations 2 and 3), applied to blink duration, saccade amplitude, gait time-delay-embedding eigenspectra, and balance path length. Fusion is a simple averaging of per-feature risk scores at the session level. This machinery converts continuous, noisy wearable signals into a scalar risk score that can be correlated with cumulative blast metrics across threshold levels.","core_discovery":"On its own terms, the paper establishes that a dose-response model can be built from wearable physiology alone: the 'dose' is a cumulative blast-exposure metric (blast count, cumulative peak pressure, positive impulse, or time-averaged sound level) computed at seven peak-pressure thresholds, and the 'response' is a continuous risk score derived from online z-scored changes in blink duration, saccade amplitude, gait dynamics, and balance path length. At a 160 dB SPL (0.25 psi) event threshold, blast count produced the highest dose-response correlation, and correlations dropped sharply when only stronger events (170 dB SPL and above) were retained. The paper also identifies one instructor whose blink and balance risk scores rose rapidly and consistently after low-level exposure across multiple years, while the subject's exposure was unremarkable compared with peers, supporting the idea that individual susceptibility varies and can be detected physiologically.","pith_inferences":["An immediate extension the paper leaves implicit: since same-day ANAM change scores were not significantly predicted by blast metrics in the 23 available sessions, the fused physiology risk score could be tested as a faster, same-day surrogate for the cognitive changes that ANAM captures only when administered pre- and post-training.","The 0.25 psi result is a group-level correlation, not an individual injury threshold; a conservative reading is that dosimeters capable of recording events below 1 psi are necessary to build meaningful cumulative dose metrics, even if individual injury thresholds differ.","Because the susceptible subject's blink and balance scores rose consistently after roughly 12 low-level events across separate years, a prospective protocol could screen for susceptibility by watching for repeated rapid rises in those scores, rather than waiting for a long-term reaction-time decline.","The sharp drop in correlation when events below 170 dB are excluded suggests that commercial blast gauges with trigger thresholds near 1 psi may systematically miss the exposure events that matter most for physiological response; replicating this with another dosimeter would provide a direct test."],"forward_implications":["If the dose-response model holds, a wearable can flag an individual's elevated physiological response during a training session, enabling intervention before a post-session cognitive test would reveal a problem.","Cumulative blast measures (event count or accumulated peak pressure above 0.25 psi) carry information beyond exposure duration alone, so safety limits based only on time on the range or on a single peak-pressure ceiling are incomplete.","The same physiology can expose individual susceptibility: the case-study subject's risk score rose steeply after about 12 low-level events on multiple training days spanning several years, while his measured exposure was average for the cohort.","Fusing accelerometry with one EOG channel improved prediction at the 160 dB threshold, while adding a second EOG channel did not improve accuracy in the sessions where it was available.","Across the four dose metrics, blast exposure was consistently associated with shorter blink durations, reduced gait complexity, and increased high-frequency movement during low-motion periods."],"supporting_citations":[{"why":"Supplies the accelerometry-based gait and low-movement feature extraction (time-delay embedding eigenspectra and path length) that the current study extends and fuses with EOG.","marker":"[8]"},{"why":"Prior demonstration that eye-tracking features change across overpressure periods; this paper builds on it with online change scoring and multimodal fusion.","marker":"[6]"},{"why":"Describes the body-worn MNOISE dosimeter and the waveform processing used to compute cumulative blast-exposure metrics and thresholds.","marker":"[32]"},{"why":"Provides the recursive online z-scoring change-detection method used to convert raw physiological features into continuous change scores.","marker":"[37]"},{"why":"External evidence that cumulative blast impulse predicts neurobehavioral symptoms, cited to support the expectation that energy-based metrics may gain importance in larger datasets.","marker":"[38]"},{"why":"The current military 4 psi single-exposure guideline against which the paper's 0.25 psi threshold result is explicitly contrasted.","marker":"[14]"},{"why":"End-user evaluation of body-worn blast sensors used to motivate the need for cumulative rather than single-exposure safety limits.","marker":"[17]"}],"fun_headline_variants":["Fused wearable signals estimate blast dose in real time","Eye and gait data flag low-level blast exposure","Blast exposure predicted by fused physiology wearables","Wearable sensors spot subtle blast effects from 0.25 psi","Risk score from blinks and balance tracks blast exposure"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that cumulative blast overpressure, not fatigue, heat, noise, or physical exertion, is what drives the extra physiological change the model picks up beyond exposure duration.","fun_headline_variants_meta":{"raw":{"variants":["Fused wearable signals estimate blast dose in real time","Eye and gait data flag low-level blast exposure","Blast exposure predicted by fused physiology wearables","Wearable sensors spot subtle blast effects from 0.25 psi","Risk score from blinks and balance tracks blast exposure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000543,"raw_usage":{"total_tokens":2600,"prompt_tokens":943,"completion_tokens":1657,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":1579}},"tokens_in":559,"tokens_out":1657,"duration_ms":11356,"temperature":1.0,"reasoning_tokens":1579,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:28:53.376681+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect matched sessions where total time, physical workload, and heat are held constant while the number of blast events above 0.25 psi varies, for example by changing standoff distance or charge size within the same breaching drill. The paper predicts the fused risk score should track blast count even when session duration and exertion are matched; if the score instead tracks duration or workload, the claim that 0.25 psi events directly drive physiology collapses.","supporting_citations":[{"cited_title":"Using body-worn accelerometers to detect physiological changes during periods of blast overpressure exposure","cited_arxiv_id":null,"evidence_quote":"Supplies the accelerometry-based gait and low-movement feature extraction (time-delay embedding eigenspectra and path length) that the current study extends and fuses with EOG."},{"cited_title":"Changes in Eye Tracking Features Across Periods of Overpressure Exposure","cited_arxiv_id":null,"evidence_quote":"Prior demonstration that eye-tracking features change across overpressure periods; this paper builds on it with online change scoring and multimodal fusion."},{"cited_title":"Development and evaluation of a body-worn dosimeter for continuous and impulsive noise","cited_arxiv_id":null,"evidence_quote":"Describes the body-worn MNOISE dosimeter and the waveform processing used to compute cumulative blast-exposure metrics and thresholds."},{"cited_title":"On the real-time prevention and monitoring of exertional heat illness in military personnel","cited_arxiv_id":null,"evidence_quote":"Provides the recursive online z-scoring change-detection method used to convert raw physiological features into continuous change scores."},{"cited_title":"Cumulative blast impulse is predictive for changes in chronic neurobehavioral symptoms following low level blast exposure during military training","cited_arxiv_id":null,"evidence_quote":"External evidence that cumulative blast impulse predicts neurobehavioral symptoms, cited to support the expectation that energy-based metrics may gain importance in larger datasets."},{"cited_title":"Department of Defense Requirements for Managing Brain Health Risks from Blast Over- pressure","cited_arxiv_id":null,"evidence_quote":"The current military 4 psi single-exposure guideline against which the paper's 0.25 psi threshold result is explicitly contrasted."},{"cited_title":"An End-User Evaluation of Blast Overpressure and Accelerative Impact Body-Worn Sensors","cited_arxiv_id":null,"evidence_quote":"End-user evaluation of body-worn blast sensors used to motivate the need for cumulative rather than single-exposure safety limits."}],"review_version":1}