{"id":"11f8510b-26f9-41b8-973b-3bf66a715150","arxiv_id":"2412.00411","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Seismocardiography from a single chest-worn accelerometer performs comparably to ECG and BVP for binary valence and arousal classification, and adding accelerometry-derived respiration enables a one-sensor emotion recognition setup.","lead":"Using chest-worn accelerometer data from the EmoWear dataset, the authors show that seismocardiography, the chest vibrations caused by heartbeats, can classify emotional valence and arousal about as well as standard ECG or blood-volume-pulse signals. The result suggests emotion recognition could run on a single commodity accelerometer, the same sensor already inside phones and wearables.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hyperparameter and classifier selection on the full EmoWear dataset (Sec. III-C) before LOVO evaluation inflates F1 and invalidates the significance tests supporting the SCG+ADR 'working framework' claim.","rationale":"The reader's stated weakest assumption is the reliability of AO-peak detection, but their rationale also identifies the more consequential validation flaw: model and hyperparameter selection on the entire dataset before LOVO evaluation. I focus on the validation flaw because it directly undermines the statistical significance that supports the headline claim of a 'working single-accelerometer framework.' The paper is transparent about this selection procedure (Sec. III-C and Limitations), which is a credit, but the transparency does not remove the bias. The relative SCG vs ECG/BVP comparison may be less biased because all modalities went through the same selection process, so the paper's core comparative finding could survive. However, the absolute 'working framework' claim rests on F1 margins of only ~0.05 over ratio baseline, and these margins could vanish under a properly nested validation scheme. The proposed concrete test would settle whether the significance results are artifacts of selection leakage. The verdict remains CONDITIONAL because the relative claim is plausible and the paper provides a useful benchmark, but the central absolute claim requires re-analysis before it can be accepted.","tokens_in":25500,"tokens_out":3272,"duration_ms":32590,"concrete_test":"Re-run the EmoWear experiments with nested LOVO: for each training fold, perform grid search over classifiers and C on an inner leave-one-video-out split of the training videos, then evaluate the selected model on the held-out video. Compute macro-F1 and the one-sample t-test for SCG+ADR (rows 20-21) under this protocol. If F1 drops by more than ~0.02 or the t-test no longer reaches p<0.05, the 'working framework' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C states that the choice of SVM/LR over other classifiers and the hyperparameter grid search were optimized to maximize macro-F1 on the entire EmoWear dataset, prior to the leave-one-video-out evaluation. This means the LOVO CV estimates performance of a pipeline selected using the test videos; the one-sample t-tests in Table VI (rows 20-21) therefore do not test generalization of a fixed pipeline. The margins over the best baseline are small: SCG+ADR SVM/LR F1 = 0.550-0.564 vs ratio baseline 0.500, and the NB variant is non-significant. A proper nested validation could erase these margins. The relative SCG vs ECG/BVP comparison is less affected because the same selection procedure was applied to all cardiac sources, but the central 'working single-accelerometer framework' claim depends on the absolute significance, which is compromised. The feature-selection threshold and minimum feature count may also have been tuned on the full dataset, compounding the leakage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces seismocardiography (SCG) and accelerometry-derived respiration (ADR), both obtained from a single chest-worn accelerometer, as modalities for emotion recognition. The authors replicate the peripheral-signal emotion classification pipeline of the DEAP study on DEAP, obtaining similar performance, and then apply the same pipeline to the EmoWear dataset, comparing SCG against ECG and BVP as cardiac sources. They report macro-F1 scores around 0.55-0.59 for SCG, similar to ECG and BVP, and report that SVM and LR classifiers using SCG+ADR significantly exceed baseline voting, leading to the claim that a single chest-worn accelerometer provides a working emotion recognition framework. The paper also contains a critical review of methodological weaknesses in prior DEAP-based studies, including class imbalance, F1 reporting, and validation practices.","tokens_in":25586,"tokens_out":6700,"duration_ms":67489,"significance":"If the results survive a leakage-free evaluation, the contribution is meaningful: it is the first evaluation of SCG as an emotion-recognition modality, it establishes initial benchmarks on a public dataset, and it demonstrates that cardiac and respiratory information can be obtained from a single chest-worn accelerometer. The study's strengths include the use of macro-averaged F1, the inclusion of multiple baselines, a DEAP replication as a sanity check, and the use of a publicly available dataset. The reported effect sizes are modest, however, and the central single-accelerometer claim rests on significance tests that are compromised by the current model-selection procedure; the contribution is therefore best viewed as promising but not yet fully supported.","major_comments":[{"comment":"The grid-search optimization described in Section III-C maximizes macro-F1 on the entire EmoWear dataset before the LOVO evaluation. This means the held-out videos in each LOVO fold have already influenced the choice of classifier family (SVM/LR over NB, k-NN, trees, boosting, and neural networks) and the hyperparameters (regularization C, solver). The LOVO estimates in Table VI, including rows 20-21 for SCG+ADR, are therefore not unbiased estimates of a fixed pipeline, and the one-sample t-tests against the 0.500 baseline do not test generalization of a pre-specified method. Because the margins over baseline are small (macro-F1 0.550-0.564 vs 0.500), selection leakage could account for the reported significance. I recommend nested cross-validation, or fixing the pipeline on the DEAP data alone and applying it unchanged to EmoWear.","section":"Section III-C, 'Classifiers'"},{"comment":"The paper does not state whether the Fisher score threshold (Eq. 1) and the minimum ranked feature count of 15 are computed inside each LOVO training fold or on the entire dataset before splitting. If feature selection is performed on all videos of a subject, including the held-out video, then the test labels are used to choose features and the reported F1 values are optimistically biased. This is not a minor detail: the discussion notes that only RSP features were selected in some configurations, so the selected feature set is small and potentially unstable. The authors should clarify the timing of feature selection and, if it is not already nested, re-run the experiments with feature selection performed on the training folds only.","section":"Section III-C, 'Feature Selection'"},{"comment":"The HR and HRV features from SCG are derived from AO peaks detected with the Massaroni algorithm, but the manuscript provides no validation of these detections on the EmoWear recordings. The EmoWear dataset includes simultaneous ECG, so the authors can report detection agreement (e.g., F1 of AO peaks against ECG R-peaks) or the correlation between SCG-derived IBI/HRV and ECG-derived IBI/HRV. Without such a check, the reader cannot distinguish a genuinely cardiac SCG feature set from one that is partly driven by motion artifacts or missed/false detections. Given that the paper's central equivalence claim is that SCG carries HRV information comparable to ECG, this validation is load-bearing.","section":"Section III-B, 'Seismocardiography'"}],"minor_comments":[{"comment":"The phrase 'gate-analysis' appears to be a typo for 'gait-analysis'.","section":"Section V.A"},{"comment":"The markers '†' and '‡' in the pipeline figure are not defined in the caption or in the surrounding text; please add a legend explaining their meaning.","section":"Figure 2"},{"comment":"Row 22 places 'Baseline' in the classifier column, which is confusing; consider aligning baseline rows with the setup columns for clarity.","section":"Table VI"},{"comment":"The claim that SCG is 'less affected by motion artifacts than other accelerometer-based methods' is not tested in this study; if retained, it should be supported with evidence or softened.","section":"Section V.A"},{"comment":"The statement that EmoWear is 'the only one to provide chest-worn accelerometer data validated for both SCG and ADR purposes' relies on the authors' own dataset paper; an independent validation citation would strengthen the claim.","section":"Section I"}],"recommendation":"major_revision","confidential_remarks":"The paper is built around the authors' own EmoWear dataset and ColEmo software; the dataset is public and this is disclosed, so I do not see a circularity problem. The main issue is that the model-selection procedure is described as using the full EmoWear dataset before the LOVO evaluation, which compromises the central significance claims. A leakage-free re-analysis is needed before the paper can be accepted as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful result here is the relative one: seismocardiography from a chest-worn accelerometer classifies binary valence and arousal about as well as ECG and BVP in the EmoWear dataset. That claim mostly holds up, and it is the first time SCG has been tested as an HRV-carrying cardiac signal for emotion recognition. The DEAP replication is a reasonable validity check, and the first EmoWear benchmarks are a real contribution. The paper is also unusually transparent about its choices, which makes the problems easier to see.\n\nThe soft spot is exactly what the stress-test note flags. Section III-C says classifier and hyperparameter selection was done by grid search maximizing macro-F1 on the entire EmoWear dataset before the LOVO evaluation. So the one-sample t-tests in Table VI are not testing a fixed pipeline; they are testing a pipeline selected partly on the held-out videos. The SCG+ADR margins over the ratio baseline are small (F1 0.550-0.564 vs 0.500), and the NB variant is non-significant. A nested validation could erase those margins. This mainly affects the absolute \"single chest-worn accelerometer works\" claim. The SCG vs ECG/BVP comparison is less affected because the same selection procedure was applied to all cardiac sources, so I would trust that relative comparison more. The feature-selection threshold and minimum feature count may also have been tuned on the full dataset, and the subject exclusions based on a 10% rating-balance threshold are post hoc. Those are secondary but worth noting.\n\nThe AO-peak detection assumption is real but not the main problem; the paper uses a published algorithm and the comparative design means any systematic detection error would likely hurt all SCG conditions rather than selectively boost them.\n\nWho gets value from this paper: people working on wearable affective computing, especially anyone building on EmoWear. The benchmark tables and the critical literature review are genuinely useful. The central relative finding is probably right, but the absolute claim needs a cleaner evaluation before it should be cited as evidence that a single accelerometer is a working emotion-recognition setup.\n\nMy recommendation: send it to peer review, but require a nested validation scheme or a re-framing of the SCG+ADR result as exploratory. Also ask the authors to run feature selection inside each fold and release the exact hyperparameters and code. The paper is worth refereeing seriously, and the relative result will likely survive the revision.","headline":"SCG is a plausible HRV-carrying modality for emotion recognition and the relative comparison to ECG/BVP is the solid part; the absolute 'single accelerometer works' claim needs a nested validation before it is taken at face value.","tokens_in":26240,"tokens_out":1860,"would_cite":true,"duration_ms":21325,"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":"The paper argues that seismocardiography (the chest-wall vibration from each heartbeat) is as informative for emotion recognition as ECG or blood-volume-pulse signals, and that combining it with respiration from the same chest-worn…","keywords":["seismocardiography","emotion recognition","accelerometry-derived respiration","heart rate variability","wearable sensors","affective computing","EmoWear dataset","DEAP dataset"],"falsifier":"Record the same subjects with a chest-worn accelerometer and a reference ECG while they walk, talk, and move their torso; if the detected vibration peaks lose alignment with the ECG peaks and the resulting heart-rhythm features diverge from ECG-derived features by more than the distance between emotion classes, the central parity claim would be refuted.","tokens_in":25170,"feed_emoji":"❤️","tokens_out":8744,"duration_ms":80330,"temperature":0.7,"pith_summary":"The paper sets out to prove that seismocardiography — the chest-wall vibration caused by each heartbeat — can serve as a stand-alone signal for emotion recognition, replacing the electrical heart signal (ECG) and the optical pulse signal (BVP). Using the EmoWear dataset and a classification pipeline validated by reproducing the DEAP benchmark, the authors obtain macro-F1 scores around 0.55–0.59 for binary valence and arousal with SCG-based features, in the same range as ECG- and BVP-based setups. They then show that combining SCG with respiration recovered from the same accelerometer (ADR) still yields classifiers whose results are significant against random, majority, and ratio baselines. If the claim holds, affective computing loses its need for extra electrodes or optical sensors: the accelerometer already inside many wearables could supply both cardiac and respiratory information. The paper presents its EmoWear results as first benchmarks for this new modality.","feed_headline":"Chest accelerometer alone matches ECG and BVP for emotion reading","feed_subtitle":"SCG plus accelerometer-derived respiration classifies valence and arousal from a single chest-worn sensor with ECG-level results.","key_machinery":"The load-bearing mechanism is the detection of aortic-valve-opening (AO) peaks in the SCG signal: the mechanical event of the aortic valve opening, visible as a chest-wall vibration shortly after the ECG R-peak. These AO peaks are treated as mechanical analogues of ECG R-peaks, from which inter-beat intervals and heart-rate-variability features are computed and passed to the classifiers. The detection chain applies band-pass filtering at 10–20 Hz, Hilbert-transform envelope extraction, a second 0.5–2 Hz band-pass, and then peak picking. Respiratory context comes from ADR, obtained by band-pass filtering the same accelerometer signal at 0.15–0.35 Hz and detrending, so that both cardiac and respiratory inputs come from one sensor.","core_discovery":"The central claim, on the paper's own terms, is that seismocardiography (SCG) is a viable and previously unexplored physiological modality for recognizing emotional valence and arousal. Using the EmoWear dataset and a pipeline validated by replicating the DEAP single-trial classification study, the authors train subject-dependent naive Bayes, SVM, and logistic regression classifiers on heart-rate and heart-rate-variability features extracted from AO peaks in the SCG signal. SCG-based setups land in the same macro-F1 range as ECG- and BVP-based setups (for example, SCG plus all peripherals with SVM reaches 0.587 valence and 0.579 arousal, versus 0.584/0.579 for ECG and 0.585/0.573 for BVP in the same configuration). Combining SCG with accelerometry-derived respiration (ADR) yields results that are significantly above baseline voting for SVM and logistic regression, although naive Bayes does not reach significance with that pairing. The paper concludes that a single chest-worn accelerometer can serve as a physiological gateway for emotion recognition, and positions its EmoWear results as the first benchmarks for this dataset.","pith_inferences":["If the parity with ECG and BVP holds, emotion recognition becomes a zero-hardware add-on to accelerometers already worn for activity tracking, fall detection, and gait analysis.","A direct validation of AO-peak detection against ECG during natural motion would settle the surrogate question; the paper does not report beat-to-beat timing error, so that check remains open.","Because the same sensor records body motion, motion artifacts in SCG could in principle be modeled and removed using the accelerometer's own movement channel — an option ECG and BVP do not offer as naturally.","A subject-independent or cross-dataset replication would test whether SCG's parity with ECG and BVP extends beyond the 42 EmoWear subjects analyzed here."],"forward_implications":["SCG-based emotion recognition reaches the same macro-F1 range as ECG- and BVP-based recognition for both valence and arousal, so a single accelerometer can reproduce the information that established cardiac pipelines provide.","SCG combined with ADR, both derived from one chest-worn accelerometer, produces SVM and logistic regression classifiers whose F1 distributions are significantly above baseline voting for valence and arousal.","Deep CNN and LSTM models overfit on this dataset size, so classical feature-based classifiers remain the working choice for EmoWear-scale data until larger datasets or pre-training become available.","The EmoWear F1 heatmaps and averaged results provide subject-level benchmarks that future emotion-recognition studies can compare against."],"supporting_citations":[{"why":"Supplies the benchmark DEAP classification pipeline and reference results that the study replicates to validate its implementation.","marker":"[25]"},{"why":"Provides the EmoWear dataset with chest accelerometer, ECG, BVP, and RSP recordings used for all main experiments.","marker":"[18]"},{"why":"Provides the AO-peak detection algorithm (band-pass, Hilbert envelope, second band-pass, peak pick) from which SCG heart-rate-variability features are derived.","marker":"[75]"},{"why":"Supplies the signal-processing and feature-extraction routines from which cardiac and respiratory features are computed.","marker":"[71]"},{"why":"Validates accelerometry-derived respiration against an established respiratory reference, supporting the ADR signal used in the single-sensor combination.","marker":"[22]"},{"why":"Defines the valence-arousal circumplex model that provides the classification label space for the study.","marker":"[12]"},{"why":"Provides the EmoWear CSV data package used in the analysis, including the phase-two elicit-assess-walk cycles.","marker":"[65]"}],"fun_headline_variants":["One chest accelerometer matches ECG and BVP for emotion reading","Chest sensor alone: emotion recognition on par with ECG and BVP","Emotion detection with just a chest accelerometer, matching ECG and BVP","Single chest-worn accelerometer equals ECG and BVP for affect recognition","One chest sensor: emotion recognition on par with ECG and BVP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the vibration peak picked out in each heartbeat stands in reliably for the electrical spike that ECG measures, so that the heart-rhythm variability features computed from the chest accelerometer carry the same emotion information as ECG-derived features; if body movement or sensor shift breaks that peak detection, SCG's parity with ECG and BVP no longer follows.","fun_headline_variants_meta":{"raw":{"variants":["One chest accelerometer matches ECG and BVP for emotion reading","Chest sensor alone: emotion recognition on par with ECG and BVP","Emotion detection with just a chest accelerometer, matching ECG and BVP","Single chest-worn accelerometer equals ECG and BVP for affect recognition","One chest sensor: emotion recognition on par with ECG and BVP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000693,"raw_usage":{"total_tokens":3206,"prompt_tokens":1088,"completion_tokens":2118,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":704,"completion_tokens_details":{"reasoning_tokens":2024}},"tokens_in":704,"tokens_out":2118,"duration_ms":15006,"temperature":1.0,"reasoning_tokens":2024,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:24:47.081670+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the same subjects with a chest-worn accelerometer and a reference ECG while they walk, talk, and move their torso; if the detected vibration peaks lose alignment with the ECG peaks and the resulting heart-rhythm features diverge from ECG-derived features by more than the distance between emotion classes, the central parity claim would be refuted.","supporting_citations":[{"cited_title":"DEAP: A Database for Emotion Analysis; Using Physiological Signals,","cited_arxiv_id":null,"evidence_quote":"Supplies the benchmark DEAP classification pipeline and reference results that the study replicates to validate its implementation."},{"cited_title":"EmoWear: Wearable Physiological and Motion Dataset for Emotion Recognition and Context Awareness,","cited_arxiv_id":null,"evidence_quote":"Provides the EmoWear dataset with chest accelerometer, ECG, BVP, and RSP recordings used for all main experiments."},{"cited_title":"Heart Rate And Heart Rate Variability Indexes Estimated By Mechanical Signals From A Skin- Interfaced IMU,","cited_arxiv_id":null,"evidence_quote":"Provides the AO-peak detection algorithm (band-pass, Hilbert envelope, second band-pass, peak pick) from which SCG heart-rate-variability features are derived."},{"cited_title":"NeuroKit2: A Python toolbox for neurophysiological signal processing,","cited_arxiv_id":null,"evidence_quote":"Supplies the signal-processing and feature-extraction routines from which cardiac and respiratory features are computed."},{"cited_title":"Adaptive Accelerom- etry Derived Respiration: Comparison with Respiratory Inductance Plethysmography during Sleep,","cited_arxiv_id":null,"evidence_quote":"Validates accelerometry-derived respiration against an established respiratory reference, supporting the ADR signal used in the single-sensor combination."},{"cited_title":"A circumplex model of affect,","cited_arxiv_id":null,"evidence_quote":"Defines the valence-arousal circumplex model that provides the classification label space for the study."}],"review_version":1}