{"id":"be4c35a2-31c6-4d27-9c29-b3b781861323","arxiv_id":"2505.04660","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Synthetic fall data from LLMs helps LSTM fall detectors on low-frequency waist datasets (UMAFall +56.83%) but hurts on high-frequency or wrist datasets; diffusion data matches real data best yet does not reliably improve detection.","lead":"This paper tests whether AI language models and a diffusion model can generate fake accelerometer data to help train smartwatch fall detectors. It finds the answer depends on the real dataset being augmented, with synthetic data helping on low-frequency waist-worn datasets and hurting on some wrist-worn ones.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Few-shot examples are drawn from the same subjects that can appear in the test split; the +56.83% UMAFall result may be an artifact of subject leakage rather than dataset characteristics.","rationale":"I agree with the reader's overall CONDITIONAL verdict, but I identify a different weakest assumption. The reader's weakest_assumption focuses on the 1/46 s time base used to convert text-to-motion joint positions to acceleration (Eq. 4, Section 5.1). That is a genuine concern for the text-to-motion arm of the study, since no resampling or temporal alignment to the baseline sampling rates is reported, and the paper's frequency-dependent conclusions would be weakened if the time bases are mismatched. However, the few-shot leakage concern is more load-bearing for the paper's central quantitative claim, because the +56.83% UMAFall improvement is the single most striking number supporting the claim that dataset characteristics determine synthetic-data effectiveness. If the five few-shot subjects can appear in the test set, then the model may be evaluated on data whose real accelerometer patterns were explicitly included in the synthetic training samples, inflating F1 in a way that has nothing to do with sensor placement, fall representation, or sampling rate. The paper's own Section 7 even states that synthetic data effectiveness varies with dataset characteristics, and the strongest empirical support for that statement is the UMAFall versus SMM comparison; leakage would directly undermine that comparison. The diffusion baseline comparison is also not fully fair because Diffusion-TS is trained on each real dataset before being compared with it, but that concern is about the interpretation of an expected result, not about the validity of the central fall-detection claim. The lack of error bars is important but secondary: even with standard deviations, a leaked test set would still bias the means. I therefore recommend keeping the CONDITIONAL verdict, with the specific condition that few-shot evaluations be repeated under a subject-disjoint split and with variance reported. If the leakage-free rerun reproduces the qualitative pattern, the central claim would be substantially strengthened; if it does not, the paper would need to be revised or rejected.","tokens_in":26241,"tokens_out":4341,"duration_ms":46880,"concrete_test":"Rerun the UMAFall, SMM, KFall, and SisFall few-shot conditions with an explicit subject-level holdout: choose the five few-shot subjects first, exclude them from the 12-subject pool, then perform the 8/2/2 leave-two-out split on the remaining subjects. Report mean and standard deviation of F1 over the five iterations. If GPT4-FS on UMAFall no longer shows a large positive delta, or if the SMM versus UMAFall ranking changes, the reported dataset-dependence is an artifact of leakage. Also report the overlap between shot subjects and test subjects in the original protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline evidence for dataset-dependent benefit is the few-shot text-to-text result, especially UMAFall GPT4-FS (+56.83%, Table 3). Section 3.1.2 and Section 5.2 state that few-shot data are generated by providing the LLM with fall samples from 'five randomly selected subjects' of the corresponding baseline dataset. Section 4 describes a leave-two-out subject split in which 8 of 12 subjects are training, 2 validation, and 2 test, with the split re-randomized over five iterations. Nowhere is it reported that the five few-shot subjects are excluded from validation or test. Because few-shot prompting feeds the model actual accelerometer windows of those subjects, any test subject overlapping the shot set has its own fall patterns embedded in the synthetic training data. The LSTM can then memorize near-copies of held-out falls, inflating F1. This is not a minor implementation detail: the largest claimed improvement (+56.83% for UMAFall) and the contrast with SMM (best method -4.05%) are exactly the numbers used to conclude that dataset characteristics determine synthetic-data effectiveness. Without a subject-disjoint split, the central claim is not yet supported. The Delta-t = 1/46 s time-base issue raised by the reader is real for text-to-motion data, but the leakage issue directly threatens the largest quantitative result and is easier to test.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates the use of large language models (text-to-motion: T2M-GPT, SATO, ParCo; text-to-text: GPT-4o, GPT-4, Gemini-1.5-Flash-8B) and a time-series diffusion model (Diffusion-TS) to generate synthetic fall accelerometer data, and tests whether augmenting four real-world fall datasets (SmartFallMM, KFall, UMAFall, SisFall) with these synthetic data improves LSTM-based fall detection. The authors report that the effect of synthetic data depends on dataset characteristics: UMAFall improved by up to +56.83% with GPT-4 few-shot data, while all synthetic methods degraded SmartFallMM performance, with T2M showing the smallest decline (-4.05%). They also report that Diffusion-TS produces the closest distributional alignment to real data (lowest JSD, highest Coverage) but does not consistently improve downstream performance. An ablation study claims to show that sensor placement, prompt strategy, and synthetic data quantity influence results.","tokens_in":26491,"tokens_out":7971,"duration_ms":69156,"significance":"If the findings were supported, the paper would provide useful guidance for using generative models to augment scarce fall data in wearable health monitoring. The study is broad in scope, comparing multiple generation paradigms on four public datasets, and it includes an unusually candid limitations section (Section 7) acknowledging that the reported gains are insufficient for real-world deployment. However, the central dataset-dependent claim is currently undermined by a likely subject-leakage issue in the few-shot experiments, a structurally biased Diffusion-TS comparison, and a time-base mismatch in the text-to-motion pipeline. These issues are fixable in principle, but the quantitative conclusions as presented are not reliable. The paper would need a re-analysis with subject-disjoint few-shot sampling, with Diffusion-TS evaluated on held-out data, and with proper temporal alignment before the empirical claims can be accepted.","major_comments":[{"comment":"The few-shot synthetic data are generated by providing the LLM with fall samples from 'five randomly selected subjects' per baseline dataset (Sections 3.1.2 and 5.2), while classifier evaluation uses a leave-two-out subject split with 8 training, 2 validation, and 2 test subjects, re-randomized over five iterations (Section 4). The paper never states that the five shot subjects are excluded from the validation and test folds. Because the few-shot prompt includes actual accelerometer windows of those subjects, any test subject overlapping the shot set has its own fall patterns embedded in the synthetic training data, which the LSTM can memorize and exploit. This directly threatens the headline UMAFall GPT4-FS result (+56.83%, Table 3) and the dataset-dependent claim in Section 7. The experiment must be re-run with shot subjects chosen from the training fold only, and the results re-reported for all datasets.","section":"Section 3.1.2 / Section 4"},{"comment":"Diffusion-TS is trained on the fall data of each real dataset and then evaluated for alignment with that same dataset (Section 3.2.2: 'We trained the Diffusion-TS model for each dataset'; Section 5.2 computes JSD and Coverage against the four real datasets). A generative model trained on the test distribution will naturally achieve low JSD and high Coverage; the finding that Diffusion-TS 'consistently outperforms' LLMs in alignment (Section 5.2 Findings) is therefore a property of training on the target data, not a general property of diffusion models. To support RQ2, the authors should compare all methods on held-out data, train Diffusion-TS only on a training subset and evaluate on a held-out split, or clearly reframe the claim as an upper-bound comparison rather than a model-level conclusion.","section":"Section 3.2.2 / Section 5.2"},{"comment":"Equation (4) defines acceleration as a(f) = (p(f+1) - p(f))/Δ t^2, which is dimensionally incorrect; this is a scaled forward difference, not a second derivative. A correct discrete second derivative would be (p(f+1) - 2p(f) + p(f-1))/Δ t^2. Additionally, the text-to-motion models output frames at Δ t = 1/46 s, and no resampling, filtering, or temporal alignment to the baseline sampling rates (20, 32, 100, 200 Hz) is described before the synthetic accelerometer data are compared with and added to the real windows. The frequency-dependent conclusions in Sections 5.2 and 7 (e.g., LLM data 'loses specificity' at higher sampling rates) may be artifacts of mismatched time bases rather than properties of the generators. The authors should correct the formula and describe or implement the necessary temporal alignment.","section":"Section 5.1, Eq. (4)"},{"comment":"The text in Section 5.3 claims that SATO's high gains are 'justified by its lowest JSD' for UMAFall (0.6405) and SisFall (0.8128), but in Figure 6 SATO has the highest JSD among text-to-motion models on both datasets (UMAFall: T2M 0.5483, ParCo 0.6176, SATO 0.6405; SisFall: T2M 0.6921, ParCo 0.7301, SATO 0.8128). The stated numbers contradict the supporting text, so the explanatory link between distributional alignment and downstream F1 gains is not established as written. The authors should correct the misstatements and provide a quantitative analysis (e.g., a correlation or a controlled comparison) if they wish to claim such a link.","section":"Section 5.3 / Figure 6"},{"comment":"The ablation in Section 6.1 changes the real baseline modality (SMM to right hip, UMAFall to waist) but for the text-to-text models it uses 'the same zero-shot synthetic data' and few-shot data generated with fall samples from the new datasets, without specifying that the prompt or output location was changed. Since the text-to-text pipeline generates left-wrist accelerometer data (Figure 2(b)), the synthetic text-to-text data remain wrist-specific while the real data are hip/waist; thus the experiment does not control sensor placement for text-to-text models. The conclusion that baseline dataset characteristics, including sensor placement, determine synthetic-data effectiveness is not supported for text-to-text models; the authors must either regenerate text-to-text data for the new sensor location or restrict the claim to text-to-motion and Diffusion-TS results.","section":"Section 6.1 / Table 4"},{"comment":"All conclusions are based on five iterations of an 8/2/2 subject split with 12 subjects, and Table 3 reports only average F1-scores with no standard deviations, confidence intervals, or significance tests. Many of the reported differences (e.g., SMM: T2M -4.05% vs Diffusion -8.11%; SisFall: several models within ±1-2%) are small relative to the likely variance from a two-subject test set, so the dataset-dependent ranking may not be stable. The authors should report per-fold results and perform a basic statistical comparison (e.g., paired bootstrap or Wilcoxon across folds) to support the claim that the ranking is not noise.","section":"Section 4 / Table 3"}],"minor_comments":[{"comment":"The title page has a typo ('AI-G ENERATED FALL DATA') and Section 3.2.1 has the heading 'Basline Datasets' instead of 'Baseline Datasets'.","section":"Title and Section 3.2.1"},{"comment":"The counts of synthetic datasets and models are inconsistent: Section 3.2.3 says 13 augmented models per dataset (1 diffusion + 3 text-to-motion + 9 text-to-text), Section 5.2 says 28 synthetic datasets in total, but Table 3 lists 10 augmented models per dataset (1 + 3 + 6), which would be 40 across four datasets. Please correct the counts and clarify which text-to-text configurations were actually evaluated.","section":"Sections 3.2.3, 5.2, and Table 3"},{"comment":"Model names are used inconsistently: 'GPT-4', 'GPT4', and 'GPT-4o' appear interchangeably in the text, tables, and figures; please standardize the notation.","section":"Throughout"},{"comment":"The phrase '12 (4×3) sets of FS accelerometer data' is inconsistent with the six FS text-to-text conditions shown in Table 3; please clarify the intended number of few-shot datasets.","section":"Section 5.2"},{"comment":"The KS-test results are reported only as average p-values per model and condition; consider providing the full range or a compact table to indicate variability across axes and repetitions.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript's central quantitative claims hinge on the few-shot subject-split issue, which I expect the authors can fix by re-running with subject-disjoint shots. The Diffusion-TS comparison needs reframing to avoid the train-on-test circularity, and the time-base mismatch in Eq. (4) needs correction. The manuscript also contains internal count inconsistencies (13 vs 10 vs 28 vs 40 models/datasets) that make the experimental protocol hard to audit; the authors should provide a precise enumeration. I would encourage the authors to include standard deviations across the five folds, as the reported F1 differences are often small. I do not see grounds for rejection, but the current version is not publishable without substantive reanalysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful empirical map of whether synthetic data from LLMs and diffusion helps wearable fall detection, and the paper is honest about its own limits. But the two headline numbers—Diffusion-TS as the best-aligned, and the +56.83% UMAFall gain from GPT-4 few-shot—rest on assumptions the paper doesn't check.\n\nWhat is actually new: the dataset-conditional finding—that sensor placement, sampling rate, and fall representation determine whether synthetic data helps—is not in the prior work they cite. They run the same generation and evaluation pipeline across four public datasets (SMM, KFall, UMAFall, SisFall) and include an ablation that flips sensor location, which strengthens the qualitative point. They also state the applicability limits plainly: the best F1 is still below 0.9, prompt engineering is labor-intensive, and LLM API variability makes real-time generation impractical. That is the right tone for this subfield.\n\nWhere it gets soft, in order of severity:\n\n1. Subject leakage. Few-shot data is generated from five subjects per dataset, and the leave-two-out split randomly chooses test subjects over five iterations. The paper never says the five shot subjects are excluded from validation/test. If any overlap occurs, the LSTM can memorize near-copies of held-out falls, and the +56.83% becomes an artifact. This is the largest claimed improvement, so it needs a subject-disjoint split or a report of overlap rates.\n\n2. The diffusion comparison is structurally biased. Diffusion-TS is trained on the real fall data of each dataset and then reported as closest in JSD/Coverage to that same data. That is expected and not a fair comparison against zero-shot LLM data. It also weakens the 'Diffusion-TS is stable' narrative.\n\n3. Time-base mismatch. Text-to-motion uses a fixed delta_t = 1/46 s, while real datasets run at 20, 32, 100, 200 Hz. No resampling or alignment is described, so frequency-dependent conclusions might be an artifact.\n\n4. No variance. Five iterations but only mean F1, no error bars or significance tests. Some differences (e.g., small SisFall gains) could be noise.\n\nNone of this kills the qualitative central claim; the pattern across datasets and the ablation are plausible. But the quantitative evidence as reported is not sufficient.\n\nWho benefits: researchers deciding whether to invest in LLM synthetic data for fall detection, or comparing generative approaches. It deserves a serious referee, but I'd ask for a leakage-controlled split, variance reporting, and a fair diffusion baseline before trusting the numbers.","headline":"Useful empirical map of synthetic fall data, but the headline numbers rest on a likely subject leak and a biased diffusion comparison.","tokens_in":27024,"tokens_out":2316,"would_cite":false,"duration_ms":22686,"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":"Synthetic fall data improves wearable fall detection only when the target dataset matches it in sensor placement and fall representation, with F1 gains up to +56.83 percent and the best-case loss at -4.05 percent on an incompatible dataset.","keywords":["fall detection","synthetic data generation","large language models","text-to-motion generation","text-to-text generation","diffusion models","time-series analysis","LSTM"],"falsifier":"Resample each synthetic dataset to the sampling rate of its matching baseline (or low-pass filter real data to 23 Hz) and re-run the F1 comparisons; if the reported pattern of low-frequency datasets gaining and high-frequency datasets losing disappears or reverses, the sampling-rate conclusion is an artifact of mismatched time bases rather than a property of LLM-generated data.","tokens_in":26000,"feed_emoji":"⌚","tokens_out":13200,"duration_ms":113779,"temperature":0.7,"pith_summary":"This paper asks whether large language models can generate synthetic accelerometer data for wearable fall detection, and whether adding that data to real training data improves an LSTM fall detector. The authors generate falls with text-to-motion models (T2M, SATO, ParCo), text-to-text LLMs (GPT-4o, GPT-4, Gemini), and a diffusion time-series model, then augment four real datasets (SmartFallMM, KFall, UMAFall, SisFall). Their central finding is that the effect of synthetic data depends on the baseline dataset's characteristics, especially sensor placement and whether the data captures only the impact phase of a fall or the whole transition into it. Gains range from +56.83% on UMAFall with GPT-4 few-shot data to a -4.05% best-case drop on SmartFallMM, where every synthetic source hurt. The practical stakes are that synthetic data is one of the few ways around the scarcity and ethical difficulty of collecting real falls from elderly people, and this study shows that the same generator can help or hurt depending on the target dataset.","feed_headline":"Synthetic fall data boosts one dataset 57%, hurts another 4%","feed_subtitle":"Gains depend on matching sensor placement; synthetic data can replace scarce real falls only when aligned.","key_machinery":"The load-bearing machinery is the conversion from generated 3D joint motion to accelerometer time series, together with the comparison protocol around it. Text-to-motion models (T2M, SATO, ParCo) produce 3D joint trajectories from prompts; the paper extracts specific joints from the SMPL 22-joint body model (left wrist index 20, right wrist index 21, waist/pelvis index 0) and computes synthetic acceleration as the discrete second derivative of position, $a(f) = (p(f+1) - p(f))/\\Delta t^2$, with $\\Delta t = 1/46$ s. Text-to-text LLMs are prompted directly for CSV accelerometer data, while Diffusion-TS is trained on real falls to generate synthetic time series. All synthetic sources are mixed into LSTM training at a 60% daily-activity / 20% real-fall / 20% synthetic-fall ratio and evaluated on held-out subjects with F1. The step doing the argumentative work is the assumption that these heterogeneous time bases—46 Hz generated frames versus 20, 32, 100, or 200 Hz real recordings—can be compared and combined without resampling or temporal alignment.","core_discovery":"The paper's central claim is that the usefulness of LLM-generated synthetic fall data for training a wearable fall detector is set by the match between the synthetic data and the baseline dataset, not by the generator alone. Concretely, on UMAFall—a 20 Hz dataset with right-wrist sensor data and only three fall types—GPT-4 few-shot synthetic data raised LSTM F1 by 56.83% over training on real data alone, while on SmartFallMM—a 32 Hz left-wrist dataset whose fall windows contain only the impact phase—every synthetic source lowered F1, with the least damage being T2M's -4.05%. The authors attribute this to two dataset characteristics: sensor placement (left wrist captures low-variance motion, while waist and hip placements carry more fall-relevant movement) and fall representation (impact-only windows do not align with synthetic sequences that model gradual transitions into falls). They also find that text-to-motion models track joint-specific movement better than text-to-text models, that none of the LLMs encode age or gender differences in the generated signals, and that diffusion-based synthetic data aligns most closely with real distributions yet does not consistently improve detection.","pith_inferences":["The paper leaves implicit that the 46 Hz frame rate of text-to-motion output band-limits synthetic acceleration below roughly 23 Hz, which may explain why high-frequency datasets reject synthetic data; resampling synthetic data to each baseline rate, or low-pass filtering real data to 23 Hz, would separate a time-base artifact from a real property of LLM-generated falls.","A selection rule the paper does not propose is to choose the synthetic source with the lowest Jensen-Shannon divergence only when its fall-phase structure matches the baseline (impact-only versus gradual transition); this could turn the observed dataset dependence into a practical recipe.","The ablation's sensor-placement result suggests a simple deployment guideline the paper stops short of stating: generate or extract synthetic data from the exact joint where the wearable will sit, rather than from a generic fall description.","Because the paper's own applicability note says even the best F1 stays below 0.9 and quantization erodes gains, a more promising test than direct mixing is to use LLM synthetic data for pre-training or regularization and then fine-tune on a small amount of real data."],"forward_implications":["Augmentation with synthetic LLM data should be recommended only after checking the target dataset's sensor site and fall-window structure; the paper's results show this is not a generator-only property.","Text-to-motion models are the safer LLM family for synthetic falls because they preserve joint-specific movement, while text-to-text models should be reserved for low-rate baselines.","High-frequency datasets are the failure mode for few-shot text-to-text data, which can lower F1 (for example GPT-4o few-shot on SisFall, -7.10%) even when zero-shot data from the same model helps.","Diffusion-based synthetic data produces the best-aligned windows across all four baselines but does not reliably improve detection, so distributional alignment alone is not sufficient for an augmentation benefit.","Fine-tuning text-to-motion models on real fall data, which the paper proposes as future work, is the most direct route to synthetic data that matches both sensor placement and fall dynamics."],"supporting_citations":[{"why":"Supplies the T2M-GPT text-to-motion generator used to produce 3D fall motion sequences from prompts.","marker":"[23]"},{"why":"Supplies the SATO diffusion-based text-to-motion generator used for synthetic fall data.","marker":"[32]"},{"why":"Supplies the ParCo part-coordinated text-to-motion generator used for synthetic fall data.","marker":"[33]"},{"why":"Supplies Diffusion-TS, the diffusion time-series generator that serves as the alignment and stability baseline.","marker":"[44]"},{"why":"Provides the KFall dataset, one of the four real baselines whose augmentation results support the dataset-dependence claim.","marker":"[41]"},{"why":"Provides the UMAFall dataset, the low-frequency baseline where the largest F1 gain (GPT-4 few-shot, +56.83%) is observed.","marker":"[42]"},{"why":"Provides the SisFall dataset, the 200 Hz baseline where few-shot LLM data causes instability and performance drops.","marker":"[43]"},{"why":"Defines the 22-joint SMPL body model whose joint indices (left wrist 20, right wrist 21, waist 0) determine which synthetic joint data is extracted.","marker":"[52]"},{"why":"Defines the Coverage metric used to quantify how much of the real fall distribution the synthetic data represents.","marker":"[54]"},{"why":"Defines Jensen-Shannon Divergence, the metric used to measure distributional alignment between real and synthetic falls.","marker":"[55]"}],"fun_headline_variants":["Synthetic fall data: 57% gain on one set, 4% loss on another","Sensor placement decides if LLM fall data helps or hurts","Right-wrist fall data gains 57%, left-wrist loses 4% with LLM","Diffusion matches real falls but doesn't improve detection","Synthetic fall data works only if it matches sensor placement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison assumes that acceleration computed from 3D joint positions at a fixed 46 Hz frame rate is directly comparable to real accelerometer readings sampled at 20, 32, 100, or 200 Hz, with no resampling or temporal alignment.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic fall data: 57% gain on one set, 4% loss on another","Sensor placement decides if LLM fall data helps or hurts","Right-wrist fall data gains 57%, left-wrist loses 4% with LLM","Diffusion matches real falls but doesn't improve detection","Synthetic fall data works only if it matches sensor placement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3726,"prompt_tokens":1021,"completion_tokens":2705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":2609}},"tokens_in":637,"tokens_out":2705,"duration_ms":19612,"temperature":1.0,"reasoning_tokens":2609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:38:57.872257+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Resample each synthetic dataset to the sampling rate of its matching baseline (or low-pass filter real data to 23 Hz) and re-run the F1 comparisons; if the reported pattern of low-frequency datasets gaining and high-frequency datasets losing disappears or reverses, the sampling-rate conclusion is an artifact of mismatched time bases rather than a property of LLM-generated data.","supporting_citations":[{"cited_title":"Parco: Part-coordinating text-to-motion synthesis, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the ParCo part-coordinated text-to-motion generator used for synthetic fall data."},{"cited_title":"Diffusion-ts: Interpretable diffusion for general time series generation, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies Diffusion-TS, the diffusion time-series generator that serves as the alignment and stability baseline."},{"cited_title":"A large-scale open motion dataset (kfall) and benchmark algorithms for detecting pre-impact fall of the elderly using wearable inertial sensors","cited_arxiv_id":null,"evidence_quote":"Provides the KFall dataset, one of the four real baselines whose augmentation results support the dataset-dependence claim."},{"cited_title":"Santoyo-Ramón, and Jose M","cited_arxiv_id":null,"evidence_quote":"Provides the UMAFall dataset, the low-frequency baseline where the largest F1 gain (GPT-4 few-shot, +56.83%) is observed."},{"cited_title":"Sisfall: A fall and movement dataset","cited_arxiv_id":null,"evidence_quote":"Provides the SisFall dataset, the 200 Hz baseline where few-shot LLM data causes instability and performance drops."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the 22-joint SMPL body model whose joint indices (left wrist 20, right wrist 21, waist 0) determine which synthetic joint data is extracted."},{"cited_title":"Reliable fidelity and diversity metrics for generative models, 2020","cited_arxiv_id":null,"evidence_quote":"Defines the Coverage metric used to quantify how much of the real fall distribution the synthetic data represents."},{"cited_title":"Endres and Johannes E","cited_arxiv_id":null,"evidence_quote":"Defines Jensen-Shannon Divergence, the metric used to measure distributional alignment between real and synthetic falls."}],"review_version":1}