{"id":"b23775d6-a1f3-4a31-b8a8-bb02277a10c7","arxiv_id":"2608.13283","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Frozen egocentric video features can discriminate freezing of gait above chance, but underperform IMU-based models, and the evidence for complementary visual information is only qualitative.","lead":"Using cameras worn as glasses, this study tested whether visual context can help detect freezing of gait in people with Parkinson's disease during daily activities at home. The video signal alone performed worse than wearable motion sensors, but the two signals seemed to carry some complementary information, motivating future combined systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of non-redundant ego-vision information rests on qualitative comparisons; a direct quantitative fusion or error-correlation test is missing and is the single load-bearing gap.","rationale":"I agree with the reader's verdict and reasoning. The reader identified exactly the load-bearing gap: the independence claim is supported only by qualitative examples and post hoc explanations, not by a quantitative fusion or error-correlation test. My stress-test pass confirms this is the central issue. The paper itself is honest about the limitation—Section 5 explicitly states 'fusion of IMU and ego-vision was not explored here'—and the supplementary stop-window analysis directly undercuts the contextual-disambiguation motivation. The standalone quantitative comparisons (Table 2, Table 3) are sensible: LOSO protocol, Holm-corrected Wilcoxon tests, multiple window lengths, and stride ablations are all present. The statistical testing on the main comparison is a real strength. However, the headline claim in the abstract and conclusion goes beyond the evidence. The concern is not that ego-vision is useless; the concern is that the specific assertion of non-redundancy is untested. A fusion experiment is the natural and feasible next step because the feature sets already exist and the linear-probe framework is already in place; alternatively, an error-correlation analysis over the 13 LOSO folds would be nearly free to compute. Because the main quantitative results are sound and the overclaim is explicitly flagged by the authors as preliminary, CONDITIONAL remains the right verdict rather than REJECT. The stress-test does not find a more fundamental flaw: no math errors, no protocol leakage (LOSO is correctly applied), no missing baseline that would invalidate the standalone comparisons. The single load-bearing concern is the unsupported inferential leap from 'ego-vision alone is above chance' to 'ego-vision adds independent context to IMU sensing.'","tokens_in":18514,"tokens_out":1843,"duration_ms":16975,"concrete_test":"Run a simple early-fusion experiment on the existing features: concatenate the per-window V-JEPA2 embedding (or its top-k PCA projection) with the Chronos-2 or TCN penultimate features, train the same L2 logistic regression under the same LOSO protocol, and report F1, AUROC, and FPR on the annotated stop windows. If fused performance does not exceed the best unimodal IMU model on AUROC/F1 and does not reduce stop-window FPR below the IMU-only value, the claim of independent context information is not supported. Alternatively, if fusion is deemed out of scope, compute the per-subject error correlation (e.g., overlap of false-positive windows or Spearman correlation of per-subject F1 between V-JEPA2 and the IMU models); near-zero or negative correlation on held-out subjects would support complementarity, whereas strong positive correlation would indicate redundancy.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's title and conclusion promise context-aware multimodal understanding, and the abstract states that ego-video 'may capture FOG-relevant information independent of IMUs' and 'support[s] the use of pretrained ego-video representations to add contextual information' to IMU sensing. The quantitative evidence in Table 2 only compares standalone modalities; no fusion experiment, error-correlation analysis, or paired statistical test on prediction errors is reported. The only quantitative test of the specific contextual benefit—FPR on annotated stop windows (Supplementary Table 3)—shows V-JEPA2 (24.9%) is worse than Chronos-2 (9.7%) and the acc-only TCN (11.0%), directly contradicting the strongest motivating example. The remaining evidence is Section 4.3's qualitative per-subject examples and post hoc explanations (low light for sub014, object interaction for sub015). Because the dataset has only 13 subjects and two subjects contribute no FOG-positive windows, these illustrative cases could easily reflect sampling noise rather than a true complementary signal. The conclusion 'ego-vision provides information that is not simply redundant with inertial sensing' is therefore stronger than the evidence supports; the load-bearing assumption is that the qualitative complementarity generalizes to the cohort, and the paper explicitly acknowledges fusion was not explored (Section 5). This is an overclaim rather than an internal inconsistency, and it is directly addressable by a quantitative test.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies whether frozen representations from pretrained egocentric-video foundation models contain information useful for detecting freezing of gait (FOG) in Parkinson's disease during home-based activities of daily living, with the longer-term motivation of adding contextual information to IMU-based wearables. Using synchronized egocentric video, five IMUs, and expert-annotated FOG labels from 13 participants, the authors evaluate linear probes on several frozen video and time-series foundation models and compare them with fully trained IMU-based temporal convolutional networks under leave-one-subject-out cross-validation. The IMU TCN achieves the best event-detection performance (F1 42.3, AUROC 83.0), while the best video representation, V-JEPA2, reaches F1 32.6 and AUROC 77.2. The paper reports that ego-video features are above chance and, based on qualitative per-subject prediction plots, argues that ego-vision captures FOG-relevant information partly independent of IMUs. The authors are explicit that multimodal fusion was not explored and that vision did not reduce false positives on voluntary-stop windows in the aggregate.","tokens_in":18745,"tokens_out":3307,"duration_ms":37622,"significance":"If the central claim holds, this is a useful first demonstration that frozen ego-vision representations carry clinically meaningful FOG-related signal in a home setting, complementing inertial sensing and motivating future multimodal systems. The paper's strengths include a realistic data-collection protocol, leave-one-subject-out evaluation with per-subject metric averaging, non-parametric significance testing with Holm correction, multiple window-length and stride ablations, and a direct supplementary analysis of the specific clinical failure case of voluntary stopping. The careful quantitative protocol is a solid foundation. However, the load-bearing conclusion that ego-vision provides information independent of IMUs rests on qualitative examples and post hoc explanations rather than on a quantitative test of complementarity, and the only direct stop-window test in the supplement contradicts the motivating example. The paper's contribution is therefore best read as an exploratory study whose main claim needs either additional quantitative support or more cautious framing.","major_comments":[{"comment":"The central claim that ego-vision captures FOG-relevant information independent of IMUs is not supported by a quantitative test. Section 4.3 relies on selected per-subject prediction plots and post hoc explanations (low light for subject 014, object interaction for subject 015), and the abstract and conclusion infer non-redundancy from these examples. No fusion experiment, error-correlation analysis, or paired disagreement statistic is reported. With 13 subjects and two subjects contributing no FOG-positive windows, these illustrative cases may reflect sampling noise. I recommend adding a quantitative complementarity analysis, for example McNemar's test on per-window disagreement between V-JEPA2 and the IMU baselines, a correlation of per-subject error rates, or a simple late-fusion classifier whose performance can be compared with the unimodal baselines. This is the load-bearing gap for the paper's stated motivation.","section":"§4.3 and Abstract"},{"comment":"The motivating example that visual context disambiguates voluntary stopping from freezing is directly contradicted by the reported false-positive rates on annotated stop windows: V-JEPA2 (24.9%) is worse than Chronos-2 (9.7%) and than the acc-only TCN (11.0%). The manuscript acknowledges this in the limitations and conclusion, but the abstract still states that the results support using ego-video to add contextual information to wearable sensing. The abstract and conclusion should be tempered to reflect the fact that the only quantitative context-specific test performed did not show a benefit, unless additional evidence is provided that some other aspect of context (e.g., doorway or hotspot triggers) is captured better by vision.","section":"Supplementary Table 3 and §5/§6"},{"comment":"The claim that V-JEPA2 is competitive with the IMU foundation models is based on descriptive comparisons without a statistical test between the two families. The asterisks in Table 2 indicate only differences from the TCN (acc only) baseline. Given the large per-subject standard deviations, a statement such as 'V-JEPA2 surpassed UniMTS on F1' should either be supported by a paired comparison with correction or be explicitly labeled as a descriptive observation. A Wilcoxon test on the 11 subjects with FOG would be feasible and would strengthen the interpretation of the model rankings.","section":"Table 2 and §3.4"}],"minor_comments":[{"comment":"The Chronos-2 linear probe uses a 23,040-dimensional feature vector after mean-pooling, whereas the video models use 512-1024 dimensions; this asymmetry should be acknowledged as a potential factor in the relative performance of the IMU and vision probes, especially since only about 10,000 training windows are available.","section":"§3.3"},{"comment":"The table header shows 'F1↑ Recall↑ FPR↓ AUPRC↑' and then 'D H DL Overall' without a repeated 'AUROC' label; adding a subheading such as 'AUROC by task' would improve readability.","section":"Table 2"},{"comment":"The sentence 'All three video encoders exceeded the single-frame baseline on AUROC' is slightly ambiguous because VideoMAE-v2, EgoVideo, and V-JEPA2 are compared with DINOv3, while the text later says 'all three'; consider naming the models explicitly.","section":"§4.1"},{"comment":"The statement 'We therefore treat the three lengths as separate operating points rather than directly comparable conditions' is helpful, but the window-size ablation in Table 3 does not include the fully trained TCN baselines; including them would make the comparison more complete.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is methodologically careful and honest about many limitations, but the main selling point in the abstract overstates the evidence for visual-context complementarity. The missing quantitative complementarity analysis is fixable within the manuscript's scope, so I see this as a major-revision rather than a rejection. If the authors can add a fusion or error-correlation experiment, or substantially soften the 'independent information' claim, the paper would be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first real test of egocentric video for FOG detection itself, not just trigger environments, and the evaluation is mostly careful. The problem is that the abstract and conclusion claim ego-vision adds context to IMU sensing, while no fusion or error-correlation experiment is run, and the one quantitative test of the motivating scenario goes the other way. The paper is honest about this in the limitations, but the headline claim outruns the evidence.\n\nWhat's genuinely new: the paper evaluates frozen ego-video foundation-model features (DINOv3, VideoMAE-v2, EgoVideo, V-JEPA2) for detecting freezing episodes, as opposed to the prior Trigger-Finder work on detecting FOG-triggering environments. That is a legitimate gap. The experimental design is sound: leave-one-subject-out evaluation, per-subject metric averaging, Friedman and Wilcoxon tests with Holm correction, plus stride and window-length ablations in the supplement. The features come from public pretrained models and labels from expert annotation, so there is no circularity. The paper also credits its own HERMES framework only as the data-collection tool, which is appropriate.\n\nSoft spots: the dataset is small (13 subjects, two with no FOG windows), heterogeneous, and not released, so the numbers are suggestive rather than definitive. The bigger issue is the gap between the abstract's phrasing and what is tested. There is no fusion experiment and no paired test on prediction errors. Supplementary Table 3, the only quantitative test of voluntary stopping, shows V-JEPA2 (24.9% FPR on stop windows) is worse than Chronos-2 (9.7%) and the acc-only TCN (11.0%), directly contradicting the motivating example. The evidence for complementary info is qualitative and post hoc, which on this sample could easily be noise. To the authors' credit, Section 5 explicitly states fusion was not explored and the stop-window result is reported plainly; the overclaim is in the abstract and conclusion, not in the body.\n\nWho this is for: people working on multimodal FOG detection or context-aware wearable monitoring for Parkinson's disease. It is a solid exploratory study with a directly addressable gap. It deserves a serious referee, not a desk reject, with a request to either add a simple fusion or error-correlation test or soften the conclusion to match the evidence.\n\nRecommendation: engage the work; require the quantitative complementarity test or a softened claim before acceptance.","headline":"First egocentric-video FOG detection study with a careful LOSO protocol, but the abstract's context-add claim outruns the evidence and needs a fusion test or a softer conclusion.","tokens_in":19318,"tokens_out":2220,"would_cite":false,"duration_ms":24034,"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":"Ego-video alone spots freezing of gait, though IMUs stay stronger","keywords":["freezing of gait","Parkinson's disease","egocentric vision","context-aware motion understanding","foundation models","IMU sensing","home-based activities of daily living","multimodal sensing"],"falsifier":"Compute, under the same leave-one-subject-out protocol, the subject-level correlation between the prediction errors of the best ego-vision model and the best IMU model, or train a simple late fusion and check whether F1 and AUROC exceed 42.3 and 83.0 respectively; if the errors are highly correlated or fusion gives no gain, the claimed independence fails. Equally direct: re-score the reported qualitative examples (subjects 013, 006, 008, 015, 014) with independent annotators and verify that the apparent disagreements with IMU models reproduce; if they do not, the evidence for complementarity collapses.","tokens_in":18305,"feed_emoji":"🎥","tokens_out":7487,"duration_ms":73848,"temperature":0.7,"pith_summary":"This paper asks whether a first-person video worn during everyday tasks can add the context that inertial sensors miss when detecting freezing of gait (FOG) in Parkinson's disease, because a stopped IMU trace can mean intentional stopping, object interaction, or a true freeze. Using synchronized ego-video, five body-worn IMUs, and expert annotations from 13 people with Parkinson's recorded at home, the authors show that frozen features from pretrained video models detect FOG above chance, with the best vision representation reaching AUROC 77.2, below the 83.0 of a fully trained IMU network but comparable to the best inertial foundation-model features. Qualitative per-subject analyses suggest the vision signal is not simply redundant with inertial sensing: it sometimes separates true freezes from voluntary stops better, and sometimes fails badly in low light or atypical movement. The paper's conclusion is that egocentric video is worth adding to wearable-sensor systems for context-aware monitoring, and that multimodal fusion is the natural next step.","feed_headline":"Ego-video alone spots freezing of gait, though IMUs stay stronger","feed_subtitle":"First-person video adds independent context cues for detecting freezing at home, pointing to camera-augmented wearables.","key_machinery":"The central machinery is a probe evaluation of frozen representations from pretrained foundation models: ego-video models (a single-frame image model, two third-person video models, and an egocentric video model) and time-series models for the IMU signals, each reduced to one per-window embedding and classified by a regularized logistic-regression linear probe under leave-one-subject-out with per-subject metric averaging. The supporting object is the synchronized home dataset—egocentric smart-glasses video, five body-worn IMUs, and expert FOG annotations in 13 Parkinson's patients performing doorway, hotspot, and daily-life tasks—which lets the authors compare modalities on the same windows. The design isolates the contribution of each modality's pretrained representation with a single shared classifier, and the controlled backbone pair (single-frame versus 16-frame video) isolates the value of temporal context.","core_discovery":"On its own terms, the paper establishes that egocentric video recorded during home activities contains predictive, partially non-redundant information about freezing of gait in Parkinson's disease. In a leave-one-subject-out evaluation over 13 participants, a linear probe on frozen V-JEPA2 features detected FOG with F1 32.6 and AUROC 77.2, while a temporal convolutional network trained from scratch on accelerometer signals reached F1 42.3 and AUROC 83.0; the best inertial foundation model, Chronos-2, scored F1 38.7 and AUROC 82.9. The vision result is clearly above chance and beats the single-frame image baseline (AUROC 53.6), showing temporal ego-video carries FOG-relevant signal. Qualitative side-by-side predictions show episodes where vision agrees with IMU models, some where it separates freezes from voluntary stopping better, and some where it fails (low light, object interaction without visible hands). The authors interpret this as evidence that ego-vision provides context independent of kinematics, motivating multimodal context-aware clinical monitoring, while acknowledging that a direct stop-window test shows vision alone does not reduce false alarms on voluntary stopping relative to inertial baselines.","pith_inferences":["Beyond the paper: the decisive test of complementarity is a fusion experiment the paper leaves open—combining V-JEPA2 ego-video features with IMU signals under the same LOSO protocol should beat 42.3 F1 if the modalities are truly independent; if it does not, the independence claim is not supported.","Beyond the paper: a per-window or per-subject error-correlation analysis between the vision and IMU predictions would quantify how much unique information each modality carries; the paper reports only qualitative agreement, so this calculation would settle whether the apparent independence generalizes.","Beyond the paper: if interpretability methods show that the ego-video signal tracks environmental triggers such as doorways or narrow passages rather than the freezing motion itself, ego-vision may be more valuable for predicting FOG context than for detecting episodes, which would point to different fusion architectures.","Beyond the paper: the stop-window false-alarm result suggests visual context helps in some recordings but not systematically; a conditional analysis per task type (doorway, hotspot, daily life) could reveal exactly where the visual context disambiguates voluntary stopping from true freezing."],"forward_implications":["Ego-video alone detects FOG above chance in home ADL: the strongest frozen vision representation reaches AUROC 77.2, below the IMU-trained TCN's 83.0 but ahead of chance and of a single-frame image baseline.","Egocentric vision is not redundant with inertial sensing: qualitative agreements and disagreements across modalities suggest the visual stream carries context-containing cues, and in individual recordings it separates freezes from voluntary stops better than the IMU models.","Vision alone is not sufficient: low-light recordings and subjects with atypical movement produce large vision errors, and on annotated voluntary-stop windows the best vision model false-alarms more than both the trained TCN and the strongest IMU foundation model.","Frozen foundation-model features with a linear probe reach performance comparable to a TCN trained from scratch on the IMU stream, supporting their use in clinical settings with little labeled data.","Longer windows raise F1 but lower AUROC for every model, meaning temporal alignment matters and comparisons across window lengths should be read as different operating points, not as monotonic improvement."],"supporting_citations":[{"why":"Supplies the strongest ego-video representation (64-frame temporal window) used for the main vision results.","marker":"[3]"},{"why":"Provides the egocentric-video pretrained representation that is competitive with the IMU foundation models.","marker":"[31]"},{"why":"Supplies the motion-specialized IMU features and defines the 10 s temporal context used in the ablation.","marker":"[51]"},{"why":"Supplies the stronger general-purpose IMU foundation baseline (F1 38.7, AUROC 82.9 at 2 s).","marker":"[2]"},{"why":"Defines the updated FOG definition and the video-based annotation standard that produces the expert labels.","marker":"[13]"},{"why":"Documents the inertial ambiguity between FOG and voluntary stopping that motivates adding context.","marker":"[47]"},{"why":"Provides the synchronized multimodal acquisition framework that aligns ego-video, IMU, and annotations.","marker":"[48]"},{"why":"Serves as the single-frame image baseline, isolating the contribution of temporal video context.","marker":"[37]"},{"why":"Supplies the 16-frame third-person video baseline used in the controlled backbone comparison.","marker":"[41]"}],"fun_headline_variants":["Ego-video adds independent FOG cues, IMU stays top","Wearables beat video for FOG, but video adds context","First-person vision reveals freezing cues beyond IMU","Home video shows promise for freezing detection context"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the qualitative cases in which ego-vision differs usefully from inertial sensing reflect a real, cohort-level complementarity; the paper does not report a quantitative test of that complementarity, so if those examples are idiosyncratic to a few subjects, the central motivation for adding ego-vision to wearable FOG monitoring is unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Ego-video adds independent FOG cues, IMU stays top","Wearables beat video for FOG, but video adds context","First-person vision reveals freezing cues beyond IMU","Home video shows promise for freezing detection context"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1556,"prompt_tokens":1047,"completion_tokens":509,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":442}},"tokens_in":663,"tokens_out":509,"duration_ms":5897,"temperature":1.0,"reasoning_tokens":442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:41:14.915826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, under the same leave-one-subject-out protocol, the subject-level correlation between the prediction errors of the best ego-vision model and the best IMU model, or train a simple late fusion and check whether F1 and AUROC exceed 42.3 and 83.0 respectively; if the errors are highly correlated or fusion gives no gain, the claimed independence fails. Equally direct: re-score the reported qualitative examples (subjects 013, 006, 008, 015, 014) with independent annotators and verify that the apparent disagreements with IMU models reproduce; if they do not, the evidence for complementarity collapses.","supporting_citations":[{"cited_title":"Nature Reviews Neurology pp","cited_arxiv_id":null,"evidence_quote":"Defines the updated FOG definition and the video-based annotation standard that produces the expert labels."},{"cited_title":"org/abs/2601.12610","cited_arxiv_id":null,"evidence_quote":"Provides the synchronized multimodal acquisition framework that aligns ego-video, IMU, and annotations."},{"cited_title":"Transactions on Machine Learning Research (2026), https://openreview.net/forum?id=2NlGyqNjns, featured Certification","cited_arxiv_id":null,"evidence_quote":"Serves as the single-frame image baseline, isolating the contribution of temporal video context."}],"review_version":1}