{"id":"8fd5f47e-d99d-482b-8013-1e1111f9201d","arxiv_id":"2507.08167","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Emotion intensity in older adults can be predicted from physiological signals with Random Forest achieving R²=0.78, but the evaluation is undermined by test-set hyperparameter tuning and camera-derived labels.","lead":"This paper reports that classical machine learning models can predict the intensity of positive, negative, and neutral emotions in older adults from wrist-worn physiological signals, with the best model (Random Forest) reaching an R-squared of 0.78.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FEA-derived emotion labels are unvalidated as ground truth; unless they track genuine emotion, the central claim that physiological signals alone yield accurate emotion recognition is unsupported.","rationale":"The reader's weakest-assumption identification is exactly the concern I find most load-bearing: the regression targets come from a commercial facial-analysis model with no validation in this population or against any independent measure. This is more fundamental than the hyperparameter-tuning flaw, because even with perfectly honest cross-validation, predicting invalid labels would not constitute emotion recognition. The TSST protocol should produce predictable changes in stress-related negative emotion, so a simple phase-contrast check on the FEA labels would settle the concern. If the labels fail that check, the paper's conclusion is unsupported. The reader's verdict of REJECT is therefore appropriate; the concern is not resolved by anything in the manuscript. I would not raise it to UNVERDICTED because the manuscript's own omissions (e.g., no validation of ground truth, test-set hyperparameter selection) are sufficient to reject the central claim, regardless of data availability.","tokens_in":8826,"tokens_out":5571,"duration_ms":62307,"concrete_test":"Validate the FEA labels on the existing data: for a random subset of participants, compare the FEA continuous intensity scores with participants' self-reported affect (e.g., SAM or PANAS) collected after each TSST phase, or with expert-coded facial expressions from the videos. Additionally, compute the mean FEA negative intensity during the stress phase (T2-T3) versus baseline (T1-T2); it should be significantly higher, while positive intensity should not. If the correlation with self-report/expert coding is low (e.g., r < 0.3) or the stress-phase contrast is absent, the ground-truth labels are invalid and the central claim fails.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim (Section 5: emotion recognition can be accurately achieved using only sensor data) rests entirely on the regression targets generated by iMotion's FEA module (Section 3.1). The paper treats FEA's continuous emotion-intensity scores as objective ground truth without any validation against self-report, expert coding, or known stimulus effects. This is load-bearing because if FEA outputs do not reflect genuine felt emotion in older adults — who may have atypical or reduced facial expressivity — then the models are learning to predict another model's outputs, not emotion. Moreover, TSST is a stress-induction protocol; the paper never shows that the FEA-labeled positive, negative, and neutral intensities vary appropriately across the stress/recovery phases (T1-T5). Without such evidence, the reported R² values (e.g., 0.782 for negative, 0.803 for positive) only quantify fit to pseudo-labels. The hyperparameter selection on test data (Section 3.3) is also a serious validity threat, but it attacks the size of the reported performance; the FEA concern attacks whether the models are predicting emotion at all.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a wearable-sensor-only framework for emotion recognition in older adults. Physiological signals from the Empatica E4 and Shimmer3 GSR+ wristbands are paired with emotion-intensity labels produced by iMotions' Facial Expression Analysis (FEA) module during a Trier Social Stress Test (TSST) protocol. The authors train seven classical regression models plus a DNN and an MLP to predict positive, negative, and neutral emotion intensities, reporting that Random Forest achieves the best R² (0.782 for negative, 0.8033 for positive) and lowest MSE. They conclude that emotion recognition can be accurately achieved using only physiological sensor data, with implications for ADRD and PTSD populations.","tokens_in":9050,"tokens_out":4250,"duration_ms":50837,"significance":"If the results were valid, the contribution would be useful: a 39-participant older-adult dataset, a physiological-only regression pipeline, and an edge-deployable classical-ML baseline are all potentially valuable for privacy-preserving affect monitoring. The paper also takes reasonable precautions in splitting by participant and normalizing after the split. However, two load-bearing issues undermine the central claim: the FEA-derived labels are never validated as ground truth for emotion, and hyperparameters were selected using the test data. As a result, the reported R² and MSE values cannot be interpreted as unbiased estimates of emotion-recognition performance. The significance of the paper therefore depends entirely on whether the target variable itself is meaningful, which the manuscript does not establish.","major_comments":[{"comment":"The ground-truth labels are produced solely by iMotions' FEA module, and the paper states that this 'automated analysis ensures a standardized, objective method for labeling emotions.' No validation is provided against self-report, expert FACS coding, or any independent measure of emotion. This is load-bearing because the central claim—that physiological signals alone can accurately recognize emotion—rests on FEA intensity scores actually reflecting felt emotion. Older adults may show reduced or atypical facial expressivity, and TSST is a stress-induction protocol, so the FEA outputs could be measuring facial muscle movements that do not correspond to the intended emotional states. The authors should provide evidence that FEA labels track emotion in this population, for example by comparing FEA intensities to self-reported affect or to known TSST phase effects.","section":"Section 3.1"},{"comment":"The hyperparameters in Table 3 were, in the authors' words, 'chosen based on the best-performing set on test data after multiple trial and error.' This is a direct use of the test set for model selection, which invalidates the R² and MSE values in Tables 4 and 5 as unbiased estimates of generalization. The comparison between Random Forest and the other models is therefore not a fair out-of-sample evaluation. The authors should use nested cross-validation or a separate validation set for hyperparameter tuning, and then re-report test performance.","section":"Section 3.3"},{"comment":"The TSST is a standardized stress-induction protocol, yet the paper never demonstrates that the FEA-derived positive, negative, and neutral intensity labels vary appropriately across the experimental phases (T1–T5). For instance, one would expect stress-related negative affect to increase during the speech and arithmetic tasks and decrease during recovery, but no such phase-wise analysis is presented. Without this evidence, it is unclear whether the labels have construct validity or whether the models are simply fitting to arbitrary variation in FEA output. This issue is directly tied to the validity of the regression targets in Eq. (1) and Eq. (2).","section":"Section 2 and Section 4"},{"comment":"The reported R² and MSE values are only means across LOSO folds, with no standard deviations, per-fold results, or significance tests. With 39 participants, the differences between Random Forest and the next-best models (e.g., Dense Network) could easily be within noise. The conclusion that 'Random Forest outperforms other models' is therefore not statistically supported. The authors should report fold-level variance and perform pairwise significance tests (e.g., corrected repeated k-fold or Wilcoxon tests).","section":"Tables 4 and 5"}],"minor_comments":[{"comment":"The term 'Leave One Sample Out' is misleading because the procedure actually leaves one subject out; it should be called 'Leave One Subject Out' cross-validation.","section":"Section 3.3"},{"comment":"Both the DNN and MLP use ReLU as the output-layer activation, but regression targets in Table 2 include negative values (e.g., Joy baseline -0.663, Fear -0.842). A ReLU output cannot produce negative predictions, which is inconsistent with the label distribution. A linear output activation should be used.","section":"Table 3"},{"comment":"The description of the 60-sample moving average is incomplete: the manuscript does not specify the sampling rate, window duration, stride, or how these smoothed samples are aggregated into regression instances. This information is needed to assess temporal leakage and reproduce the pipeline.","section":"Section 3.1"},{"comment":"The caption expands BVP as 'Blood Volume Pressure' instead of 'Blood Volume Pulse' (or 'Photoplethysmography'), and 'TMP' as 'Temperature' while the main text uses 'ST' for skin temperature; the notation should be consistent.","section":"Table 1"},{"comment":"Reference [13] is the iMotions FEA product page, not a peer-reviewed validation study. The authors should cite published work on FEA or FACS-based labeling accuracy, especially for older adult populations.","section":"References"},{"comment":"The correlation heatmap is mentioned as showing 'small correlations,' but no correlation values are reported or discussed quantitatively. A brief numerical summary would help the reader interpret why linear models perform poorly.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"The two major threats—unvalidated FEA labels and test-data hyperparameter tuning—are severe enough that the paper's central claim is not supported. If the authors can obtain an independent validation of the labels (e.g., self-report or expert coding) and re-run the experiments with a proper nested-CV protocol, a future revision might be reconsidered. The manuscript could also be repositioned as a dataset or methods-description paper rather than a claim of accurate emotion recognition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth knowing: this is a small empirical paper with a genuinely new dataset—39 older adults under TSST, wristband EDA/BVP/IBI plus camera-derived FEA intensity labels—and the methods are off-the-shelf regressors. The reported Random Forest R² of 0.78–0.80 is the best case for physiological-only emotion regression in this population. But I don't trust those numbers, and the stress-test note is right about the deeper problem.\n\nWhat's genuinely useful: older adults are underrepresented in physiological emotion recognition; TSST is a reasonable stress induction; the LOSO split by participant is correct; they synchronized wearable data with face-based labels; and the feature/emotion correlation heatmap is honest about weak linear relationships.\n\nThe FEA labels are the load-bearing problem. The paper calls them “ground truth” but never validates them against self-report, expert AU coding, or known TSST phase effects. In older adults facial expressivity can be reduced, and FEA's intensity scores are model outputs, so predicting them from wristband signals may just be predicting another model. No evidence shows that positive/negative/neutral intensities track the stress/recovery phases (T1–T5). That validation would be a short experiment, and its absence is telling.\n\nSecond, Section 3.3 states hyperparameters were chosen based on the best-performing set on test data. That invalidates the R²/MSE as unbiased generalization estimates; it inflates reported performance. They report mean scores without variance across LOSO folds; with 39 subjects, fold-to-fold spread matters and is absent.\n\nMinor: no code/data are provided, and self-citations are frequent, though the cited prior work is the relevant literature, not a problem by itself. The abstract says 40 and the body says 39 due to data corruption; that is fine but could be stated more clearly.\n\nBottom line: the dataset may be reusable, but the central claim that emotion recognition is “accurately achieved” is not supported as written. Would I send it to review? Yes—the dataset and population are worth referee time, and the label-validation and evaluation flaws are fixable in principle. As it stands, the claims should be heavily revised.","headline":"New older-adult physiology dataset with FEA-derived emotion labels, but the labels are unvalidated and the models were tuned on the test set, so the headline R² claims are not trustworthy.","tokens_in":9617,"tokens_out":1982,"would_cite":false,"duration_ms":22576,"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":"This paper aims to show that emotion intensity in older adults can be predicted from wrist-worn physiological signals alone, without cameras or facial analysis.","keywords":["emotion detection","older adults","physiological signals","wearable sensors","Trier Social Stress Test","Random Forest regression","edge-based emotion recognition"],"falsifier":"A study that records the same TSST sessions with both facial-expression analysis scores and participants' self-reported emotion intensity, then trains the Random Forest on the self-report labels; if $R^2$ drops far below 0.78 or the two labeling methods correlate weakly, the physiological signals were predicting the facial-expression algorithm's outputs, not felt emotion.","tokens_in":8599,"feed_emoji":"⌚","tokens_out":6938,"duration_ms":69279,"temperature":0.7,"pith_summary":"This paper aims to show that the emotional intensity of older adults can be read from physiological signals captured by wrist-worn sensors, removing the need for cameras, facial analysis, or active participation. The authors report on 39 adults aged 60-80 who underwent the Trier Social Stress Test while wearing two commercial sensor wristbands, with facial-expression analysis used only to create training labels. A Random Forest model then predicted Positive, Negative, and Neutral emotion intensity from the sensor data alone, reaching $R^2$ values of 0.80, 0.78, and 0.76 respectively under leave-one-subject-out cross-validation. If this holds, continuous emotion monitoring could run entirely on-device, with privacy advantages for hospitals, assisted living, and care of people with dementia, PTSD, or limited communication.","feed_headline":"No cameras needed: wrist sensors predict emotion intensity","feed_subtitle":"In 39 adults aged 60-80, sensor data alone explains up to 80 percent of emotion intensity variation.","key_machinery":"The load-bearing mechanism is a synchronized pipeline: the Trier Social Stress Test induces emotional responses; a commercial facial-expression analysis module based on the Facial Action Coding System converts those responses into continuous intensity labels for Positive, Negative, and Neutral emotion; and wrist-worn sensor devices record electrodermal activity, blood volume pulse, inter-beat intervals, temperature, and accelerometer signals. A Random Forest regression model, trained with leave-one-subject-out cross-validation on independently normalized train/test splits, maps the physiological features to the intensity labels. The Random Forest is the component that carries the predictive claim, and the paper credits its ensemble averaging with the generalization that linear models and deep networks do not reach on this small dataset.","core_discovery":"The paper's central claim is that continuous emotion intensity -- specifically Positive, Negative, and Neutral -- can be predicted from wrist-worn physiological signals alone in older adults, with no video or facial-expression input at prediction time. Using data from 39 adults aged 60-80 exposed to the Trier Social Stress Test, the authors report that a Random Forest regressor achieves $R^2$ values of 0.8033 for positive, 0.782 for negative, and 0.7636 for neutral emotion, with mean squared error as low as 0.0006, under leave-one-subject-out cross-validation. They interpret this as evidence that emotion recognition can be accurate using only sensor data, eliminating the need for visual or behavioral inputs, while remaining compatible with edge deployment.","pith_inferences":["If the facial-expression-derived labels are only a proxy for expressed facial movement, the reported accuracy is an upper bound on how well wrist sensors predict that proxy; agreement with felt emotion could be lower.","The same feature set and protocol could plausibly extend from three coarse emotion categories to the finer-grained negative emotions listed in the dataset, since the paper's 12-category labels include sadness, anger, fear, contempt, disgust, confusion, and frustration.","Because the participants were healthy older adults screened for medical conditions, translation to clinical populations with ADRD or PTSD should be validated before deployment."],"forward_implications":["Emotion monitoring in hospitals and assisted-living settings could run without cameras, avoiding the privacy cost of facial surveillance while still tracking a person's emotional state.","Because the top performer is an ensemble of decision trees rather than a deep network, the method can run on low-power edge hardware instead of cloud servers.","Continuous intensity values, not just discrete emotion labels, can be predicted, giving caregivers a graded signal rather than a yes/no emotion tag.","For people who cannot speak or actively report feelings, such as some ADRD patients, passive wrist-worn sensing could provide emotion estimates without requiring participation."],"supporting_citations":[{"why":"Supplies the Trier Social Stress Test protocol used to induce emotional responses in the experimental procedure.","marker":"[4]"},{"why":"The cited facial-expression analysis module provides the automated ground-truth labels that serve as regression targets.","marker":"[13]"},{"why":"Prior evidence that positive and negative emotion can be detected without video, supporting the sensor-only approach.","marker":"[28]"},{"why":"Shows that EDA and heart-rate signals can predict valence and arousal without video, motivating the feature set.","marker":"[23]"},{"why":"Previous physiological-signal work targeting older adults, which this paper extends by predicting three emotion categories.","marker":"[2]"},{"why":"Introduces the WESAD multimodal wearable dataset used as a benchmark context for wearable stress and affect detection.","marker":"[26]"}],"fun_headline_variants":["Wrist sensors alone predict emotion intensity in older adults","No cameras, just wristbands: predicting emotion intensity","Sensor-only emotion detection for older adults achieves 0.78 R²","Older adults' emotions readable from wrist sensors alone"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on the assumption that the facial-expression analysis software's intensity scores really measure how strongly each participant felt the emotion, so the models are trained on true emotion rather than on another algorithm's guesses.","fun_headline_variants_meta":{"raw":{"variants":["Wrist sensors alone predict emotion intensity in older adults","No cameras, just wristbands: predicting emotion intensity","Sensor-only emotion detection for older adults achieves 0.78 R²","Older adults' emotions readable from wrist sensors alone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1382,"prompt_tokens":933,"completion_tokens":449,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":382}},"tokens_in":549,"tokens_out":449,"duration_ms":5090,"temperature":1.0,"reasoning_tokens":382,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:25:22.770937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A study that records the same TSST sessions with both facial-expression analysis scores and participants' self-reported emotion intensity, then trains the Random Forest on the self-report labels; if $R^2$ drops far below 0.78 or the two labeling methods correlate weakly, the physiological signals were predicting the facial-expression algorithm's outputs, not felt emotion.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Trier Social Stress Test protocol used to induce emotional responses in the experimental procedure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The cited facial-expression analysis module provides the automated ground-truth labels that serve as regression targets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior evidence that positive and negative emotion can be detected without video, supporting the sensor-only approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that EDA and heart-rate signals can predict valence and arousal without video, motivating the feature set."},{"cited_title":"A Novel Multimodal System to Predict Agitation in People with Dementia Within Clinical Settings: A Proof of Concept","cited_arxiv_id":"2411.08882","evidence_quote":"Previous physiological-signal work targeting older adults, which this paper extends by predicting three emotion categories."}],"review_version":1}