{"id":"558a9279-549e-4b8b-874e-30fd707700bf","arxiv_id":"2607.06509","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"Three custom TENG sensors integrated into eyeglasses capture arterial pulse, jaw kinematics, and facial activity at 4.1 µW total front-end power, achieving 93.8% activity accuracy and 1.82 BPM heart rate error in a 20-participant study.","lead":"GlassTENG embeds three triboelectric nanogenerator sensors into eyeglasses to capture pulse, jaw, and facial muscle activity at 1.36 µW per channel. A smart generalist might read it because it demonstrates a path toward battery-free, continuous physiological monitoring from a device billions already wear.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The 1.82 BPM heart-rate MAE relies on per-user oracle sensor selection (§4.2), inflating the headline quantitative claim beyond what a deployable fixed-sensor configuration would achieve.","rationale":"The reader's CONDITIONAL verdict is appropriate. The paper is a legitimate engineering contribution with honest limitations. My concern about oracle sensor selection is a real methodological issue that inflates the HR claim, but it does not invalidate the core contribution — the paper does not overclaim deployability and explicitly frames battery-free operation as future work. The per-user sensor selection should be more prominently disclosed as a caveat on the 1.82 BPM number, but the overall verdict of CONDITIONAL with MODERATE confidence already captures the appropriate level of caution. The reader's identified weakest assumption (contact reliability during movement) is also valid but is a forward-looking concern about untested conditions rather than a problem with the data as presented. Both concerns point in the same direction: the validation scope is narrower than the framing suggests, which is exactly what CONDITIONAL captures. No verdict adjustment needed.","tokens_in":16561,"tokens_out":1926,"duration_ms":50570,"concrete_test":"Recompute HR MAE for each sensor individually across all 10 Study 2 participants (S1-only, S2-only, S3-only) using the same 30-second sliding-window methodology. Compare each per-sensor MAE to the oracle-selected 1.82 BPM. If the best single fixed sensor yields MAE > 4 BPM, the headline claim is substantially inflated by the selection procedure and should be reported as 'best-of-three' rather than as a system-level accuracy. Additionally, test whether peak prominence (the selection criterion used) can identify the optimal sensor without ground-truth HR: compute the correlation between per-sensor peak prominence ranking and per-sensor MAE ranking across participants. If the correlation is weak, automatic sensor selection is not viable and the oracle result is not achievable in deployment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly notes per-user optimal sensor selection as a limitation but identifies sensor-skin contact reliability as the weakest assumption. I think the sensor selection issue is more load-bearing because it directly inflates the reported quantitative HR claim with data already in the paper. §4.2 states: 'We selected the optimal sensor channel for each user based on peak prominence to maximize heart rate estimation accuracy.' This is oracle selection: for each of the 10 participants, the authors post hoc chose whichever of S1, S2, or S3 yielded the best HR estimate, using peak prominence as the selection criterion. The reported 1.82 BPM MAE and Bland-Altman limits (−5.17 to +5.58 BPM) reflect this best-of-three scenario. In a real deployment, the system has no ground-truth HR to validate which sensor is optimal, and no automatic selection mechanism is proposed or validated. If the system had to commit to a single sensor site for all users — or even select per-user without ground truth — the MAE would likely be meaningfully higher. The paper does not report per-sensor HR MAE, so we cannot assess how much the oracle selection inflates the headline number. This is distinct from the contact-reliability concern (which is speculative — about untested conditions) because it is a demonstrable methodological choice that affects the central quantitative claim as presented. The 93.8% activity classification accuracy, by contrast, uses all three sensors simultaneously in a LOSO framework and is methodologically sound, though the Talk/Eat conflation (55% Talk accuracy) is a real practical limitation the paper honestly acknowledges.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper presents GlassTENG, a system integrating three custom-fabricated triboelectric nanogenerator (TENG) sensors into a glasses frame at the angular artery (nasal bridge), superficial temporal artery (anterior temple), and temporalis muscle (posterior temple). Each sensor operates through a vertical contact-separation mechanism using PDMS and FEP layers with silver electrodes, paired with a 1.36 µW-per-channel high-impedance analog front-end (TLV8802 op-amp). The system simultaneously captures arterial pulse waveforms and classifies six jaw/upper facial activities. A 20-participant study (Study 1) evaluates activity recognition via LOSO cross-validation, achieving 93.8% accuracy with a Random Forest classifier across seven classes (including No Activity). A 10-participant subset (Study 2) validates heart rate estimation against a Polar H10 chest strap, reporting 1.82 BPM MAE with Bland-Altman limits of agreement from −5.17 to +5.58 BPM. The paper positions GlassTENG as a step toward battery-free, longitudinally worn physiological sensing from eyewear.","tokens_in":17388,"tokens_out":1346,"duration_ms":153564,"significance":"The paper makes a solid hardware contribution in demonstrating that custom TENG sensors can transduce sub-Newton physiological forces (arterial pulse at ~0.01 N, facial muscle activity at 0.05–2 N) into usable electrical signals at three anatomically distinct facial sites, with a measured front-end power consumption of 4.1 µW total — more than two orders of magnitude below the PPG/EMG alternatives catalogued in Table 1. The sensor fabrication process (§3.1, Figure 2A) is described with sufficient detail for replication, including plasma etching parameters and PVD deposition specifications. The 20-participant LOSO evaluation for activity classification follows standard practice, and the Bland-Altman analysis for HR validation is appropriate. The power comparison in Table 1 and the battery-life analysis in §6.1 provide useful context for the energy budget argument. The Talk/Eat confusion (55% accuracy for Talk, 37% misclassified as Eat) is honestly reported and discussed in §6.3.","major_comments":[{"comment":"§4.2 (Study 2, Data Processing): The HR validation uses oracle per-participant sensor selection: 'We selected the optimal sensor channel for each user based on peak prominence to maximize heart rate estimation accuracy.' This is a post hoc best-of-three selection that requires knowledge of which sensor yields the best HR estimate. In deployment, no ground-truth HR is available to perform this selection, and no automatic selection mechanism is proposed or validated. The headline 1.82 BPM MAE and Bland-Altman limits (−5.17 to +5.58 BPM) therefore reflect an upper bound on achievable accuracy. The paper does not report per-sensor HR MAE (e.g., S1-only, S2-only, S3-only across all participants), making it impossible to assess the inflation magnitude or whether a single fixed sensor would yield acceptable accuracy. This is load-bearing for the central quantitative HR claim. The authors should","section":null},{"comment":"§5.2, Table 2, Figure 7: The activity classification is described as 'six jaw and upper facial activities' in the abstract and contributions, but Table 2 and Figure 7 include seven classes (the sixth being 'No Activity'). The 93.8% accuracy is a 7-class figure. This should be clarified so that readers understand the headline figure includes a baseline class, and the per-class performance on the six target activities should be reported separately to allow comparison with prior work.","section":null}],"minor_comments":[{"comment":"Abstract and §1 (contributions list): The phrase 'six facial activity classes' should be reconciled with the seven-class setup in Table 2. Either clarify that 'No Activity' is excluded from the 'six' count or adjust the framing.","section":null},{"comment":"§3.3: The sensor characterization (Figure 2C) maps force to voltage but does not specify the number of trials or measurement variability. Adding error bars or confidence intervals would strengthen the characterization data.","section":null},{"comment":"§4.1: The HR validation study (Study 2) used only 10 participants in seated, resting conditions. While this is noted, the abstract and conclusion present 1.82 BPM MAE without this context. Consider qualifying the headline claim or noting the sample size in the abstract.","section":null},{"comment":"§3.4: The bias voltage (V_bias) and voltage divider ratio for S3 are mentioned but specific values are not provided in the text. These should be stated for reproducibility.","section":null},{"comment":"Table 1: The power values for comparison sensors are marked with an asterisk as 'estimated front-end power values based on components used.' The basis for these estimates should be briefly cited (e.g., specific component datasheets or prior measurements).","section":null},{"comment":"§5.3: The user experience survey reports a 'public comfort rating of 5.5/7' for a 'lighter, compact version,' but the actual prototype used was a protoboard. This distinction should be made clearer so readers understand the ratings are for a hypothetical form factor, not the actual prototype.","section":null},{"comment":"Figure 1: The figure caption lists sensor sites as 'nasal bridge, posterior temple, anterior temple' but the text in §3.2 refers to 'angular artery,' 'superficial temporal artery,' and 'temporalis muscle.' Aligning the anatomical and sensor-site terminology across figures and text would improve readability.","section":null},{"comment":"§6.1: The battery life comparison states GlassTENG's front-end 'would cost the battery less than half a minute of its lifetime' but the calculation is not shown. A brief derivation (battery capacity, current draw, resulting time) would help readers verify the claim.","section":null}],"recommendation":"major_revision","confidential_remarks":"The oracle sensor selection issue (§4.2) is the most substantive concern. The authors have the data to address it: they can report per-sensor HR MAE and either propose a ground-truth-free selection method or present fixed-sensor results. This should be feasible within revision. The Talk/Eat confusion is honestly reported and not a blocker, but the authors might consider whether additional sensor features or a seventh sensor modality could help — though that is beyond revision scope. The paper is a reasonable fit for the journal's HCI/sensing scope."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. Both major comments identify legitimate issues that we will address in revision. Comment 1 (oracle sensor selection in HR validation) is correct that our current methodology reflects an upper bound on achievable accuracy; we will add per-sensor HR results and an automatic selection mechanism. Comment 2 (seven-class vs. six-class framing) is a valid clarification point; we will revise the abstract/contributions and report per-class metrics separately.","responses":[{"response":"The referee is correct that our current HR validation uses oracle per-participant sensor selection, and that the headline 1.82 BPM MAE therefore represents an upper bound on achievable accuracy in deployment. We agree this needs to be addressed transparently. In the revision we will: (1) report per-sensor HR MAE (S1-only, S2-only, S3-only) across all 10 Study 2 participants, so readers can assess the inflation magnitude and whether a single fixed sensor yields acceptable accuracy; (2) explicitly label the current 1.82 BPM figure as an oracle-selected upper bound in the text and in the Bland-Altman figure caption; and (3) propose and evaluate a simple automatic sensor-selection mechanism based on peak prominence computed from the TENG signal itself (no ground-truth HR required), which is the same criterion used for oracle selection but applied without knowledge of the reference. We will report the MAE under this automatic selection alongside the per-sensor and oracle results. We note that the automatic selection criterion is already signal-based (peak prominence), so a deployment-time implementation is straightforward; we simply did not validate it separately in the current manuscript, which was an oversight.","revision_made":"yes","referee_comment":"§4.2 (Study 2, Data Processing): The HR validation uses oracle per-participant sensor selection... The headline 1.82 BPM MAE and Bland-Altman limits therefore reflect an upper bound on achievable accuracy... The paper does not report per-sensor HR MAE... The authors should [report per-sensor results and address deployment-time selection]."},{"response":"The referee is correct. The abstract and contributions list six activities, but the classification evaluation includes a seventh 'No Activity' baseline class, and the 93.8% accuracy is a 7-class figure. This mismatch is an oversight in how we framed the headline result. In the revision we will: (1) clarify in the abstract, contributions, and §5.2 that the 93.8% accuracy is a 7-class figure including the No Activity baseline; (2) report the 6-class accuracy (excluding No Activity) separately in Table 2 so that direct comparison with prior work reporting only target-activity classes is possible; and (3) add per-class precision, recall, and F1 for the six target activities in a supplementary table or alongside the existing confusion matrix. The confusion matrix in Figure 7 already contains the per-class information, but we will make the 6-class vs. 7-class distinction explicit in the text and table captions.","revision_made":"yes","referee_comment":"§5.2, Table 2, Figure 7: The activity classification is described as 'six jaw and upper facial activities' in the abstract and contributions, but Table 2 and Figure 7 include seven classes (the sixth being 'No Activity'). The 93.8% accuracy is a 7-class figure. This should be clarified so that readers understand the headline figure includes a baseline class, and the per-class performance on the six target activities should be reported separately."}],"tokens_in":16478,"tokens_out":783,"duration_ms":76604,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper puts three custom TENG sensors on a glasses frame at three anatomically targeted sites, pairs them with a genuinely low-power analog front-end (1.36 µW/channel), and validates the combination in a 20-person study. That is a real engineering contribution. The activity classification (93.8%, LOSO, Random Forest) is methodologically sound — all three sensors used simultaneously, proper cross-validation. The sensor fabrication is well-documented, the force-voltage characterization is thorough, and the power budget argument is credible. The paper is also honest about its limitations, which I appreciate. Qu et al. [37] put a TENG on glasses for hemifacial spasm detection and incidentally caught pulse; GlassTENG's contribution is the deliberate multi-site design, the co-designed front-end, and the systematic activity + HR validation. That is a legitimate delta. Now the soft spots. The stress-test concern about oracle sensor selection for HR is the one that lands hardest. §4.2 says they selected the optimal sensor channel per user based on peak prominence to maximize HR accuracy. That is best-of-three post hoc selection with no automatic mechanism proposed for deployment. The 1.82 BPM MAE and the Bland-Altman limits reflect this oracle scenario. The paper does not report per-sensor HR MAE, so we cannot tell how much worse a fixed-sensor configuration would be. This inflates the headline number beyond what a deployable system would achieve, and it is the central quantitative claim. The reader flagged sensor-skin contact reliability as the weakest assumption. That is a real concern but it is speculative — about untested conditions. The oracle selection issue is demonstrable from the paper itself and more load-bearing for the HR claim. The Talk/Eat conflation (55% Talk accuracy) is a real practical limitation the paper acknowledges honestly. HR validation being limited to 10 seated participants is narrow but acceptable for a proof-of-concept. No code or data shared is a minus. The paper validates the analog front-end only, not the full battery-free vision — but they say this clearly in §6.3 and do not overclaim. This paper is for the UbiComp/IMWUT community — researchers working on wearable sensing, low-power systems, and TENG-based interfaces. It deserves a serious referee who can evaluate the sensor design and the HR methodology. The oracle selection issue should be a required revision point: either report per-sensor MAE or propose and validate an automatic selection mechanism. Recommend accept for peer review.","headline":"Multi-site TENG sensors on glasses with 1.36 µW/channel front-end; HR claim uses oracle sensor selection","tokens_in":17634,"tokens_out":592,"would_cite":true,"duration_ms":69425,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Glasses Sense Pulse and Jaw Motion on 4 Microwatts","keywords":["triboelectric nanogenerator","self-powered sensing","smart glasses","wearable sensors","pulse sensing","activity recognition","ultra-low-power electronics","analog front-end"],"falsifier":"If participants wearing the glasses during normal ambulatory activity (walking, talking, head turning) produce TENG signals where the pulse waveform is unrecognizable or activity classification drops below ~80%, the claim that this is a viable platform for longitudinal physiological monitoring would be undermined.","tokens_in":16612,"feed_emoji":"","tokens_out":1190,"duration_ms":112564,"temperature":0.7,"pith_summary":"This paper asks whether rich physiological sensing — arterial pulse, jaw kinematics, and upper facial activity — can be extracted from everyday glasses using only a few microwatts of power, roughly three orders of magnitude less than conventional optical or electrode-based sensors. The authors build GlassTENG, a system of three custom triboelectric nanogenerator (TENG) sensors embedded at the nasal bridge, anterior temple, and posterior temple of a glasses frame. Each TENG converts the tiny mechanical forces of arterial pulsation and muscle movement directly into electrical signals, eliminating the need for powered emitters or gels. A high-impedance analog front-end draws 1.36 microwatts per channel (4.1 microwatts total), conditioning these self-generated voltage signals for digitization. In a 20-participant study, the system classified six jaw and upper facial activities with 93.8% accuracy using leave-one-subject-out cross-validation, and estimated heart rate with 1.82 BPM mean absolute error against a chest-strap reference across 10 participants in resting conditions. The central mechanism is the co-design of flexible PDMS/FEP triboelectric layers — sensitive to forces as small as 0.01 N — with an 18-teraohm-input front-end that can read the resulting high-impedance signals at negligible power cost. The paper positions this as a sensing front-end proof-of-concept: the TENGs are self-powered for transduction, but the downstream microcontroller and any energy harvesting for fully battery-free operation remain future work.","feed_headline":"Self-Powered Glasses Sensors Capture Pulse and Jaw Activity at 4 Microwatts","feed_subtitle":"Three triboelectric nanogenerators on a glasses frame classify six facial activities at 93.8% accuracy and track heart rate within 1.82 BPM,","key_machinery":"PDMS/FEP triboelectric nanogenerator sensors with silver electrodes, operating in vertical contact-separation mode, read by a TLV8802-based unity-gain amplifier with 18 TΩ input impedance at 1.36 µW per channel","core_discovery":"The core finding is that triboelectric nanogenerator sensors, fabricated from PDMS and FEP films with silver electrodes, can be integrated into a glasses frame at three anatomically targeted sites — the angular artery, superficial temporal artery, and temporalis muscle — and produce voltage signals (50 mV to 1.5 V) of sufficient quality to simultaneously recover arterial pulse waveforms and classify six facial/jaw activities, all while the analog front-end consumes only 1.36 microwatts per channel. This is three orders of magnitude below the milliwatt-level power draw of optical PPG, EMG, or load-cell sensors used in prior eyewear systems. The 20-participant validation demonstrates that the ","pith_inferences":["The Bland-Altman limits of agreement (−5.17 to +5.58 BPM) are wide enough that clinical-grade heart rate monitoring would likely require longer averaging windows or motion-artifact rejection beyond what seated resting conditions validate; real-world ambulatory use may degrade these numbers substantially.","The nonlinearity of the TENG force-voltage response across the 0.01–5 N range means that the mapping from sensor voltage to biomechanical force is participant-dependent; per-participant normalization was used here, but a calibration-free deployment would need a model of how facial anatomy varies this transfer function across populations.","If S2 (the superficial temporal artery sensor on a dedicated downward arm) is eliminated as the authors suggest, the system loses one of three pulse sites; whether two-site pulse capture still supports the multi-site blood pressure pathway the paper envisions is an open question."],"forward_implications":["If the sensing chain holds during real-world movement, glasses-based cardiovascular monitoring could operate without the continuous power drain of optical emitters, potentially extending smart-glasses battery life by hours compared to adding PPG or EMG.","Multi-site pulse waveform capture from the angular and superficial temporal arteries could enable pulse transit time measurements between facial arterial sites, opening a pathway to cuffless blood pressure estimation from eyewear.","The 93.8% activity classification accuracy, with the primary confusion between talking and eating, suggests that adding a fourth sensor or richer temporal features could disambiguate oral activities — a capability relevant to dietary monitoring and bruxism detection.","The 1.36 µW-per-channel front-end budget is small enough that a glasses frame covered with indoor solar cells harvesting 200–300 µW could plausibly power the entire sensing chain, making fully battery-free physiological eyewear a tractable engineering target."],"fun_headline_variants":["Self-Powered Glasses Sensors Track Pulse and Facial Activity at 1.36 Microwatts","Glasses-Mounted TENG Sensors Capture Pulse and Jaw Motion at Microwatt Scale","Triboelectric Glasses Sensors Track Heart Rate and Facial Activity","Microwatt-Powered Glasses Sensors Detect Pulse and Jaw Movement","Self-Powered Eyewear Sensors Map Pulse and Facial Activity Without Gels"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The entire signal chain depends on consistent, comfortable mechanical coupling between each TENG sensor and the skin at three facial sites; the 20-participant validation was conducted in seated, resting conditions, and whether the sensors maintain adequate contact and signal quality during walking, head movement, or prolonged wear is not established.","fun_headline_variants_meta":{"raw":{"variants":["Self-Powered Glasses Sensors Track Pulse and Facial Activity at 1.36 Microwatts","Glasses-Mounted TENG Sensors Capture Pulse and Jaw Motion at Microwatt Scale","Triboelectric Glasses Sensors Track Heart Rate and Facial Activity","Microwatt-Powered Glasses Sensors Detect Pulse and Jaw Movement","Self-Powered Eyewear Sensors Map Pulse and Facial Activity Without Gels","GlassTENG: Self-Powered Eyewear Sensors Track Pulse and Jaw Kinematics"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1261,"prompt_tokens":635,"completion_tokens":626,"prompt_tokens_details":null},"tokens_in":635,"tokens_out":626,"duration_ms":26701,"temperature":1.0,"reasoning_tokens":509,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T03:19:02.051143+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If participants wearing the glasses during normal ambulatory activity (walking, talking, head turning) produce TENG signals where the pulse waveform is unrecognizable or activity classification drops below ~80%, the claim that this is a viable platform for longitudinal physiological monitoring would be undermined.","supporting_citations":[],"review_version":1}