{"id":"0f5b0912-d3e3-483c-9afc-24808c6a83d5","arxiv_id":"2507.08624","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AIRS combines smartphone 3D reconstruction, avatar-based privacy, navigation, and vision-language models to support TKR home rehabilitation; on internal tests it detected errors in 89% of 263 videos and GPT-4 Vision matched ground truth in up to 60% of 15 cases.","lead":"AIRS is a smartphone-based ambient intelligence framework for guiding knee-replacement patients through home exercises, using 3D reconstruction, an anonymizing avatar, path planning, and vision-language feedback. The paper reports an 89% detection rate of exercise errors via frame selection and names GPT-4 Vision as the best VLM for corrective suggestions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 89% error-detection claim rests on same-person DTW alignment; cross-subject use is untested and likely degrades.","rationale":"The reader's weakest_assumption and my concern are identical: the framework's comparison pipeline depends on same-performer alignment, but real deployment involves cross-subject comparisons. This is load-bearing because the headline quantitative claim (89% error detection) and the framework's central value proposition both rest on that alignment. The paper's own text confirms the assumption (Section 3.6.1: 'the player is the same on both videos, so no additional matching is necessary') and the dataset is limited to one expert acting as patient (Section 3.1). This is a correctness risk, not merely an external-validity caveat, because DTW on raw pose parameters does not automatically handle body-shape differences. The proposed concrete test would directly measure the degradation and also evaluate whether the avatar-based normalization (Section 3.2) can fix it. Since the reader's verdict already conditions acceptance on testing with real patients and addressing the same-performer issue, my read does not change the verdict; it reinforces it. The concern is not an internal inconsistency, but a missing validation step that is explicit in the paper's limitations (Section 5.2 mentions testing with patients as future work). Thus, UNCHANGED is appropriate.","tokens_in":802,"tokens_out":1062,"duration_ms":55730,"concrete_test":"Collect a dataset of at least 10 healthy volunteers with varied height, BMI, and sex, each performing the same set of TKR exercises both correctly and with the same predefined errors. Use the therapist's correct video as the reference and each volunteer's error video as the 'home' recording. Run the full Section 3.6.1 pipeline: MeTRAbs/SKEL pose extraction, DTW synchronization, and MSE/MAE-based error-frame selection, with no manual adjustment for body differences. Have two raters, blind to subject identity, judge whether the selected frame actually shows the error. If the detection rate falls meaningfully below 89%, or if the selected frames shift to different movement phases, the original claim is an artifact of same-person alignment. Also test whether avatar/BMI-based normalization restores performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DTW-selected frames make 89% of exercise errors visible is established only in a same-performer setup. Section 3.1 states that all videos, including the 'home' error recordings, were executed by the expert physical therapist, and Section 3.6.1 explicitly notes: 'Note that the player is the same on both videos, so no additional matching is necessary.' In the actual AIRS use case, the patient at home is a different person from the therapist in the reference video. DTW alignment operates on joint angles or SKEL pose parameters; it does not normalize for differences in body proportions, segment lengths, scale, or range of motion between subjects. Error-frame selection then uses raw MSE/MAE in these pose spaces, so a patient with different limb lengths will exhibit systematic offsets that are unrelated to exercise errors, potentially shifting the selected frames and reducing the 89% detection rate. The avatar and BMI-based body matching described in Section 3.2 could in principle provide cross-subject normalization, but the evaluation never uses it for matching between different people; it only demonstrates the framework with one expert playing both roles. Without cross-subject validation, the framework's core promise—automatic comparison of a clinical reference to patient home recordings—remains unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents AIRS, an ambient-intelligence framework for home rehabilitation after total knee replacement. The framework combines smartphone-based real-time 3D reconstruction of the environment, user localization and navigation, avatar-based privacy-preserving visualization, and vision-language-model (VLM) feedback. The evaluation is based on 263 videos of 43 exercises, all performed by an expert physical therapist, with error detection via DTW-aligned frames and VLM scoring of correction suggestions. The authors report an 89% error-detection rate for DTW-selected single-camera frames and conclude that GPT-4 Vision substantially outperforms open-source VLMs in matching therapist ground-truth corrections.","tokens_in":11035,"tokens_out":5439,"duration_ms":60128,"significance":"If the framework worked as claimed, it would be a useful integrated proof-of-concept for smartphone-based, AI-assisted home rehabilitation: it combines many components (3D reconstruction, navigation, avatar anonymization, VLM feedback) in one pipeline, includes accessibility considerations, and provides a substantial dataset and code release. The paper's strengths are its breadth of integration, the explicit attention to privacy and AI Act compliance, and the attempt to evaluate both frame-level error visibility and VLM output quality. However, the two headline quantitative claims are not yet established: the 89% detection figure is obtained in a same-performer setting, and the VLM comparison uses model-based judges that overlap with the tested systems. These limitations are central, not peripheral, because the framework's core promise is automatic comparison between a clinical reference video and a patient's home recording.","major_comments":[{"comment":"The 89% error-detection rate does not support the claimed home-use scenario because all 'patient' and 'home' videos were performed by the same expert physical therapist (Section 3.1), and Section 3.6.1 explicitly states that 'the player is the same on both videos, so no additional matching is necessary.' In the actual AIRS workflow, the home user is a different person from the therapist in the reference video; DTW on raw joint-angle or SKEL time series and MSE/MAE-based frame selection do not normalize for body proportions, segment lengths, or joint ranges. The avatar body matching described in Section 3.2 could in principle provide cross-subject normalization, but the evaluation never uses it for matching between different people. A cross-subject experiment, or a clear re-scoping of the 89% claim to a 'same-performer proof-of-concept,' is required before the framework's core promise can be considered supported.","section":"Sections 3.1, 3.6.1, and 4.2"},{"comment":"The VLM feedback results are partly self-referential: GPT-4 Vision is one of the evaluated models and is also the judge that decides whether outputs are semantically equivalent to the ground truth (Section 4.3), which can inflate its own score. The embedding-based similarity metric uses LLaMA3-8B embeddings, another model, and the 15-video subset was deliberately selected for high error counts. Reported percentages (e.g., the 60% hit rate and the 'significant margin' in Table 2) are presented without confidence intervals, inter-annotator agreement, or an independent human baseline. The claim that GPT-4 Vision leads by a significant margin should be re-evaluated with human raters or with a judge model that is not among the compared systems.","section":"Section 4.3, Tables 1 and 2"},{"comment":"The 'ratio of detected errors is as high as 89%' is based on two reviewers who inspected DTW-selected frames, but the paper reports only the raw agreement count (245 agreed, 18 consensus) and does not report inter-rater reliability (e.g., Cohen's kappa), a confidence interval for the 89% proportion, or the annotation criteria used to decide whether an error is visible. Without these details, the headline detection rate is not statistically grounded; the authors should provide the full confusion-style breakdown and a pre-specified annotation protocol.","section":"Section 4.2"},{"comment":"The limitations discussion acknowledges VLM inconsistency and TKR-specific generalization, but it does not address the cross-subject matching problem, which is more fundamental than a future-work item. Moreover, the VLM evaluation lacks a comparison against simpler baselines (e.g., random frame selection or an always-choose-the-most-frequent-error baseline), so it is unclear how much of the reported performance is attributable to DTW-based frame selection and prompt design rather than to the VLMs themselves. Adding such baselines and a human-judged sample would substantially strengthen the central claims.","section":"Section 5.2 and Section 3.6.3"}],"minor_comments":[{"comment":"There are typos such as 'excercise' and 'resepctively'; please proofread the manuscript thoroughly.","section":"Abstract and Section 1"},{"comment":"The 'Conclusion and Outlook' section is empty in the submitted text; either fill it with a substantive conclusion or remove the heading.","section":"Section 6"},{"comment":"The text ends with an isolated 'm' ('Supplementary Materials.m'), which appears to be a stray character from a missing file extension.","section":"Section 3.4.2"},{"comment":"'minimal ellipsis' should be 'minimal ellipse'.","section":"Section 3.4.1"},{"comment":"The caption contains 'an excercise at home' and does not clearly describe what the left and right images show; please revise for clarity.","section":"Figure 3 caption"},{"comment":"The tables are referenced but the captions are sparse, and abbreviations such as SMPL, EL, and BL are not defined in the captions; please add fuller captions and note the sample size in each cell.","section":"Tables 1 and 2"},{"comment":"The paper repeatedly refers to Supplementary Materials for implementation details, but the arXiv version does not include them; this makes several reproducibility-related claims unverifiable. Please provide the supplementary material or clearly state where it can be accessed.","section":"Supplementary Materials"},{"comment":"Reference formatting is inconsistent (e.g., 'et al.' vs. 'et. al.'), and some entries mix arXiv identifiers into the title or URL fields; please unify the bibliography style.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a system-integration proof-of-concept rather than a controlled clinical study. The editors should consider whether the journal's scope expects quantitative validation in realistic conditions; here, the same-performer dataset and model-as-judge evaluation are the main gaps. The empty Section 6 suggests the manuscript is not yet in a polished, camera-ready state. The citation and authorship patterns appear unremarkable, but the paper leans heavily on arXiv preprints for several technical components."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine integration paper, not a clinical validation. What's new is the full AIRS pipeline — RT-3DR, navigation, avatar-based privacy, and VLM feedback — plus a small evaluation of four VLMs on 15 deliberately selected high-error videos. The avatar layer and accessibility features are sensible. Credit where due: the authors openly admit in Section 5.2 that single-camera detection misses subtle errors and that VLM feedback needs therapist oversight. The 263-video dataset is real, and they properly build on their own prior navigation work.\n\nThe soft spots are the ones you flagged, and the stress-test note lands. The strongest quantitative claim, 89% error visibility, comes from DTW-selected frames where the same expert performs both the clinical and \"home\" takes. The authors even say \"no additional matching is necessary\" (Section 3.6.1). Real patients differ in body proportions, segment lengths, and joint ranges, so a cross-subject test is the first thing a reviewer will ask for. That is untested, and it matters because the framework's core promise is automatic comparison between a therapist's reference and a patient's home recording. This is a load-bearing gap, not a minor caveat.\n\nThe VLM evaluation is weaker still. Fifteen videos is a small sample, and using GPT-4 Vision both as a tested model and as the judge of semantic equality with ground truth (Section 4.3) is circular; the LLaMA3 embeddings add another model into the loop. The authors' own numbers are modest — at most 60% hits — and they say GPT-4 leads, which is fine within that setup but not evidence for real-world feedback quality.\n\nThere are also manuscript-level issues: Section 6 is empty, and the abstract says software components are available while the body says code will be released upon publication. That mismatch hurts reproducibility and is easy to fix.\n\nOverall, the architecture is plausible and the authors are honest about limits. The flaws are addressable: release code, run a small cross-subject pilot, replace the self-judging metric with human raters, and finish the paper. I would send it to review, but the review should treat the current evaluation as a feasibility study rather than a validated system. I would not cite it as evidence yet.","headline":"A genuine systems integration for home rehab, but both headline numbers rest on same-performer data and model-as-judge metrics; worth a conditional review.","tokens_in":649,"tokens_out":2482,"would_cite":false,"duration_ms":44215,"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":"A smartphone-only rehabilitation framework can expose 89% of exercise errors in single-camera video and have GPT-4-class vision models draft corrections for therapist review.","keywords":["ambient intelligence","home rehabilitation","vision-language models","3D reconstruction","dynamic time warping","avatar privacy","total knee replacement","smartphone guidance"],"falsifier":"Run the same DTW frame-selection protocol on recordings in which the home performer is a real patient and the reference video is made by the therapist; if the share of errors visible in the selected frames drops well below 89%, or if VLM semantic agreement with therapist ground truth falls below the reported 60%, the central claim fails.","tokens_in":10637,"feed_emoji":"📱","tokens_out":9938,"duration_ms":102782,"temperature":0.7,"pith_summary":"This paper proposes a smartphone-based support system for home rehabilitation after total knee replacement. The phone reconstructs the living space in 3D, guides the patient to a safe practice spot, records the exercise, replaces the patient's body with a privacy-protecting avatar, and aligns the home motion with a clinic reference video. The load-bearing evidence is two numbers: Dynamic Time Warping makes about 89% of exercise errors visible in single-camera frames, and the best vision-language model reaches about 60% semantic agreement with the therapist's ground-truth corrections under the best prompt combination. The authors argue that this is enough to make AI a useful pre-filter for the therapist, not a replacement for human judgment. If correct, the framework provides a concrete template for privacy-compliant, machine-guided rehabilitation at home.","feed_headline":"Smartphone AI spots 89% of rehab exercise errors","feed_subtitle":"A privacy-preserving avatar pipeline then uses GPT-4-class vision to draft corrections for the therapist.","key_machinery":"The engine of the argument is a four-stage pipeline. First, Dynamic Time Warping aligns the time series of joint angles or SKEL pose parameters between the clinic reference and the home recording, and the frame with the largest mean-square or mean-absolute error is selected as the most erroneous frame for review. Second, a body-matched avatar (SMPL parameters refined by SKEL) replaces the patient's image, preserving biomechanical detail while hiding identity. Third, the selected image and skeleton pairs are fed to a vision-language model through zero-shot Chain-of-Thought prompting, optionally enriched with error lists and body-region hints, to generate correction instructions. Fourth, correctness is measured by embedding cosine similarity and by asking GPT-4 Vision whether the generated instruction means the same thing as the therapist's ground truth. The same avatar, together with a Gaussian-Splatting reconstruction of the home, feeds the placement optimizer and navigation module that position the patient and camera before recording begins.","core_discovery":"The paper's central claim is that a single smartphone running the AIRS pipeline can make home rehabilitation after total knee replacement machine-guidable. The pipeline reconstructs the room, plans a safe exercise area and camera pose, navigates the patient into position, records the exercise, replaces the person with a body-matched SMPL/SKEL avatar to protect privacy, and synchronizes the home video with the clinic reference using Dynamic Time Warping. On the 263-video dataset, the DTW-selected frames made 89% of the documented errors visible to human reviewers, with subtle balance errors as the main misses. In the vision-language study on 15 error videos, GPT-4 Vision outperformed the open-source models by a significant margin on embedding similarity and reached 60% semantic agreement with the therapist's ground-truth corrections in the best prompt configuration, which combined image, SMPL, and body-region context. The authors conclude that such models are not yet reliable enough to advise patients directly, but are useful for flagging errors and drafting corrections for therapist review.","pith_inferences":["A natural extension is to treat the DTW-selected (clinic frame, home error frame) pairs as labeled training data for a smaller on-device vision model, which would remove the cloud dependency and latency of a GPT-4 call.","An independent panel of therapists rating correction adequacy would give a sharper test of the 60% semantic-agreement number than a same-family model judge does.","Once cross-subject skeleton alignment is solved, the same error frames could drive closed-loop therapy adaptation, such as automatically suggesting easier ranges of motion or more repetitions rather than just flagging mistakes."],"forward_implications":["A patient with only a smartphone can be guided to record exercise videos that expose most mistakes to a remote therapist: 89% of the errors in the 263-video dataset are visible in DTW-selected single-camera frames.","Vision-language feedback at the current level, around 60% semantic agreement with therapist ground truth for GPT-4 Vision under the best prompts, is useful for pre-filtering and drafting but not for unsupervised patient advice.","Replacing the patient's image with a biomechanically matched avatar satisfies the privacy requirements of the EU AI Act while still allowing motion analysis.","The room-reconstruction, placement-optimization, and navigation modules transfer to other rehabilitation exercises and to users with visual or hearing impairments.","Undetectable balance errors in single-camera recordings motivate RGBD-equipped smartphones as the next hardware step for home rehabilitation."],"supporting_citations":[{"why":"This supplies the dataset: 263 videos of 43 total-knee-replacement exercises with expert-labeled errors and therapist ground-truth corrections.","marker":"(Baranyi and et. al., 2022)"},{"why":"This provides MeTRAbs, the single-image 3D skeleton extraction used to turn recorded frames into time series for comparison.","marker":"(Sárandi and et al., 2022)"},{"why":"This provides SKEL, which refines SMPL skeletons into biomechanically accurate joint rotations used for error scoring.","marker":"(Keller and et al., 2023)"},{"why":"This provides the SMPL body parametrization that makes the privacy-preserving, body-matched avatars possible.","marker":"(Loper and et al., 2015)"},{"why":"This provides 3D Gaussian Splatting, the real-time room reconstruction method used to model the home environment.","marker":"(Kerbl and et al., 2023)"},{"why":"This supplies the zero-shot Chain-of-Thought prompting technique used to elicit vision-language-model reasoning in the feedback experiments.","marker":"(Wei and et al., 2022)"},{"why":"This earlier smartphone application for total-knee-replacement rehabilitation is the baseline the framework extends with navigation, placement optimization, and VLM feedback.","marker":"(Zhao and et al., 2024b)"},{"why":"This hierarchical localization method is extended with a streaming image matcher to give real-time user localization inside the reconstructed home.","marker":"(Sarlin and et al., 2019)"}],"fun_headline_variants":["Smartphone AI flags 89% of home rehab errors","Phone AI catches 89% of knee rehab mistakes","Privacy avatar + AI spot rehab errors in home","AI vision guides rehab, hides identity, flags errors","Home rehab AI: 89% error detection, avatar privacy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the same person performs both the clinical reference and the home error videos, so skeletons align without cross-subject normalization, a condition that real patients, whose bodies differ from the therapist's, do not satisfy and that the dataset never tests.","fun_headline_variants_meta":{"raw":{"variants":["Smartphone AI flags 89% of home rehab errors","Phone AI catches 89% of knee rehab mistakes","Privacy avatar + AI spot rehab errors in home","AI vision guides rehab, hides identity, flags errors","Home rehab AI: 89% error detection, avatar privacy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3387,"prompt_tokens":990,"completion_tokens":2397,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2319}},"tokens_in":606,"tokens_out":2397,"duration_ms":18685,"temperature":1.0,"reasoning_tokens":2319,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:13:57.503011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same DTW frame-selection protocol on recordings in which the home performer is a real patient and the reference video is made by the therapist; if the share of errors visible in the selected frames drops well below 89%, or if VLM semantic agreement with therapist ground truth falls below the reported 60%, the central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This supplies the dataset: 263 videos of 43 total-knee-replacement exercises with expert-labeled errors and therapist ground-truth corrections."},{"cited_title":"(2023) From skin to skeleton: T owards biomechanically accurate 3 D digital humans","cited_arxiv_id":null,"evidence_quote":"This provides SKEL, which refines SMPL skeletons into biomechanically accurate joint rotations used for error scoring."},{"cited_title":"(2015) SMPL : A skinned multi-person linear model","cited_arxiv_id":null,"evidence_quote":"This provides the SMPL body parametrization that makes the privacy-preserving, body-matched avatars possible."},{"cited_title":"(2022) Chain-of-thought prompting elicits reasoning in large language models","cited_arxiv_id":null,"evidence_quote":"This supplies the zero-shot Chain-of-Thought prompting technique used to elicit vision-language-model reasoning in the feedback experiments."},{"cited_title":"(2019) From coarse to fine: R obust hierarchical localization at large scale","cited_arxiv_id":null,"evidence_quote":"This hierarchical localization method is extended with a streaming image matcher to give real-time user localization inside the reconstructed home."}],"review_version":1}