{"id":"bc021cb9-222b-4e37-b87d-4b7880ad4214","arxiv_id":"2505.24507","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A wearable-oriented fall detection plus fall impact time estimation system achieves 82.6% sensitivity, 98.4% specificity on falls, 94.4% specificity on daily activities, and a 160ms RMSE for time of impact on SisFall.","lead":"The paper combines a recurrent neural network with a Kolmogorov-Arnold network to detect falls and estimate the time until impact, testing on the public SisFall dataset. It reports high specificity for daily activities, moderate sensitivity for actual falls, and a time-to-impact error of about 160 milliseconds.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fall-detection metrics are evaluated on random sequence splits with subject demographics as inputs, so the reported TPR/TNR may reflect subject identity rather than generalization to unseen users; a subject-independent split is needed before accepting the central claim.","rationale":"The central claim is an empirical performance claim about a fall detector and a time-to-impact estimator. For the fall detector, the only evaluation that supports 'new users' would require subjects in the test set to be absent from training. The current protocol shuffles sequences and splits 60/20/20, so the same subject's five repetitions can appear on both sides; with age/height/weight/gender provided as inputs, identity is a trivially learnable shortcut. This is not a matter of disagreement with community norms; it is an internal mismatch between the stated plug-and-play use case and the evaluation. The same pattern affects the KAN section (§4.2), where the held-out 'last repetition' still comes from subjects used in training, and §4.1's feature selection appears to use all data before splitting. The 160ms RMSE is further qualified by the authors themselves in §8 as distorted by a ~140ms constant error at small time-to-impact. None of this proves the method is bad; it means the headline numbers cannot yet be read as evidence of generalization. A subject-independent re-evaluation, plus a check on feature-selection leakage, would settle it. The reader's conditional verdict is consistent with this assessment, so no change is recommended.","tokens_in":7427,"tokens_out":5729,"duration_ms":77787,"concrete_test":"Re-run the FDNN evaluation with leave-one-subject-out (or repeated subject-disjoint 70/15/15 splits) on the 1798 fall sequences, keeping the architecture and demographics unchanged. Report mean and per-subject TPR/TNR. Additionally, run the same subject-independent protocol with demographic features removed. If mean TPR drops materially below 82.6%, or if removing demographics changes performance by more than the variability across folds, the reported metrics are inflated by subject identity leakage and the central generalization claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive weakness is in §3: after the fall-only dataset (1798 sequences, 24 subjects) is shuffled, it is divided 60/20/20 into train/validation/test with no subject partitioning. Because each subject contributed five repetitions of each activity, the test set almost certainly contains sequences from subjects also present in training. The network input (§3, Table 3) includes age, height, weight, and gender, i.e., quasi-static subject identifiers. The reported average TPR (82.6%) and TNR (98.4% falls, 94.4% ADL) are therefore consistent with the model recognizing the subject—through demographics or sensor-specific characteristics—rather than detecting fall dynamics on unseen users. This matters because the stated aim is a plug-and-play device for diverse workers. The conclusion treats these numbers as evidence of generalization, but they only certify interpolation to new sequences of known subjects. A secondary leakage path exists in §4.1: correlation/mRMR feature selection appears to be run on the whole dataset before the KAN train/validate/test split, so the test repetition could influence feature choice; this should be checked as well.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-stage wearable system for mitigating workplace falls: a recurrent neural network (FDNN) with LSTM layers that detects an ongoing fall, and a Kolmogorov-Arnold Network (KAN) that estimates the time to impact. The FDNN is trained on 1798 fall sequences from the SisFall dataset and the KAN is trained on five selected inertial features. The authors report an average true positive rate of 82.6% and true negative rate of 98.4% on fall sequences, a true negative rate of 94.4% on activities of daily living, and a root mean square error of approximately 160 ms for time-to-impact estimation. The paper also presents per-subject and per-activity breakdowns and a qualitative example of the KAN output.","tokens_in":7664,"tokens_out":7708,"duration_ms":85083,"significance":"If the reported metrics held under a rigorous, subject-independent evaluation protocol, the proposed combination of an LSTM-based fall detector and a KAN time-to-impact estimator could be a worthwhile contribution to real-time fall prevention for industrial workers. The use of the public SisFall dataset, the per-activity and per-subject result tables, and the explicit treatment of time-to-impact estimation are strengths. However, the current evaluation protocol does not support the central claim of generalization to new users, and the feature selection procedure for the KAN appears to leak information from the test fold. The paper is a brief report rather than a full methods paper; no code or detailed methodology is provided, so independent reproduction would be difficult.","major_comments":[{"comment":"In §3, after shuffling the 1798 fall sequences from 24 subjects, the dataset is split 60/20/20 into train/validation/test without subject partitioning, while the model inputs include age, height, weight, and gender (Table 3). Since each subject performed five repetitions of each activity, the test set almost certainly contains sequences from subjects also present in the training set. The reported average TPR (82.6%) and TNR (98.4%) are therefore consistent with the model recognizing the subject—through demographics or sensor-specific characteristics—rather than detecting fall dynamics on unseen users, which is the stated plug-and-play goal. Please re-evaluate with a subject-independent split (e.g., leave-one-subject-out) and, ideally, with an ablation that excludes the demographic features.","section":"§3"},{"comment":"The correlation and mRMR feature selection in §4.1 is described before the train/validation/test split in §4.2, suggesting that it is applied to the whole dataset. The test repetition (the fifth repetition for each subject) is then used to compute the RMSE that supports the central claim of approximately 160 ms. If feature selection used the test repetition to choose features, the reported RMSE is optimistically biased. Please move feature selection inside the cross-validation loop, or explicitly state and demonstrate that the test fold did not influence feature choice.","section":"§4.1–§4.2"},{"comment":"The paper reports an average RMSE of about 160 ms for time-to-impact estimation, but the Conclusions state that for very small time-of-impact values the estimate is distorted by a constant error of about 140 ms in the presented example. This implies a systematic and large error in the most safety-critical period immediately before impact. Please report the RMSE and bias as a function of time-to-impact, especially for the last 200 ms before impact, and discuss how the system would trigger a mitigation mechanism given this error structure.","section":"§7 and Conclusions"}],"minor_comments":[{"comment":"The abstract cites 2706 ADL sequences in SisFall, while §6 states that the FDNN was tested on 2701 ADL sequences; please reconcile this discrepancy.","section":"Abstract vs. §6"},{"comment":"Table 3 lists 19 features (four demographic, nine inertial, four quaternion, and two derived), but the text in §3 says the input has 18 entries; please correct the count or the table.","section":"Table 3 and §3"},{"comment":"The architecture description references Layers 1–9 but the paper does not include the actual figure; a textual description of the tensor shapes and connections would help the reader follow the design.","section":"§3, Figure 1"},{"comment":"No comparison with existing fall detection or time-to-impact methods is provided; a quantitative comparison on a common protocol would help position the contribution relative to prior work.","section":"§5–§6"},{"comment":"Details of the training procedure are incomplete: the optimizer, learning rate, and early-stopping criterion for the FDNN are not specified, and the k-fold scheme for the KAN parameter search (number of folds, fold construction) is not described.","section":"§3 and §4.2"},{"comment":"The fall-interval definition relies on a threshold on the standard deviation of acceleration; please provide a sensitivity analysis or justify the threshold choice, since the time-of-impact labels depend on it.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a preliminary extended abstract rather than a full journal paper. The core technical risk is the evaluation protocol: the FDNN results are obtained under a random sequence split with demographic inputs, which risks subject-identity memorization, and the KAN feature selection may leak test-fold information. Both issues are fixable with additional experiments, but the paper would need substantial revision to support its generalization claims. I also note the paper does not provide code, which limits reproducibility, and the treatment of the time-to-impact error near impact needs deeper analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a plausible system paper – an LSTM fall detector plus a KAN that estimates time to impact, tested on SisFall – but the headline fall-detection accuracy (TPR 82.6%, TNR 98.4%) is not evidence of generalization to new users because of how the train/test split was done. The concern is right: the 1798 fall sequences come from 24 subjects, each with five repetitions, and the 60/20/20 split is random per sequence. With age, height, weight, and gender as inputs, the model can essentially read the subject's identity off the demographics and the per-sensor signature. So the reported numbers likely measure interpolation to new sequences of known people, not performance on an unseen worker. The stated aim is a plug-and-play device for diverse workers, so this matters.\n\nWhat the paper does well: it is clearly written, uses the public SisFall dataset, reports per-subject and per-activity results, and it is honest about two known confounds – the subjectivity of the Musci annotation and the fact that the 160 ms RMSE is distorted by a ~140 ms floor near impact. The KAN time-to-impact piece is a reasonable continuation of the authors' Fall-KAN work, and the integration with an LSTM detector is a legitimate engineering step.\n\nThe other soft spots are secondary but real. Feature selection for the KAN (correlation/mRMR) appears to be run on the whole dataset before the train/validate/test split, which is a mild leakage path even if the selected features look generic. There are no error bars or confidence intervals, no baseline comparison in the results (despite a mention in the methods), and no measurement of actual inference latency on the wearable target. None of these are fatal by themselves; combined, they push the paper into 'promising but unproven' territory.\n\nWho is it for: people working on pre-impact fall detection and wearable edge systems. The paper deserves a serious referee, but it should come back with a hard requirement: rerun the FDNN evaluation with a subject-independent split, report per-subject performance separately, and add error bars and a baseline. If the numbers survive that, the paper is meaningful. As it stands, I would not cite the fall-detection accuracy as evidence of generalization.\n\nMy recommendation: send to peer review, but expect major revision. It is not desk-reject material; the integration is sensible and the flaw is fixable.","headline":"The fall-detection evaluation is the weak link: a random sequence split with demographics as inputs means the reported accuracy may be identity memorization rather than generalization to unseen users; the KAN time-to-impact part is more plausible but still needs a cleaner evaluation.","tokens_in":8248,"tokens_out":3162,"would_cite":false,"duration_ms":36858,"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 two-stage neural pipeline detects workplace falls and estimates impact time within about 160 ms.","keywords":["fall detection","workplace safety","LSTM","Kolmogorov-Arnold network","time-of-impact estimation","wearable sensors","SisFall dataset","IMU"],"falsifier":"A subject-independent rerun of the same experiment -- train on some subjects' falls and ADLs, test on complete held-out subjects -- would settle the generalization question: if TPR, TNR, or RMSE degrades sharply, the model was memorizing subjects. Complement it with a fresh deployment recording falls from workers not in SisFall and checking whether the 160 ms impact-time estimate still appears before ground contact.","tokens_in":7216,"feed_emoji":"🦺","tokens_out":7488,"duration_ms":89327,"temperature":0.7,"pith_summary":"This paper is trying to establish that a wearable AI system can both recognize a fall as it begins and estimate when the person will hit the ground, before impact. If true, office and small-business workers could wear a device that gives a real-time warning or triggers protection, reducing the long absences falls currently cause. The proposed evidence is a two-stage pipeline tested on the public SisFall dataset: an LSTM-based recurrent network detects fall sequences with 82.6% average sensitivity and 98.4% specificity, while a Kolmogorov-Arnold network estimates time of impact with roughly 160 ms root-mean-square error. The work is aimed at an agile, working-age population, where fast reactions make false alarms a major usability issue.","feed_headline":"Fall-detection AI predicts impact 160 ms ahead, catches 82.6%","feed_subtitle":"Two-stage wearable network separates real falls from daily motion; workers get a real-time warning.","key_machinery":"The mechanism that carries the argument is a two-stage wearable sensing pipeline. The first stage is the Fall Detection Neural Network, an LSTM-based recurrent network that consumes 18 inputs -- raw accelerations, gyroscope rates, quaternion orientation, body angle $\\theta$ relative to gravity, and subject demographics -- and emits the probability of falling. The second is a Kolmogorov-Arnold Network (KAN), a function representation of the time-to-impact map as sums of univariate piecewise linear basis functions fit with the Newton-Kaczmarz method, using five features chosen by correlation and minimum-redundancy maximum-relevance selection. The SisFall recordings and the Musci et al. fall labels supply the shared ground truth that connects detection accuracy and impact-time error.","core_discovery":"The central claim is that a fall can be recognized and timed from waist-worn accelerometer and gyroscope data in real time using a two-stage neural architecture. The first stage, the Fall Detection Neural Network, consumes 18 channels -- raw accelerations from two accelerometers, gyroscope rates, quaternion orientation, body angle relative to gravity, plus the subject's age, height, weight, and gender -- and outputs a falling probability from stacked LSTM layers. The second stage is a Kolmogorov-Arnold Network that estimates remaining time to impact from five selected features using piecewise linear basis functions fit by the Newton-Kaczmarz method. On SisFall the detector reports an average true positive rate of 82.6%, a true negative rate of 98.4% on fall sequences, 94.4% on activities of daily living, and the impact-time estimator reaches roughly 160 ms RMSE. The paper also notes that the time-of-impact estimate saturates at about 140 ms very close to the ground, which defines how much warning the system can actually give.","pith_inferences":["A subject-independent test -- holding out whole people rather than random sequences -- is the natural next check; because same-subject sequences appear in both training and test and demographics are inputs, the reported numbers may partly reflect person recognition rather than fall dynamics.","The observed ~140 ms floor in the impact-time estimate sets a physical constraint: an airbag, brace, or posture assist must deploy within about a tenth of a second of the alarm to change the outcome.","A joint policy that uses the detector's confidence and the KAN's lead time together could tune false alarms against warning time; the paper explicitly leaves this combination for future work."],"forward_implications":["The detector can run on the 200 Hz waist-worn sensor data of SisFall, so a plug-and-play wearable device is a plausible deployment target.","Because the KAN needs only five features, three already in the dataset and two derived from accelerometer and gyroscope, the impact-time stage can work with a simpler sensor configuration than the full 18-channel detector.","The time-of-impact estimate saturates at about 140 ms close to the ground, so any protective actuation has a small, finite warning window to exploit.","Some activities and fall types are systematically harder: fall F10 (forward while trying to sit down) lowers detection, and high-intensity ADLs such as D04 cause the false-positive rate to rise, indicating where the approach needs more data.","The age, height, weight, and gender inputs let the model condition on body type, which matters for a workforce that is more agile than the elderly populations most fall studies target."],"supporting_citations":[{"why":"Supplies the SisFall dataset: 2706 ADLs and 1798 falls from 38 subjects on three waist sensors.","marker":"[18]"},{"why":"Provides the manual fall-period labeling criteria and the LSTM/RNN architecture the fall detector is built around.","marker":"[19]"},{"why":"Introduces the preceding KAN-based time-of-impact estimator that this work extends and evaluates on the full fall set.","marker":"[14]"},{"why":"Supplies the minimum-redundancy maximum-relevance algorithm that selects the KAN's input features.","marker":"[21]"},{"why":"Supplies the Newton-Kaczmarz method used to construct the Kolmogorov-Arnold representation from data.","marker":"[22]"},{"why":"Supplies the Kalman-filter orientation estimation that turns accelerometer and gyroscope data into quaternion body orientation.","marker":"[20]"}],"fun_headline_variants":["Fall AI: 82.6% recall, 160ms impact estimate","Two-stage neural net catches falls, times impact within 160ms","Fall-detection AI: 98.4% specificity, 160ms accuracy","Wearable AI predicts fall impact within 160ms","RNN+KAN fall AI: real-time warning, 160ms precision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the reported accuracy reflects learning what a fall looks like, not recognizing the specific people in the dataset, because training and test sequences come from the same subjects and the model is given each subject's age, height, weight, and gender.","fun_headline_variants_meta":{"raw":{"variants":["Fall AI: 82.6% recall, 160ms impact estimate","Two-stage neural net catches falls, times impact within 160ms","Fall-detection AI: 98.4% specificity, 160ms accuracy","Wearable AI predicts fall impact within 160ms","RNN+KAN fall AI: real-time warning, 160ms precision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001142,"raw_usage":{"total_tokens":4733,"prompt_tokens":931,"completion_tokens":3802,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":3706}},"tokens_in":547,"tokens_out":3802,"duration_ms":38019,"temperature":1.0,"reasoning_tokens":3706,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:20:07.207530+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A subject-independent rerun of the same experiment -- train on some subjects' falls and ADLs, test on complete held-out subjects -- would settle the generalization question: if TPR, TNR, or RMSE degrades sharply, the model was memorizing subjects. Complement it with a fresh deployment recording falls from workers not in SisFall and checking whether the 160 ms impact-time estimate still appears before ground contact.","supporting_citations":[{"cited_title":"Using the two accelerometers and one gyroscope, the orientation of the subject was calculated through a six-axis Kalman filter [20] and expressed as a quaternion","cited_arxiv_id":null,"evidence_quote":"Provides the manual fall-period labeling criteria and the LSTM/RNN architecture the fall detector is built around."},{"cited_title":"Recognition of Physiological Patterns during Activities of Daily Living Using Wearable Biosignal Sensors,","cited_arxiv_id":null,"evidence_quote":"Introduces the preceding KAN-based time-of-impact estimator that this work extends and evaluates on the full fall set."},{"cited_title":"Elderly Fall Detection Using Wearable Sensors: A Low Cost Highly Accurate Algorithm,","cited_arxiv_id":null,"evidence_quote":"Supplies the minimum-redundancy maximum-relevance algorithm that selects the KAN's input features."},{"cited_title":"Fall Risk Reduction for the Elderly by Using Mobile Robots Based on Deep Reinforcement Learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the Newton-Kaczmarz method used to construct the Kolmogorov-Arnold representation from data."},{"cited_title":"Elderly fall detection with an accelerometer using lightweight neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Kalman-filter orientation estimation that turns accelerometer and gyroscope data into quaternion body orientation."}],"review_version":1}