{"id":"2214fb00-cd7d-41ec-ab94-a648d28441aa","arxiv_id":"2507.08175","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors report that a simulated 4-qubit quantum kernel SVM achieves over 80% F1 in three-class emotion recognition from wearable physiological signals in 39 older adults, claiming improvements over classical models.","lead":"This paper reports a comparison of classical and quantum-style machine learning models for recognizing emotions from wristband sensor data in 39 older adults. The authors claim the quantum-kernel SVM, simulated on a normal computer, outperforms classical models in most emotion classes, with F1 scores above 0.80.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The comparison is invalid as reported: hyperparameters are tuned on the test set, and Table V contradicts the claim of superiority across all emotion classes.","rationale":"The reader's weakest assumption is that the iMotions FEA labels are valid ground truth for older adults. I agree that this is a serious problem: Section III-A derives labels from a commercial computer-vision tool with no human validation, and reduced facial expressivity in older adults makes age bias plausible. However, I find an even more immediate, internal problem that does not depend on external label validity. The paper admits in Section III-B3 that hyperparameters were selected using the test set, which invalidates the reported generalization numbers. In addition, Tables IV and V show the quantum SVM's Neutral F1 (0.775 at 640 training samples, 0.824 at 1600) below classical Random Forest (0.851 at 2000 samples), directly contradicting the claim that the quantum model surpasses classical counterparts across all emotion categories. Thus the central claim fails on the paper's own evidence. The proposed concrete test uses nested cross-validation and matched sample sizes to determine whether any quantum advantage survives a valid comparison; if it does not, rejection is warranted. This does not change the reader's REJECT verdict.","tokens_in":8809,"tokens_out":8033,"duration_ms":88390,"concrete_test":"Re-run the full comparison with nested 5-fold cross-validation: select hyperparameters on the training folds only, then evaluate on the held-out test split. Report per-class F1, precision, recall with mean and standard deviation for the quantum kernel SVM and all classical baselines at matched training sizes (e.g., 640 and 1600, plus the 2000-sample configuration used for classical baselines). If the quantum SVM's Neutral-class F1 does not exceed Random Forest, or if the 36% recall improvement and >80% F1 values disappear, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B3 states that model hyperparameters were chosen based on the best-performing set on test data after multiple trials and errors. Selecting on the test set invalidates every generalization estimate in Tables IV and V, so the quantum-vs-classical comparison is not a valid comparison of fitted models. Independently, the paper's own numbers contradict the headline: at train:640 the quantum SVM reports Neutral F1 = 0.775 (Table V), below Random Forest's Neutral F1 = 0.851 (Table IV), and at train:1600 the quantum Neutral F1 = 0.824 is still below the RF value. Because the central claim asserts improvement across all emotion categories, the reported tables alone are sufficient to refute it. The unvalidated iMotions FEA labeling (Section III-A) is an additional serious threat, but no external validation is needed to see that the stated conclusion does not follow from the paper's own results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a physiological-signal-only emotion recognition pipeline for older adults. Data from 39 participants under the Trier Social Stress Test are labeled as Neutral, Positive, or Negative using iMotions Facial Expression Analysis, and several classical classifiers are compared with a hybrid quantum-kernel SVM (4-qubit Belis feature map, simulated with Qiskit's AerSimulator). The authors report F1 scores in Tables IV and V and claim the quantum SVM outperforms classical baselines in every emotion class, with F1 scores over 80% and up to about 36% improvement in recall.","tokens_in":8986,"tokens_out":5270,"duration_ms":65247,"significance":"If the reported results held, the study would be a useful application of wearable sensing to an understudied population and would support privacy-preserving affect monitoring. The participant-level data split and the focus on older adults are commendable, and the kernel construction follows prior work. However, the central quantitative claim is not supported by the reported evaluation: hyperparameters were selected on the test set, and the paper's own tables contradict the 'all emotion categories' claim. As it stands, the contribution is an exploratory pipeline with unvalidated labels and a classically simulated quantum kernel, rather than a demonstration of quantum-enhanced emotion recognition.","major_comments":[{"comment":"The paper states that 'the values for these parameters were chosen based on the best-performing set on test data after multiple trials and errors.' This means the test set was used for model selection, so every accuracy, precision, recall, and F1 value in Tables IV and V is an optimistically biased estimate and is not a valid estimate of generalization. The central quantum-versus-classical comparison is therefore invalid: apparent advantages could arise from selection effects rather than from the quantum kernel. The evaluation must be redone with hyperparameters tuned on a validation split or via nested cross-validation, with the procedure reported explicitly.","section":"Section III-B3"},{"comment":"The abstract and Section V claim that the quantum-enhanced SVM surpasses classical counterparts in classification performance across all emotion categories. This is contradicted by the paper's own numbers: Random Forest achieves a Neutral F1 of 0.851, while the quantum SVM achieves only 0.775 at train:640 and 0.824 at train:1600. The abstract's statement that 'F1 scores over all classes are over 80%' is also false for the 640-sample quantum model, whose Neutral F1 is 0.775. The headline conclusion therefore does not follow from the reported results.","section":"Tables IV-V"},{"comment":"The ground-truth labels are derived solely from iMotions Facial Expression Analysis, with no validation against self-report, clinical assessment, or human FACS coding. This is a serious threat to validity, especially for older adults, whose facial expressivity is often reduced. If the commercial FEA labels are noisy or biased for this population, the classifiers are trained to predict an invalid target, and all reported F1 scores are conditional on that target. At minimum, a small human-coded validation subset or an explicit label-noise analysis is needed before the emotion-recognition claims can be interpreted.","section":"Section III-A"}],"minor_comments":[{"comment":"The index terms ('stress prediction, RNN, LSTM, cortisol, wearables') do not match the paper's content and should be replaced with terms reflecting emotion recognition, quantum kernels, and wearable physiological sensing.","section":"Index Terms"},{"comment":"The header 'train: 1600; test: 4000' appears to be a typo for test: 400, since 2000 samples with an 80/20 split would give 400 test samples. Please clarify the sample-count notation throughout.","section":"Table V"},{"comment":"The claim of 'around a maximum of 36% increase in the recall values' is not tied to a specific table entry. Using Tables IV and V, the recall gains over Random Forest are approximately 39.6% relative for Negative, 30.5% for Positive, and 9.6% for Neutral at train:1600; the exact comparison should be stated.","section":"Section IV"},{"comment":"The quantum kernel is evaluated with a classical statevector simulation (AerSimulator), so the phrase 'quantum-enhanced' should be qualified; the results are a classical simulation of a quantum kernel, not a demonstration on quantum hardware.","section":"Section III-B3"},{"comment":"The baseline values for Joy and Positive are identical, as are those for Anger and Negative, and the thresholding procedure ('outside 1st standard deviation') is not fully explained. Please clarify how the three emotion classes are aggregated from the FEA outputs and how the threshold is applied.","section":"Table II"}],"recommendation":"reject","confidential_remarks":"The manuscript's central empirical claim is invalidated by test-set hyperparameter selection and is contradicted by its own tables, so I do not see how it can meet the bar for publication without a fundamentally revised evaluation. If the authors re-run the study with proper validation, address label validity for older adults, and narrow their claims, a new submission could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper reports a new dataset of physiological signals from 39 older adults under the TSST protocol, labeled with iMotions FEA, and compares classical models to a quantum-kernel SVM. The dataset is a real contribution, and the privacy-preserving framing is sensible. But the central comparison is not valid as reported: Section III-B3 states that hyperparameters were chosen based on best performance on the test data after multiple trials, which invalidates every generalization estimate. On top of that, the abstract's claim of superiority across all emotion categories is directly contradicted by the paper's own tables: at both training sizes, the quantum SVM's Neutral F1 (0.775 and 0.824) is below Random Forest's 0.851. A 4-qubit simulated kernel may or may not help, but this paper doesn't show it.\n\nWhat's genuinely new: the application of the Belis feature map to wearable sensor data from an older-adult population, with a TSST protocol and FEA labels. That combination hasn't been published. The authors also split by participant before normalization, which is correct, and they report per-class precision/recall/F1 rather than just accuracy. Those are good instincts.\n\nThe soft spots beyond the two big ones: the ground-truth labels come solely from a commercial computer-vision tool with no validation against self-report or human coding for older faces, which may show reduced expressivity; the emotion intensity thresholding (outside 1st standard deviation) is ad hoc; and there is no code/data or confidence intervals. These are serious but the first two are enough to sink the current claims.\n\nWho should read it: people working on wearable affect recognition in aging populations and anyone curious about how quantum kernels fare on small physiological datasets. The dataset itself could be reusable if shared, but the current results should not be cited as evidence for QML advantage.\n\nMy recommendation: a serious referee should see this, because the question is legitimate and the dataset is potentially valuable, but the paper needs a major revision: proper nested cross-validation or a fixed validation set, corrected claims, and ideally label validation. I would not accept it in its present form.","headline":"A useful new physiological dataset from older adults is undermined by test-set hyperparameter tuning and a central claim contradicted by the paper's own tables.","tokens_in":9539,"tokens_out":3365,"would_cite":false,"duration_ms":34751,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantum SVM beats classical models on emotion from wearables","keywords":["emotion recognition","quantum machine learning","quantum kernel","support vector machine","wearable sensors","physiological signals","older adults","privacy-preserving"],"falsifier":"Re-annotate a random subset of the recorded faces with human FACS coding or participant self-report, retrain on the same physiological features, and re-run the 5-fold evaluation; if the quantum SVM's F1 scores against those labels fall to classical levels or to chance, the reported quantum advantage was an artifact of the automatic labeler.","tokens_in":8591,"feed_emoji":"🧠","tokens_out":9124,"duration_ms":85034,"temperature":0.7,"pith_summary":"This paper claims that emotional states can be read from wrist-worn physiological signals alone, with no camera or facial imagery, using a hybrid quantum-classical machine learning model. On data from 39 older adults undergoing a stress protocol, the authors report that a support vector machine with a quantum kernel beats seven classical models across positive, negative, and neutral classes, with F1 scores reaching 0.905 for the positive class on only 640 training samples and recall gains up to 36 percent. The stated payoff is a privacy-preserving way to monitor emotional well-being in older adults and in clinical populations such as people with Alzheimer's disease and related dementias or veterans with PTSD, where speech and facial expression may be unreliable. The paper's significance, if the result holds, is that small physiological datasets can be handled by quantum kernels to produce accurate, unobtrusive affect sensing.","feed_headline":"Quantum SVM beats classical models on emotion from wearables","feed_subtitle":"F1 reaches 0.90 on positive emotion, with up to 36 percent better recall than classical models.","key_machinery":"The central carrying object is a fidelity-based quantum kernel. Each physiological sample is encoded by a circuit $U(x)$ that applies parameterized $SU(2)$ rotations to four qubits, with two classical features mapped per qubit, and then entangles neighboring qubits with CNOT gates; the similarity between two samples is the squared fidelity $k(x_i,x_j)=|\\langle 0|U^\\dagger(x_i)U(x_j)|0\\rangle|^2$. This kernel matrix feeds a classical support vector machine, so the model is hybrid: the quantum circuit defines the geometry of the feature space, and classical optimization solves the margin problem. The kernel matrices are evaluated on a classical simulator, so the reported results do not depend on quantum hardware noise.","core_discovery":"The paper's central discovery is that a 4-qubit quantum kernel SVM, with data encoded through a feature map that places two classical features per qubit and entangles neighboring qubits, classifies emotion from wearable physiological data better than any of seven classical baselines. On a test set held out by participant, the quantum model achieves F1 scores of 0.851, 0.775, and 0.905 for negative, neutral, and positive emotion with only 640 training samples, improving to 0.872, 0.824, and 0.897 with 1,600 training samples; the best classical model, Random Forest, peaks at 0.685, 0.851, and 0.761 on the same 2,000-sample split. The authors interpret the result as evidence that quantum kernels can capture nonlinear correlations in physiological time series that classical kernels miss, and that emotion recognition need not rely on facial imagery at any point in the pipeline.","pith_inferences":["Inference: Because the quantum kernel is evaluated on a classical simulator, the paper demonstrates an advantage of a particular feature map, not of quantum hardware; the same kernel on real hardware might behave differently due to noise.","Inference: A matched comparison that tunes classical RBF-SVM hyperparameters and feature preprocessing as extensively as the quantum kernel would clarify whether the 36 percent recall gain comes from the quantum feature map or from unequal tuning.","Inference: The practical value hinges on label validity; a re-labeling study with self-report or clinical ratings on a subset would test whether the reported accuracies transfer.","Inference: The same quantum kernel could be applied to other wearable affect datasets, such as stress or cognitive-load recordings, to see whether the advantage is specific to this labeling protocol."],"forward_implications":["If the result holds, emotion monitoring in assisted-living and clinical settings can be done with wristbands alone, removing the privacy cost of cameras.","The reported performance at 640 training samples suggests quantum kernels can work on small, participant-limited datasets, which is the norm for clinical populations.","The 36 percent recall improvement implies fewer missed emotional events, which matters for detecting distress or agitation in people who cannot communicate verbally.","The approach is a stepping stone toward real-time cloud-based inference from wearable streams, assuming the kernel computation can be made fast enough.","The same pipeline, if extended beyond the three coarse classes, could provide finer-grained affect labels without changing the sensing hardware."],"supporting_citations":[{"why":"Describes the stress-inducing protocol applied to participants to elicit emotional responses during data collection.","marker":"[28]"},{"why":"Provides the automated facial expression analysis module that generates the ground-truth emotion labels for the physiological training data.","marker":"[29]"},{"why":"Introduces the quantum feature map and fidelity kernel used by the hybrid SVM to encode and compare physiological samples.","marker":"[30]"}],"fun_headline_variants":["Quantum kernel SVM beats classical for emotion from wearables","Hybrid quantum kernel outperforms classical on emotion from wearables","Quantum SVM achieves up to 36% better recall on emotional states","Quantum kernel model reads emotions from wearable sensors without cameras"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire evaluation depends on the assumption that the automated facial expression analysis tool used for labeling gives correct emotion labels for these older adults; no self-report, clinical assessment, or human coding checks that assumption.","fun_headline_variants_meta":{"raw":{"variants":["Quantum kernel SVM beats classical for emotion from wearables","Hybrid quantum kernel outperforms classical on emotion from wearables","Quantum SVM achieves up to 36% better recall on emotional states","Quantum kernel model reads emotions from wearable sensors without cameras"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000988,"raw_usage":{"total_tokens":4167,"prompt_tokens":901,"completion_tokens":3266,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":3198}},"tokens_in":517,"tokens_out":3266,"duration_ms":26889,"temperature":1.0,"reasoning_tokens":3198,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:25:05.987151+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-annotate a random subset of the recorded faces with human FACS coding or participant self-report, retrain on the same physiological features, and re-run the 5-fold evaluation; if the quantum SVM's F1 scores against those labels fall to classical levels or to chance, the reported quantum advantage was an artifact of the automatic labeler.","supporting_citations":[{"cited_title":"The trier social stress test protocol for inducing psy- chological stress,","cited_arxiv_id":null,"evidence_quote":"Describes the stress-inducing protocol applied to participants to elicit emotional responses during data collection."},{"cited_title":"Facial expression analysis (fea),","cited_arxiv_id":null,"evidence_quote":"Provides the automated facial expression analysis module that generates the ground-truth emotion labels for the physiological training data."},{"cited_title":"Quantum anomaly detection in the latent space of proton collision events at the lhc,","cited_arxiv_id":null,"evidence_quote":"Introduces the quantum feature map and fidelity kernel used by the hybrid SVM to encode and compare physiological samples."}],"review_version":1}