{"id":"74745245-e729-4eda-8c3f-6dfaa06c5322","arxiv_id":"2505.09529","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Face video from an event camera, converted to 2D event frames, lets a supervised neural network estimate heart rate with about 2 bpm error, matching or beating a standard RGB camera in a 64-subject study.","lead":"Researchers trained a neural network to read a person's heartbeat from face recordings made by an event camera, a sensor that only reports fast brightness changes. In their tests, the best event-camera model estimated heart rate with a root mean square error of 2.13 beats per minute, slightly better than a standard 30 FPS video camera baseline of 2.92 bpm.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline claim that 120 FPS event frames beat RGB rests on 10 test subjects, a single split, and no repeated training runs; per-subject RMSEs vary from 0.00 to 5.05, so the 2.13 vs 2.92 bpm difference is not statistically established.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the most load-bearing weakness is not the ECG label proxy. The label preprocessing in Section 3.2 is physiologically questionable, yet the evaluation metric is heart rate from an FFT peak; any label that is periodic at the true cardiac frequency supplies the correct target, so waveform shape and phase differences in the inverted-ECG label do not by themselves invalidate the heart-rate comparison. The decisive issue is statistical: Table 3 is built from 10 test subjects, one training run, and one data split, while Table 4 shows per-subject errors spanning 0.00 to 5.05 bpm. A difference of 0.79 bpm in aggregate RMSE could easily be driven by a few subjects and could disappear under re-splitting or re-training. This does not refute the paper's feasibility contribution; it means the central comparative claim outruns the evidence, which is exactly the sort of limitation the CONDITIONAL verdict should carry. Therefore I keep the reader's verdict unchanged while disagreeing in part with the stated weakest assumption.","tokens_in":11799,"tokens_out":6120,"duration_ms":67799,"concrete_test":"Re-run the comparison with bootstrap or leave-one-subject-out over the 10 test subjects, and retrain each model (RGB 30, Events 30/60/120) with at least 5 random seeds on repeated training/validation splits. Compute the paired per-subject RMSE difference between 120 FPS events and RGB, and report a bootstrap 95% confidence interval. If the CI includes 0 or the sign flips across seeds, the headline 'events beat RGB' is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.4 reports aggregate RMSE over only 10 test subjects (Table 3), with a single randomly chosen split and one training run per frame rate. The per-subject errors in Table 4 are highly heterogeneous: the 120 FPS event model ranges from RMSE 0.00 (subjects 2 and 7) to 5.05 (subject 5), while the RGB model is perfect on five subjects but exceeds 5 bpm on three. Under these conditions the headline comparison (120 FPS events at 2.13 bpm vs RGB at 2.92 bpm) can hinge on one or two subjects; no confidence interval, paired test, or seed averaging is reported. This is more central than the reader's label-proxy concern: since heart rate is read from the FFT peak of the post-processed output, the exact waveform or phase of the inverted-ECG label mostly cancels; what matters is whether the event stream actually carries a periodic signal at cardiac frequency in these ten recordings. If the statistical support is lacking, the verified claim narrows to a feasibility demonstration on a proprietary stationary-subject dataset, not to the stated conclusion that event cameras are a viable low-power alternative to RGB.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper investigates whether event cameras can capture cardiac pulse information from face recordings. The authors collect a proprietary dataset of 64 subjects with simultaneous event, RGB, and ECG data, represent event streams as 2D event frames, and train a modified TS-CAN to predict a processed ECG signal. They compare models trained on RGB frames and on event frames at 30, 60, and 120 FPS, evaluating heart rate via FFT of the post-processed prediction. Reported RMSEs are 2.92 bpm for RGB 30 FPS, and 3.32, 2.54, and 2.13 bpm for event frames at 30, 60, and 120 FPS, respectively. The abstract and conclusion interpret the 120 FPS event result as outperforming the RGB baseline and supporting event cameras as a viable low-power contactless cardiac monitoring sensor.","tokens_in":12054,"tokens_out":6418,"duration_ms":63338,"significance":"If the result holds, this is a useful first demonstration: it extends event-camera vitals monitoring from wrist and chest to the face, uses end-to-end learning rather than hand-crafted event-count features, and provides code and trained weights for reproducibility. The design choices (held-out subject evaluation, no subject exclusion, standard r-PPG error metrics) are appropriate for a feasibility study. The main limitation is statistical: the headline comparison rests on 10 test subjects, one split, and one training run per configuration, with highly variable per-subject errors.","major_comments":[{"comment":"The claim that the 120 FPS event model outperforms the RGB baseline (RMSE 2.13 vs. 2.92 bpm) is not statistically established. The evaluation uses a single random split (44/10/10 subjects), one training run per configuration, and a test set of only 10 subjects. Per-subject RMSE in Table 4 varies from 0.00 to 5.05 for the 120 FPS event model, so the aggregate difference could be dominated by one or two subjects. Please report confidence intervals, a paired significance test over subjects, and ideally multiple splits or seeds before drawing the comparative conclusion in the abstract.","section":"Section 4.4, Tables 3 and 4"},{"comment":"The label pre-processing inverts the ECG to emphasize s-wave peaks and then applies smoothing, band-pass filtering, clipping, resampling, and differentiation, but no evidence is given that the resulting signal tracks the facial blood-volume pulse observed by the event camera. Since heart rate is derived from the FFT peak of the post-processed output, waveform phase details may be less critical, but the claim that the model extracts 'physiological cardiac information' would be considerably stronger if the processed ECG were validated against an SpO2 or standard r-PPG reference, or if an ablation without the inversion step were reported.","section":"Section 3.2"},{"comment":"The manuscript concludes that event cameras are a viable low-power alternative to RGB for remote heart rate monitoring, but the current evidence supports a narrower feasibility claim: on a stationary, proprietary 10-subject test set, a 120 FPS event-frame model achieves competitive aggregate RMSE. Given the acknowledged lack of motion scenarios, the small test set, and the single-split evaluation, I recommend softening the conclusion to a feasibility demonstration until broader evaluation is available.","section":"Section 5, Conclusion"}],"minor_comments":[{"comment":"Equation (1) defines Wj with an ambiguous set-builder notation; please clarify that i and k are the first and last event indices in the window and write the set condition explicitly.","section":"Section 3.1.1"},{"comment":"The pre-processing description says frames are normalized between -8 and 8 and then multiplied by 255 and cast to 8-bit unsigned integers; negative values cannot be represented in this conversion, so please describe the actual normalization and typecasting procedure.","section":"Section 3.1.1"},{"comment":"The caption says 'A 10-second sample is inverted and smoothed' while the text says the figure shows a 1-second sample; these statements are inconsistent.","section":"Figure 3"},{"comment":"The paper states that validation loss was used to select optimal weights, but no validation metrics or selected epoch numbers are reported; please include them.","section":"Section 4.3"},{"comment":"The implementation details of the event-count baseline from [26] (filter order, cutoff frequencies, FFT windowing) are not provided, which limits reproducibility of the comparison.","section":"Section 5, Table 5"},{"comment":"The sentence about some subjects providing an additional minute of break data is vague; clarify how recording durations varied across the 64 subjects.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope for the journal and the core feasibility result is likely worth publishing if the statistical support is strengthened. The label-proxy concern is real but secondary because heart rate is computed from the FFT peak, so the main load-bearing issue is the single-split, 10-subject comparison. I also suggest verifying reference [36], which appears to be an incomplete self-citation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first face-based event-camera pulse extraction with an end-to-end CNN, and the code and weights are public. The narrow claim is plausible: event frames from a face contain enough periodic cardiac information for a model to estimate heart rate. The headline comparison, 120 FPS events at 2.13 bpm RMSE vs RGB at 2.92, is not statistically established.\n\nWhat's new and good: the paper does something no one has done before, and it does it straightforwardly. It bins events into frames, feeds them to a standard TS-CAN, and shows heart rate estimation works at all. It also shows that increasing event frame rate from 30 to 60 to 120 FPS monotonically improves results, which is a useful empirical finding. The authors adapt the ECG label preprocessing explicitly and transparently, and they implement the prior event-count method from [26] on their own data and report it fails badly; that is a useful negative result. Code and weights are on GitHub, so the method itself is reproducible up to the proprietary dataset.\n\nSoft spots: the evaluation is the load-bearing weak point. Everything rests on 10 test subjects, one random split, and one training run per condition. Per-subject RMSE in Table 4 is highly heterogeneous: the RGB model is perfect on five subjects and over 5 bpm on three; the 120 FPS events model is perfect on two and over 5 bpm on one. With that variance, the 2.13 vs 2.92 difference can flip based on one or two people. No confidence intervals, paired tests, or seed averaging are reported. The paper should not say the results prove that cardiac information is preserved or that event cameras are a viable alternative to RGB. What they support is a feasibility demonstration. The label-proxy concern is real but secondary: since heart rate is read from the FFT peak of a post-processed output, the exact waveform of the inverted ECG mostly cancels; the important thing is whether the event stream carries a periodic signal at cardiac frequency, and that is what needs more subjects. Also, stationary subjects only and proprietary data, both acknowledged.\n\nWho it's for: people working on event-based vision or r-PPG who want to know whether event cameras are worth exploring for vitals. It is a first-demonstration paper, not a benchmark.\n\nRecommendation: send it to review, but expect major revision. The authors need to add multiple splits or bootstrapping, report intervals, or scale to a public dataset, and temper the language. It deserves a serious referee because the problem is new and the code release is real.","headline":"First face-based event-camera pulse extraction with a CNN, and the code release is real; but the headline 120 FPS advantage over RGB rests on ten subjects and one split, so treat the comparison as preliminary.","tokens_in":12616,"tokens_out":2869,"would_cite":true,"duration_ms":28661,"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":"Event-camera face recordings preserve cardiac pulse well enough for contactless heart-rate monitoring.","keywords":["event cameras","neuromorphic vision","remote photoplethysmography","heart rate estimation","convolutional neural networks","temporal shift attention","ECG signal preprocessing","contactless health monitoring"],"falsifier":"Simultaneously record an event camera, a synchronized RGB camera, and both ECG and fingertip SpO2 from the same subjects; train the same model with ECG-derived labels, then compare its predicted pulse peaks against SpO2 pulse peaks beat by beat. If the event-frame predictions do not track SpO2 timing within a fraction of a cardiac cycle while the RGB model does, the claim that cardiac information is preserved in event streams would be falsified.","tokens_in":1470,"feed_emoji":"❤️","tokens_out":3988,"duration_ms":78059,"temperature":0.7,"pith_summary":"The paper sets out to show that an event camera pointed at a person's face captures enough of the cardiac pulse to estimate heart rate without contact. It trains a supervised CNN on 2D event frames built from asynchronous brightness-change events, using a processed ECG signal as the target, and reports that event-frame models match or beat an RGB-video baseline when the event frames are generated at 60 or 120 FPS. The best event model reaches 2.13 bpm RMSE, below the RGB baseline's 2.92. If the result holds, event cameras, which use little power and have high dynamic range, become a plausible low-power sensor for remote health monitoring.","feed_headline":"Faster event frames beat RGB video at face heart-rate tracking","feed_subtitle":"A CNN trained on 120 FPS event frames reaches 2.13 bpm RMSE, beating the 30 FPS RGB baseline of 2.92 bpm.","key_machinery":"The machinery is the event frame: events, each carrying timestamp, coordinates, and polarity, are binned into non-overlapping windows of fixed duration and summed by polarity into a single-channel 2D image, making the asynchronous event stream compatible with frame-based network architectures. A modified TS-CAN, a two-branch convolutional attention network with temporal shift modules, takes the same single-channel event frame into both branches and learns to output a normalized signal difference that post-processing integrates into a pulse waveform. On the label side, the ECG signal is inverted to emphasize s-wave peaks, smoothed, bandpass filtered, clipped, resampled to the frame rate, and differentiated, so the supervision target is designed to align with the blood-pulse timing the event stream is assumed to encode.","core_discovery":"The paper claims that physiological cardiac information in the facial region is effectively preserved within the event stream, and that a simple 2D event-frame representation lets a standard convolutional network recover it. On its local dataset, the event-frame model at 120 FPS reaches RMSE 2.13 bpm, MAE 1.58 bpm, and Pearson correlation 0.97, outperforming the 30 FPS RGB model (RMSE 2.92) and the 60 FPS event model (RMSE 2.54). The 30 FPS event model is slightly worse than RGB but still within the accepted error range for remote photoplethysmography. The authors also report that an event-count-only method from prior work fails on face data, with RMSE above 25 bpm, while their spatial event-frame method does not.","pith_inferences":["If the finding generalizes to natural head motion, event cameras could be built into vehicle cabins or wearables without some of the illumination constraints that affect RGB remote-PPG, because event cameras have high dynamic range.","A direct test of the label assumption would be to compare the model's predicted pulse peaks against a simultaneous fingertip SpO2 waveform, whose timing reflects blood arrival in the extremities rather than electrical heart activity.","Synthetic generation of event streams from public RGB remote-PPG datasets could be used to scale training data and to test whether the model learns the same optical blood-volume signal that RGB methods exploit.","Beat-by-beat waveform comparison, not just heart-rate error, would clarify whether event frames preserve pulse shape or only periodicity."],"forward_implications":["Contactless pulse monitoring can run on event-camera hardware, which offers microsecond-level latency and lower power than frame-based video capture.","Existing frame-based remote-PPG network architectures can be reused for event data simply by feeding them temporally binned event frames, lowering the barrier to neuromorphic health sensing.","Increasing the event-frame sampling rate from 30 to 120 FPS improves heart-rate accuracy in the reported experiments, so shorter temporal bins appear to be preferable for this task.","The failure of event-count-only methods on face data suggests that spatial information is necessary for face-based event pulse estimation.","The achieved RMSE values fall within the accepted threshold of 5 bpm for remote-PPG methods, so the accuracy is in the clinically relevant range."],"supporting_citations":[{"why":"Establishes that facial blood flow can be measured remotely with consumer cameras, the baseline phenomenon this work extends to event cameras.","marker":"[8]"},{"why":"Supplies the TS-CAN architecture whose modified two-branch variant is trained on the event frames.","marker":"[9]"},{"why":"Provides the deep convolutional attention network design and the normalized frame-difference motion representation used for the RGB baseline and derivative training target.","marker":"[12]"},{"why":"Reports the wrist-based event camera heart-rate results used as a comparison point for the face-based method.","marker":"[25]"},{"why":"Gives the event-count-only baseline method that is reimplemented and shown to be unsuitable for face-based pulse estimation.","marker":"[26]"},{"why":"Describes the event camera sensor used to capture the event streams in the experiments.","marker":"[28]"},{"why":"Supplies the post-processing pipeline, including detrending and filtering, applied to model outputs before heart-rate calculation.","marker":"[31]"},{"why":"Defines the acceptable RMSE threshold of 5 bpm used to judge whether the estimated heart rates are accurate enough.","marker":"[35]"}],"fun_headline_variants":["120 FPS event frames beat 30 FPS RGB for face heart-rate tracking","Face pulse tracking: 120 FPS event frames top RGB video","Event cameras read heart rate from face, 120 FPS beats RGB","Contactless pulse monitor: 120 FPS event frames outdo RGB video","Event frames at 120 FPS deliver accurate pulse from face, beating RGB"],"cache_read_input_tokens":14720,"weakest_assumption_plain":"The paper's central claim rests on the assumption that the processed ECG signal used as the training label tracks the timing and shape of the optical blood-volume pulse that the event camera sees in the face; if that mapping is wrong, the reported errors measure something other than cardiac information in the event stream.","fun_headline_variants_meta":{"raw":{"variants":["120 FPS event frames beat 30 FPS RGB for face heart-rate tracking","Face pulse tracking: 120 FPS event frames top RGB video","Event cameras read heart rate from face, 120 FPS beats RGB","Contactless pulse monitor: 120 FPS event frames outdo RGB video","Event frames at 120 FPS deliver accurate pulse from face, beating RGB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0018,"raw_usage":{"total_tokens":7090,"prompt_tokens":947,"completion_tokens":6143,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":6044}},"tokens_in":563,"tokens_out":6143,"duration_ms":38877,"temperature":1.0,"reasoning_tokens":6044,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:29:05.235375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simultaneously record an event camera, a synchronized RGB camera, and both ECG and fingertip SpO2 from the same subjects; train the same model with ECG-derived labels, then compare its predicted pulse peaks against SpO2 pulse peaks beat by beat. If the event-frame predictions do not track SpO2 timing within a fraction of a cardiac cycle while the RGB model does, the claim that cardiac information is preserved in event streams would be falsified.","supporting_citations":[{"cited_title":"Remote plethysmographic imaging using ambient light","cited_arxiv_id":null,"evidence_quote":"Establishes that facial blood flow can be measured remotely with consumer cameras, the baseline phenomenon this work extends to event cameras."},{"cited_title":"Multi-task temporal shift attention networks for on-device contactless vitals measurement","cited_arxiv_id":null,"evidence_quote":"Supplies the TS-CAN architecture whose modified two-branch variant is trained on the event frames."},{"cited_title":"Deepphys: Video-based physiological measurement using convolutional attention networks","cited_arxiv_id":null,"evidence_quote":"Provides the deep convolutional attention network design and the normalized frame-difference motion representation used for the RGB baseline and derivative training target."},{"cited_title":"Periodicity intensity reveals insights into time series data: three use cases","cited_arxiv_id":null,"evidence_quote":"Reports the wrist-based event camera heart-rate results used as a comparison point for the face-based method."},{"cited_title":"Non-invasive vital parameter detection using neuromorphic cameras exploring the use of neuromor- phic cameras for biomedical applications","cited_arxiv_id":null,"evidence_quote":"Gives the event-count-only baseline method that is reimplemented and shown to be unsuitable for face-based pulse estimation."},{"cited_title":"A 128×128 120 db 15µs latency asynchronous temporal contrast vision sensor","cited_arxiv_id":null,"evidence_quote":"Describes the event camera sensor used to capture the event streams in the experiments."},{"cited_title":"rPPG-Toolbox: Deep Remote PPG Toolbox","cited_arxiv_id":"2210.00716","evidence_quote":"Supplies the post-processing pipeline, including detrending and filtering, applied to model outputs before heart-rate calculation."},{"cited_title":"Non-contact, automated cardiac pulse measurements using video imaging and blind source separation","cited_arxiv_id":null,"evidence_quote":"Defines the acceptable RMSE threshold of 5 bpm used to judge whether the estimated heart rates are accurate enough."}],"review_version":1}