acceptodds
Under review as a conference paper at ICLR 2027

BEATFORMER: A Delay-Aware Cardiovascular Foundation Model for Beat-Resolved ECG–PPG Representation Learning

Abstract

Electrocardiography and photoplethysmography record the same heartbeat at two ends of the cardiovascular system, separated by a beat-to-beat delay that reflects ejection dynamics, blood pressure and arterial stiffness. Existing foundation models tokenize the signals into fixed-length patches, obscuring this delay variation and conflating within-beat morphology with across-beat rhythm. We introduce BEATFORMER, a multimodal foundation model that organizes synchronized ECG and PPG as sequences of paired heartbeats anchored on shared R peaks. A delay-aware fusion module estimates a posterior over the beat-specific ECG–PPG delay and uses its confidence to gate bidirectional cross-modal attention. An orthogonal beat–phase Transformer then models within-beat morphology and across-beat dynamics using physical-time and rhythm-aware positional biases. Pretrained on 84.76 million valid cardiac cycles, BEATFORMER outperforms existing foundation models on 9 of 15 physiological-signal classification and regression tasks, and predicts pulse arrival time (PAT) with a mean absolute error of 11.48 , 24.6% below the strongest baseline. Across 11 masked-signal reconstruction conditions, it reduces mean RMSE by 11.3% relative to the respective strongest baselines and achieves the highest median signal-to-noise ratio improvement in 13 of 15 denoising conditions. These results support beat-resolved ECG-PPG modeling for diverse physiological prediction and waveform recovery tasks. Code is available at: https://anonymous.4open.science/r/anonymous-model-release-6541/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cardiovascular foundation model, ECG–PPG representation learning, beat-resolved modeling, delay-aware cross-modal fusion

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_dUR8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes BEATFORMER, a 73.8M-parameter foundation model for synchronized ECG and PPG that represents a window as a sequence of paired beats anchored on ECG R peaks, each beat resampled to 128 samples and encoded as eight phase tokens per modality. A delay-aware fusion module estimates a posterior over the ECG-to-PPG lag for each beat and uses its confidence to bias bidirectional cross-modal attention, followed by factorized within-beat and across-beat attention with physical-time biases. Pretraining on about 84.8M beats, almost all from ICU and operating-room recordings, combines masked reconstruction with PAT, PPG morphology and HRV regression. Evaluation covers 15 frozen-backbone classification and regression tasks on nine datasets against six foundation models (five of them PPG-only), PAT on four datasets, and fine-tuned reconstruction and denoising on VTaC. BEATFORMER is best on 9 of 15 tasks, with a PAT MAE of 11.48 ms on Li ECG–PPG.

Critical strengths
  1. The main strength, to me, is a design that follows the physiology: beats anchored on R peaks with phase tokens, an explicit ECG-PPG lag, and positional biases in physical time with RR information, instead of arbitrary fixed patches.
  2. The evaluation is broad and honest: 15 tasks on nine datasets plus reconstruction and denoising, subject-disjoint splits, PAT on four datasets including the ones where it loses, an explicit statement that recovery assumes clean beat segmentation, and code released.
Critical weaknesses
  1. Novelty for ICLR is incremental, and the scope stays cardiac... (i) In my opinion the methodological novelty is incremental for ICLR. Each block has a known counterpart: beat-level tokenization anchored on R peaks has already been done for ECG, the within-beat and across-beat attention is axial (factorized) attention, the delay posterior is a soft cross-correlation over eight lags turned into an attention bias, and the pretraining is multi-task regression of handcrafted physiological targets (PAT, PPG morphology, HRV). The combination is well thought out, but to me it is careful engineering for one pair of signals. Also, the claim of being the first to put PAT supervision into pretraining needs softening, since App A itself describes a model with a PAT-anchored delay aligner. (ii) The scope is still one physiological system: synchronized ECG and PPG, cardiac only, and the model cannot run without both signals (no unimodal variant, App. h), while most wearables record PPG continuously and ECG only on demand, if at all. If the general idea is event-anchored tokens with an uncertain lag between modalities, it would need to be shown on other paired signals with physiological delays, e.g. ECG-PCG (already in SensSmartTech, which the paper uses), ECG-arterial pressure, or beyond cardiology. As it stands, it may fit better more applied venues.
  2. The delay-aware part is not isolated (i) The module in the title, delay-aware fusion, is never ablated: App. G.6 removes six components, not this one, and evaluates only HR and BP, not PAT or recovery, with one run per variant. I also wonder what "beat-specific delay" can resolve when the lag posterior moves in steps of 1/8 beat, around 100 ms at 75 bpm, while the beat-to-beat changes in PAT that motivate the paper are much smaller. This needs to be done thoroughly imo. (ii) Five of the six baselines see only PPG, while BEATFORMER sees ECG and PPG. For HR, PAT and VT false-alarm detection the ECG is decisive, and the RR intervals are given to the model as input features (the beat-context descriptor), so the HR results are close to by construction. Against the only paired baseline, CSFM, it loses on 3 of 15 tasks, including VT false alarms. A late-fusion baseline, e.g. an ECGFounder or ECG-FM embedding concatenated with a PPG foundation model embedding, is needed to show that the delay-aware fusion adds something beyond having both signals. In the end, this is your selling point. (iii) The PAT evidence is weaker than it reads. The reference is computed by an algorithm from the same two signals (R peak to PPG foot), close to the pretraining target, so the probe mainly shows that the embedding kept what it was trained to predict, and the PPG-only models cannot see the R peak at all. On UQ and ErgoBP the PAT errors are around 90 ms with window-level SDs above 120 ms, and a PPG-only model is better on ErgoBP. Could device-dependent ECG–PPG offsets in monitor data explain this? Am I missing something? (iv) The statistics do not support most of the margins. There is a single pretraining run, and the authors' own sensitivity analysis moves SBP MAE between 10.42 and 11.02 mmHg, more than the gap to the best PPG-only baseline (10.75 vs 10.89). The fold stds in Table 3 are also large compared with most of the gaps. (v) The recovery results lean on the setup. Beats are segmented on the clean signals before masking or corruption (stated in App. E.2), the noise is defined in per-beat coordinates, and BEATFORMER is fine-tuned with 11 to 22 times more optimizer updates than CSFM, VCT-M2AE and MOMENT at the same learning rates; CSFM ends with negative SNR improvements, which looks more like a broken adaptation than a fair comparison. In denoising, the from-scratch Small Transformer on the same beat packing is within about 2 dB and wins two conditions, so imo most of the gain over the other baselines comes from the beat packing with clean segmentation, not from pretraining.
Questions for authors

My main concerns are the two weaknesses above, so I copy them here. I'd be open to raising the score if the delay-aware fusion is ablated, the comparisons are made like for like and the scope goes beyond ECG and PPG, and the below points are defended

  1. Novelty for ICLR is incremental, and the scope stays cardiac... (i) In my opinion the methodological novelty is incremental for ICLR. Each block has a known counterpart: beat-level tokenization anchored on R peaks has already been done for ECG, the within-beat and across-beat attention is axial (factorized) attention, the delay posterior is a soft cross-correlation over eight lags turned into an attention bias, and the pretraining is multi-task regression of handcrafted physiological targets (PAT, PPG morphology, HRV). The combination is well thought out, but to me it is careful engineering for one pair of signals. Also, the claim of being the first to put PAT supervision into pretraining needs softening, since App A itself describes a model with a PAT-anchored delay aligner. (ii) The scope is still one physiological system: synchronized ECG and PPG, cardiac only, and the model cannot run without both signals (no unimodal variant, App. h), while most wearables record PPG continuously and ECG only on demand, if at all. If the general idea is event-anchored tokens with an uncertain lag between modalities, it would need to be shown on other paired signals with physiological delays, e.g. ECG-PCG (already in SensSmartTech, which the paper uses), ECG-arterial pressure, or beyond cardiology. As it stands, it may fit better more applied venues.
  2. The delay-aware part is not isolated (i) The module in the title, delay-aware fusion, is never ablated: App. G.6 removes six components, not this one, and evaluates only HR and BP, not PAT or recovery, with one run per variant. I also wonder what "beat-specific delay" can resolve when the lag posterior moves in steps of 1/8 beat, around 100 ms at 75 bpm, while the beat-to-beat changes in PAT that motivate the paper are much smaller. This needs to be done thoroughly imo. (ii) Five of the six baselines see only PPG, while BEATFORMER sees ECG and PPG. For HR, PAT and VT false-alarm detection the ECG is decisive, and the RR intervals are given to the model as input features (the beat-context descriptor), so the HR results are close to by construction. Against the only paired baseline, CSFM, it loses on 3 of 15 tasks, including VT false alarms. A late-fusion baseline, e.g. an ECGFounder or ECG-FM embedding concatenated with a PPG foundation model embedding, is needed to show that the delay-aware fusion adds something beyond having both signals. In the end, this is your selling point. (iii) The PAT evidence is weaker than it reads. The reference is computed by an algorithm from the same two signals (R peak to PPG foot), close to the pretraining target, so the probe mainly shows that the embedding kept what it was trained to predict, and the PPG-only models cannot see the R peak at all. On UQ and ErgoBP the PAT errors are around 90 ms with window-level SDs above 120 ms, and a PPG-only model is better on ErgoBP. Could device-dependent ECG–PPG offsets in monitor data explain this? Am I missing something? (iv) The statistics do not support most of the margins. There is a single pretraining run, and the authors' own sensitivity analysis moves SBP MAE between 10.42 and 11.02 mmHg, more than the gap to the best PPG-only baseline (10.75 vs 10.89). The fold stds in Table 3 are also large compared with most of the gaps. (v) The recovery results lean on the setup. Beats are segmented on the clean signals before masking or corruption (stated in App. E.2), the noise is defined in per-beat coordinates, and BEATFORMER is fine-tuned with 11 to 22 times more optimizer updates than CSFM, VCT-M2AE and MOMENT at the same learning rates; CSFM ends with negative SNR improvements, which looks more like a broken adaptation than a fair comparison. In denoising, the from-scratch Small Transformer on the same beat packing is within about 2 dB and wins two conditions, so imo most of the gain over the other baselines comes from the beat packing with clean segmentation, not from pretraining.
12 comments
AuthorsOpenReview ↗

Response to Reviewer dUR8 (1/6): Methodological contribution and ECG-PCG transfer

We thank the reviewer dUR8 for recognizing the physiological organization of the model. We address the concerns and questions below, including methodological novelty, ECG--PCG adaptation, delay-aware fusion, dual-input comparisons, PAT reference quality, and the recovery evaluation.

Methodological contribution and PAT supervision

We thank the reviewer for recognizing that the combination is well thought out. Our research is indeed motivated by practical problems. In the early stages, we also spent a lot of time collecting data ourselves in real-world settings to understand the challenges these signals present in practice and inform our model design.

We acknowledge that several components are closely related to existing methods. Our emphasis is on carefully designing and empirically evaluating their integration, guided by established physiological understanding, to address paired ECG--PPG correspondence, temporal offsets, and cardiac dynamics. Rather than claiming that every component is individually new, our primary aim is to address these problems effectively.

We also thank the reviewer for pointing out that our original priority claim was insufficiently precise. We narrow this claim as follows:

BEATFORMER is a multimodal cardiovascular foundation model to explicitly formulate paired ECG–PPG representations and incorporate beat-specific PAT supervision directly into its pretraining objective.

Our intended contribution is to model ECG and PPG as physiologically corresponding paired beats and use beat-specific PAT supervision to encourage these paired representations to preserve the timing relationship between the two signals, rather than broadly claim the first use of PAT as a pretraining target.

Extension to ECG--PCG

We agree with the reviewer that our model was designed to align physiologically corresponding events in ECG and PPG, using ECG R peaks as shared anchors for paired-beat segmentation. When only PPG is available, the absence of these ECG anchors makes beat segmentation more challenging, and our current architecture also performs less well when adapted to PPG-only inputs. However, the paired-event formulation can be explored with other physiologically related signal pairs. We thank the reviewer for suggesting ECG--PCG as such an extension. Following this suggestion, we evaluated the transfer of BEATFORMER, pretrained on ECG--PPG, to ECG--PCG under both frozen and fine-tuned settings.

\text{Model}&\begin{array}{c} \text{SensSmartTech} \\\\ \text{AUROC} \uparrow \end{array}&\begin{array}{c} \text{SensSmartTech} \\\\ \text{HR MAE (bpm)} \downarrow \end{array}&\begin{array}{c} \text{CinC 2016} \\\\ \text{AUROC} \uparrow \end{array} \\\\ \hline \text{CNN, from scratch}&0.8355\pm0.0682&9.7888\pm1.9244&0.7456\pm0.0080 \\\\ \text{Transformer, from scratch}&0.7764\pm0.0750&12.6297\pm2.7008&0.6905\pm0.0259 \\\\ \text{ECGFounder+PANNs, frozen}&0.9197\pm0.0677&4.3330\pm1.4896&\mathbf{0.8544\pm0.0007} \\\\ \text{BEATFORMER, frozen}&0.8733\pm0.0241&9.2565\pm0.7168&0.5768\pm0.0073 \\\\ \text{ECGFounder+PANNs, fine-tuned}&0.9362\pm0.0430&3.8191\pm0.6319&0.8373\pm0.0188 \\\\ \text{BEATFORMER, fine-tuned}&\mathbf{0.9586\pm0.0399}&\mathbf{3.2740\pm0.5481}&0.7745\pm0.0260 \\\\ \end{array}$$ Despite being pretrained on ECG--PPG rather than ECG--PCG, BEATFORMER adapts to the new signal pair through fine-tuning, achieving the best classification and HR results among the compared learned models on SensSmartTech. On CinC 2016, fine-tuning improves its AUROC from 0.5768 to 0.7745, exceeding the CNN and Transformer trained from scratch, although it remains below ECGFounder+PANNs. These results support the applicability of our paired-signal modeling framework beyond ECG--PPG, with the CinC 2016 experiment providing exploratory record-level evidence.
AuthorsOpenReview ↗

Response to Reviewer dUR8 (2/6): Delay-aware fusion and lag resolution

Ablation of delay-aware fusion

We thank the reviewer for pointing out that the original ablations did not directly examine the fusion module. We have added comparisons covering HR, PAT, blood pressure, and classification tasks. Specifically, we compare the full model with two variants: one removes the delay bias while retaining cross-modal attention, and the other retains the delay bias but fixes the posterior confidence to γk=1\gamma_k=1. All three variants were trained for five epochs using the same training data and optimization settings.

\text{Dataset / task / metric}&\begin{array}{c} \text{Without cross-modal} \\\\ \text{attention} \end{array}&\begin{array}{c} \text{Without delay bias} \\\\ \text{(} B_k=0 \text{)} \end{array}&\begin{array}{c} \text{Fixed confidence} \\\\ \text{(} \gamma_k=1 \text{)} \end{array}&\text{Full model} \\\\ \hline \text{Simultaneous / HR MAE (bpm)} \downarrow&15.305\pm7.148&14.634\pm5.901&13.553\pm4.492&\mathbf{11.354\pm4.511} \\\\ \text{Li ECG--PPG / PAT MAE (ms)} \downarrow&13.568\pm2.381&13.377\pm1.766&13.621\pm1.728&\mathbf{12.227\pm2.602} \\\\ \text{PPG-DaLiA / Activity accuracy (\\%)} \uparrow&43.84\pm10.92&43.72\pm9.80&44.54\pm8.12&\mathbf{47.57\pm8.40} \\\\ \text{Simultaneous / Load accuracy (\\%)} \uparrow&54.23\pm10.64&54.44\pm11.73&56.08\pm10.32&\mathbf{57.64\pm11.61} \\\\ \text{Li ECG--PPG / SBP MAE (mmHg)} \downarrow&10.730\pm1.937&10.597\pm1.805&10.759\pm1.242&\mathbf{10.411\pm1.386} \\\\ \text{Li ECG--PPG / DBP MAE (mmHg)} \downarrow&7.657\pm0.918&7.504\pm0.597&7.397\pm0.673&\mathbf{7.367\pm0.576} \\\\ \end{array}$$ Removing the delay bias or fixing the confidence to $\gamma_k=1$ degrades mean performance across all six tasks. These comparisons support the downstream utility of the delay bias and its adaptive modulation by posterior confidence. We alos thank the reviewer for raising the concern about temporal resolution. We conducted an ablation comparing 4, 8, and 16 lag candidates. $$\begin{array}{lrrr} \text{Dataset / task / metric}&\text{4 candidates}&\text{8 candidates}&\text{16 candidates} \\\\ \hline \text{PTT PPG / HR MAE (bpm)} \downarrow&5.608&\mathbf{5.556}&5.570 \\\\ \text{Li ECG--PPG / PAT MAE (ms)} \downarrow&13.041&\mathbf{12.002}&12.233 \\\\ \text{Li ECG--PPG / SBP MAE (mmHg)} \downarrow&11.209&\mathbf{10.687}&10.960 \\\\ \text{Li ECG--PPG / DBP MAE (mmHg)} \downarrow&7.613&\mathbf{7.305}&7.460 \\\\ \end{array}$$ Increasing the number of lag candidates from 4 to 8 improves mean performance on all four tasks. Increasing it further to 16 yields similar performance without additional improvement. We therefore adopt 8 lag candidates for their lower computational cost relative to 16 lag candidates. However, we agree that this does not guarantee the accuracy of the module's internal PAT-related delay estimates. The preceding fusion ablations support the downstream utility of this module. These findings are consistent with our motivation to introduce PAT-related physiological timing information into paired ECG--PPG representation learning through delay guidance. This may suggest that even coarse-grained timing information can benefit downstream tasks.
AuthorsOpenReview ↗

Response to Reviewer dUR8 (3/6): ECG-PPG late-fusion comparisons

Comparison with ECG--PPG late fusion We compared BEATFORMER with ECGFounder and AnyPPG as unimodal baselines, and with ECGFounder+AnyPPG and CLEF-Medium+AnyPPG using concatenation-based late fusion across all 15 tasks.

\text{Dataset / task / metric}&\text{BEATFORMER}&\text{ECGFounder}&\text{AnyPPG}&\text{Founder+AnyPPG}&\text{CLEF+AnyPPG} \\\\ \hline \text{PTT PPG / HR MAE (bpm)} \downarrow&\mathbf{4.025\pm3.177}&4.335\pm3.411&5.623\pm4.652&4.337\pm3.517&5.030\pm5.247 \\\\ \text{Simultaneous / HR MAE (bpm)} \downarrow&\mathbf{9.861\pm4.114}&16.122\pm8.679&15.711\pm5.848&13.461\pm8.920&10.197\pm5.113 \\\\ \text{Li ECG--PPG / PAT MAE (ms)} \downarrow&\mathbf{11.475\pm2.185}&24.161\pm3.888&16.418\pm2.503&16.709\pm1.944&16.957\pm2.240 \\\\ \text{BUT PPG / Quality AUROC} \uparrow&\mathbf{0.839\pm0.033}&0.622\pm0.070&0.801\pm0.040&0.751\pm0.047&0.751\pm0.030 \\\\ \text{WESAD / Stress AUROC} \uparrow&0.919\pm0.027&0.887\pm0.066&\mathbf{0.922\pm0.026}&0.921\pm0.035&0.899\pm0.023 \\\\ \text{PTT PPG / Activity accuracy} \uparrow&\mathbf{0.680\pm0.145}&0.591\pm0.128&0.550\pm0.129&0.660\pm0.140&0.600\pm0.117 \\\\ \text{PPG-DaLiA / Activity accuracy} \uparrow&\mathbf{0.517\pm0.108}&0.463\pm0.123&0.457\pm0.067&0.500\pm0.133&0.467\pm0.123 \\\\ \text{ScientISST MOVE / Activity accuracy} \uparrow&0.656\pm0.245&0.625\pm0.282&0.674\pm0.207&0.646\pm0.251&\mathbf{0.686\pm0.247} \\\\ \text{Simultaneous / Load accuracy} \uparrow&\mathbf{0.569\pm0.098}&0.500\pm0.094&0.452\pm0.082&\mathbf{0.569\pm0.087}&0.553\pm0.073 \\\\ \text{SensSmartTech / Before--after AUROC} \uparrow&0.871\pm0.049&\mathbf{0.913\pm0.050}&0.787\pm0.085&0.890\pm0.065&0.863\pm0.061 \\\\ \text{VTaC / False-alarm detection AUROC} \uparrow&0.8435\pm0.0040&0.9262\pm0.0040&0.780\pm0.043&\mathbf{0.9295\pm0.0026}&0.8358\pm0.0084 \\\\ \text{PPG-DaLiA / Age MAE (years)} \downarrow&\mathbf{6.449\pm3.096}&8.376\pm2.970&6.810\pm2.177&9.289\pm1.570&9.323\pm5.027 \\\\ \text{PPG-DaLiA / BMI MAE (kg/m} ^2 \text{)} \downarrow&1.459\pm0.523&1.497\pm0.513&\mathbf{1.309\pm0.522}&1.596\pm0.418&1.625\pm0.407 \\\\ \text{Li ECG--PPG / SBP MAE (mmHg)} \downarrow&\mathbf{10.752\pm1.474}&10.822\pm1.606&10.975\pm1.528&11.790\pm1.711&11.671\pm1.640 \\\\ \text{Li ECG--PPG / DBP MAE (mmHg)} \downarrow&\mathbf{7.114\pm0.505}&7.838\pm0.854&7.206\pm0.896&8.001\pm1.075&7.518\pm0.828 \\\\ \end{array}$$ Consistent with the original main-table results, BEATFORMER maintains its advantages on HR, PAT, and BP prediction even when compared with baselines that also use both ECG and PPG through late fusion. On VTaC, ECGFounder+AnyPPG performs better, while ECGFounder alone already achieves an AUROC of 0.9262. ECGFounder's large-scale supervised pretraining on ECG diagnostic labels, including rhythm-related information, may contribute to its strong performance on this task. Overall, these results support the effectiveness of BEATFORMER's pretrained representations relative to simple ECG--PPG late fusion.
Reviewer_9ofwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BEATFORMER models ECG and PPG together, with an explicit estimate of the delay between the two signals. It anchors paired beats to ECG R peaks and uses a beat-specific delay posterior and its confidence to guide bidirectional cross-modal attention. A beat–phase Transformer then models waveform shape within beats and changes across beats using physical-time and rhythm information. Pretraining combines masked reconstruction with PAT, PPG morphology, and HRV objectives. The frozen backbone achieves the best mean result on nine of 15 prediction tasks across nine datasets. For reconstruction and denoising, the authors jointly fine-tune the backbone and a recovery head.

Critical strengths
  1. The beat-based hierarchy fits ECG–PPG structure, and physical-time encoding retains rhythm information after resampling. Removing within-beat attention, across-beat attention, or cumulative-time encoding increases errors on all four ablation tasks, directly supporting these architectural choices.
  2. Across 15 prediction tasks from nine datasets, BEATFORMER achieves the best mean result on nine tasks. Its subject-macro PAT MAE on Li ECG–PPG is about 11.48 ms, compared with 15.22 ms for paired-input CSFM. Subject-disjoint evaluation with a frozen backbone supports the transfer value of the representations.
Critical weaknesses
  1. PAT supervision trains a prediction head on beat representations, and downstream PAT regression uses frozen window representations. Neither directly validates the fusion module's internal delay estimate or confidence. Removing PAT supervision also differs from removing the delay bias. The results support PAT-related information in the representations, but do not isolate the delay module's contribution.
  2. The reported 11.3% reduction in mean reconstruction RMSE compares models in their native normalized coordinates. A comparison on a common reference amplitude scale is missing. Because RMSE depends on amplitude scaling, scale differences may affect the magnitude and ranking of the reconstruction advantage. This concern does not apply to the prediction results in Table 3.
Questions for authors
  1. How are PPG events handled when they occur after the beat's right boundary, defined as the midpoint between consecutive R peaks? How often does this occur, and can subsequent across-beat modeling compensate?
  2. Beat-context descriptors include rhythm and physical-time information, while heart-rate and HRV targets are directly related to R–R intervals. How well would a lightweight predictor using only these descriptors perform, for example on heart-rate estimation?
  3. What is the contribution of pretraining to waveform recovery when compared with the same BEATFORMER architecture trained from random initialization under the same task-adaptation budget? The Small Transformer baseline changes both architecture and capacity.
5 comments
AuthorsOpenReview ↗

Response to Reviewer 9ofw (1/3): Delay-aware fusion and common-scale reconstruction

We thank the reviewer 9ofw for recognizing the beat-based hierarchy. We address all questions below, including analyses of boundary crossings and the predictive information already present in the timing inputs.

Contribution of delay-aware fusion

We thank the reviewer for distinguishing PAT-related information in the learned representations from the accuracy of the internal delay estimate. But our design is motivated by practical representation learning: we use PAT as a physiological supervisory signal to encourage the paired ECG--PPG representations to preserve their temporal relationship and support downstream prediction. We agree that removing PAT supervision is different from ablating the fusion module.

To examine the contribution of the delay bias and posterior-confidence modulation in delay-aware fusion, we compared the full model with two variants: one removes the delay bias while retaining cross-modal attention, and the other retains the delay bias but fixes the confidence to γk=1\gamma_k=1. All three variants were trained for five epochs using the same training data and optimization settings.

\text{Dataset / task / metric} &\begin{array}{c} \text{Without cross-modal} \\\\ \text{attention} \end{array} &\begin{array}{c} \text{Without delay bias} \\\\ \text{(} B_k=0 \text{)} \end{array} &\begin{array}{c} \text{Fixed confidence} \\\\ \text{(} \gamma_k=1 \text{)} \end{array} &\text{Full model} \\\\ \hline \text{Simultaneous / HR MAE (bpm)} \downarrow &15.305\pm7.148 &14.634\pm5.901 &13.553\pm4.492 &\mathbf{11.354\pm4.511} \\\\ \text{Li ECG--PPG / PAT MAE (ms)} \downarrow &13.568\pm2.381 &13.377\pm1.766 &13.621\pm1.728 &\mathbf{12.227\pm2.602} \\\\ \text{PPG-DaLiA / Activity accuracy (\\%)} \uparrow &43.84\pm10.92 &43.72\pm9.80 &44.54\pm8.12 &\mathbf{47.57\pm8.40} \\\\ \text{Simultaneous / Load accuracy (\\%)} \uparrow &54.23\pm10.64 &54.44\pm11.73 &56.08\pm10.32 &\mathbf{57.64\pm11.61} \\\\ \text{Li ECG--PPG / SBP MAE (mmHg)} \downarrow &10.730\pm1.937 &10.597\pm1.805 &10.759\pm1.242 &\mathbf{10.411\pm1.386} \\\\ \text{Li ECG--PPG / DBP MAE (mmHg)} \downarrow &7.657\pm0.918 &7.504\pm0.597 &7.397\pm0.673 &\mathbf{7.367\pm0.576} \\\\ \end{array}$$ The full model achieves the best mean performance on all six reported tasks. Removing the delay bias increases PAT MAE from 12.227 to 13.377 ms and Simultaneous HR MAE from 11.354 to 14.634 bpm. Retaining the delay bias but fixing the confidence to $\gamma_k=1$ also increases these errors to 13.621 ms and 13.553 bpm, respectively. The classification and BP results show the same direction of improvement. These comparisons support the downstream utility of the delay bias and its adaptive modulation by posterior confidence. **Reconstruction on a common amplitude scale** We thank the reviewer for pointing out that RMSE values expressed in different normalized coordinates require a common amplitude scale for comparison. Using the original epoch-10 checkpoints, we re-evaluated all five models with matched target regions, scoring points, and reference waveforms. The main evaluation uses a common reference waveform cleaned with NeuroKit and the reference standard deviation of each channel over the complete 10-second window as the common amplitude scale. Each model's predictions are first transformed back from its own normalization and then mapped to this common coordinate system. We apply no scaling or offset fitted between predictions and references. The evaluation covers all 11 reconstruction conditions in the paper. Each of the seven phase-masking conditions contains 2,654 windows. The nominal 0.5-, 1-, 2-, and 4-second gaps contain 2,654, 2,653, 2,650, and 2,627 windows, respectively. $$\begin{array}{lrrr} \text{Model}&\text{ECG RMSE} \downarrow&\text{PPG RMSE} \downarrow&\text{Joint RMSE} \downarrow \\\\ \hline \text{BEATFORMER}&\mathbf{0.6055}&\mathbf{0.4523}&\mathbf{0.5377} \\\\ \text{Small Transformer}&0.7784&0.5407&0.6737 \\\\ \text{CSFM}&0.7491&0.5452&0.6559 \\\\ \text{VCT-M2AE}&0.7425&0.5141&0.6399 \\\\ \text{MOMENT}&0.7882&0.4895&0.6593 \\\\ \end{array}$$ For each condition, we pool squared errors over the scoring points and take the square root, then average the resulting RMSE values equally across the 11 conditions. Joint scoring assigns equal sampling-point weight to ECG and PPG. Under this common reference protocol, BEATFORMER reduces the overall joint RMSE by 15.96% relative to the strongest baseline, VCT-M2AE, and achieves the lowest joint RMSE in all 11 conditions. This common-reference evaluation supports BEATFORMER's advantage in overall joint reconstruction RMSE. The 15.96% reduction refers specifically to the main evaluation with a common reference waveform, window-level standard-deviation scaling, and matched scoring locations. The originally reported 11.3% was calculated in each model's native normalized coordinates.
AuthorsOpenReview ↗

Response to Reviewer 9ofw (2/3): Pretraining and timing-input information

Contribution of pretraining to waveform recovery

We thank the reviewer for asking whether pretraining improves waveform recovery beyond the contribution of the architecture. We compared randomly initialized and pretrained BEATFORMER models with the same backbone architecture, recovery-head design, training data, and optimization settings. Both groups underwent ten epochs of task adaptation using the original learning-rate schedule. Evaluation used the same clean beat segmentation, corrupted waveforms, target regions, and common amplitude scale.

\text{Task / metric}&\text{Random initialization}&\text{Pretrained initialization} \\\\ \hline \text{Reconstruction / joint RMSE} \downarrow&0.53192&\mathbf{0.49161} \\\\ \text{Denoising / joint RMSE} \downarrow&0.12454&\mathbf{0.11056} \\\\ \end{array}$$ Pretrained initialization reduces reconstruction and denoising RMSE by 7.58% and 11.23%, respectively, relative to random initialization. Equivalently, the randomly initialized models have errors 8.20% and 12.64% higher than the pretrained models. These comparisons support an additional benefit from pretrained weights under the matched ten-epoch adaptation budget. Both initializations nevertheless yield effective recovery. The architecture and task-training procedure can therefore learn recovery from random initialization, and the overall recovery performance should not be attributed entirely to pretraining. The comparison distinguishes the additional benefit of pretrained initialization from the recovery ability of the architecture and adaptation procedure under the evaluated protocol. **How much HR information is contained in the timing inputs?** We evaluated lightweight predictors using only the rhythm and physical-time components of the beat-context descriptors, including the preceding and following R--R intervals, beat duration, and the physical-time offset of each phase position relative to the R peak, without ECG or PPG waveforms. We compared ridge regression and an MLP against the full pretrained model on Simultaneous HR. Both the timing-only MLP and the full model's downstream prediction head were trained for 40 epochs. The ridge regularization parameter was selected using validation data. $$\begin{array}{lr} \text{Predictor}&\text{HR MAE (bpm)} \\\\ \hline \text{Timing-only ridge}&13.267\pm7.978 \\\\ \text{Timing-only MLP, 40 epochs}&17.114\pm13.904 \\\\ \text{Full pretrained model, 40-epoch head}&\mathbf{9.858\pm4.114} \\\\ \end{array}$$ The pretrained model's strong performance demonstrates the practical value of its learned representations.
AuthorsOpenReview ↗

Response to Reviewer 9ofw (3/3): Beat-boundary crossings and temporal robustness

PPG events beyond the beat boundary

When a PPG event crosses the right boundary of a beat, it typically enters the next beat segment. We examined this phenomenon on the PTT-PPG dataset, which contains recordings from 22 healthy participants during sitting, walking, and running. We identified the PPG maximum-upstroke point associated with each ECG R peak and defined a boundary crossing as this point falling at or beyond the right beat boundary, defined by the midpoint between the current and subsequent annotated R peaks. Among 35,409 successfully matched beats, 1,413 (3.99%) exhibited such crossings. We additionally evaluated BUT PPG dataset, which contains smartphone-camera PPG recordings from the finger or ear, together with concurrently recorded reference ECG. Across all 3,888 records from 50 participants, 1,920 of 11,386 successfully matched beats (16.86%) exhibited boundary crossings.

Provided that the next beat segment remains within the input window, the event remains available to the model, and across-beat attention is designed to incorporate information from neighboring beat segments. The reliability-weighted readout additionally aims to reduce the contribution of unreliable beat representations.

To assess sensitivity to temporal misalignment, we conducted R-peak and PPG perturbation experiments. In the R-peak experiment, all anchors within each window were shifted together by either −20-20 or +20+20 ms, followed by resegmentation of both signals and recomputation of the timing inputs; the waveforms themselves were unchanged. We additionally evaluated a mixed condition in which each window was independently left unchanged with 50% probability, shifted by −20-20 ms with 25% probability, or shifted by +20+20 ms with 25% probability. In the PPG experiment, the entire PPG signal was delayed by 2020 ms while ECG remained unchanged.

Within each experiment, the backbone and prediction head were fixed, and all conditions were evaluated on the same eligible samples. The R-peak experiment retains beats that can be validly extracted under all anchor-shift conditions, whereas the PPG experiment selects windows satisfying the boundary-margin and segment-length requirements of the signal-shift procedure. These criteria yield different retained windows or beats, so each experiment recomputes its unperturbed baseline using its corresponding retained inputs to ensure matched comparisons within that experiment. We report six selected regression and classification tasks below.

\text{Dataset / task / metric} &\begin{array}{c}\text{R-peak experiment}\\\\ \text{Unperturbed}\end{array} &\begin{array}{c}\text{R-peak experiment}\\\\ \text{Fixed shifts}\\\\ \text{Mean of }\pm20\text{ ms}\end{array} &\begin{array}{c}\text{R-peak experiment}\\\\ 50\\%\text{ unperturbed}\\\\ 25\\%\\ {-20},\\ 25\\%\\ {+20}\text{ ms}\end{array} &\begin{array}{c}\text{PPG experiment}\\\\ \text{Unperturbed}\end{array} &\begin{array}{c}\text{PPG experiment}\\\\ +20\text{ ms delay}\end{array} \\\\ \hline \text{PTT PPG / HR MAE (bpm)} \downarrow &4.394\pm3.118 &4.536\pm3.234 &4.457\pm3.168 &4.130\pm2.900 &4.464\pm3.029 \\\\ \text{Li ECG--PPG / PAT MAE (ms)} \downarrow &10.624\pm2.036 &12.839\pm2.509 &11.781\pm2.248 &10.036\pm2.042 &11.032\pm3.195 \\\\ \text{BUT PPG / Quality AUROC} \uparrow &0.8385\pm0.0303 &0.8368\pm0.0300 &0.8334\pm0.0280 &0.8308\pm0.0323 &0.8305\pm0.0310 \\\\ \text{Simultaneous / Load accuracy (\\%)} \uparrow &56.94\pm10.16 &56.79\pm9.35 &57.18\pm9.66 &54.61\pm11.47 &54.25\pm9.39 \\\\ \text{Li ECG--PPG / SBP MAE (mmHg)} \downarrow &10.673\pm1.521 &10.833\pm1.369 &10.841\pm1.483 &10.679\pm1.471 &11.209\pm1.486 \\\\ \text{Li ECG--PPG / DBP MAE (mmHg)} \downarrow &7.444\pm0.572 &7.610\pm0.548 &7.415\pm0.528 &7.799\pm0.665 &7.936\pm0.571 \\\\ \end{array}$$ These results suggest that the model can tolerate some degree of temporal misalignment, consistent with its design to retain and integrate information across neighboring beat segments.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Reviewer xX74 (1/3): R-peak robustness and CSFM denoising

We thank the reviewer xX74 for recognizing the physiology-informed design and the breadth of our evaluation. We address the concerns and questions below, including R-peak localization errors, baseline comparisons, model scale, and the role of morphology supervision in HR prediction.

Sensitivity to R-peak localization errors

We evaluated sensitivity to R-peak localization errors under fixed and mixed perturbations. For fixed shifts, all R-peak anchors within each window were shifted together by either −20-20 or +20+20 ms. For mixed perturbations, each window was independently left unchanged with 50% probability, shifted by −20-20 ms with 25% probability, or shifted by +20+20 ms with 25% probability.

The ECG and PPG waveforms themselves were unchanged. We resegmented both signals using the shifted anchors and recomputed the timing inputs. The backbone and prediction heads saved before perturbation were kept fixed, and all conditions were evaluated using the same common set of valid beats and samples.

\text{Dataset / task / metric}&\text{Unperturbed}&\begin{array}{c}\text{Fixed shifts}\\\\ \text{Mean of}\pm20\text{ms}\end{array}&\begin{array}{c}\text{50\\% unperturbed}\\\\ \text{25\\%} -20\text{, 25\\%} +20\text{ms}\end{array}\\\\ \hline \text{PTT PPG / HR MAE (bpm)}\downarrow&4.394\pm3.118&4.536\pm3.234&4.457\pm3.168\\\\ \text{Li ECG--PPG / PAT MAE (ms)}\downarrow&10.624\pm2.036&12.839\pm2.509&11.781\pm2.248\\\\ \text{BUT PPG / Quality AUROC}\uparrow&0.8385\pm0.0303&0.8368\pm0.0300&0.8334\pm0.0280\\\\ \text{Simultaneous / Load accuracy (\\%)}\uparrow&56.94\pm10.16&56.79\pm9.35&57.18\pm9.66\\\\ \text{VTaC / False-alarm detection AUROC}\uparrow&0.8377\pm0.0030&0.8260\pm0.0058&0.8288\pm0.0046\\\\ \text{Li ECG--PPG / SBP MAE (mmHg)}\downarrow&10.673\pm1.521&10.833\pm1.369&10.841\pm1.483\\\\ \text{Li ECG--PPG / DBP MAE (mmHg)}\downarrow&7.444\pm0.572&7.610\pm0.548&7.415\pm0.528\\\\ \end{array}$$ Despite changes in beat segmentation caused by $\pm20$ ms R-peak shifts, reported tasks retain performance close to their unperturbed baselines, supporting the model's robustness to anchor localization errors. **Negative SNR improvement for CSFM** We thank the reviewer for highlighting the negative SNR improvement of CSFM. We checked the original evaluation and investigated its sources through clean-input controls and error decomposition. The current CSFM recovery adaptation introduces waveform distortion even when given clean signals. In the windows examined, baseline offsets and amplitude errors account for a substantial fraction of the distortion. Negative improvement occurs when these output errors exceed the error introduced by input corruption, particularly under mild noise. Negative SNR improvement under mild noise has also been reported for a dedicated ECG denoising autoencoder, with gains of approximately $-4.09$ to $-3.93$ dB at an input SNR of $12$ dB ([Seo et al., Sensors, 2026, Section 4.2](https://doi.org/10.3390/s26165183)). To further compare waveform-shape recovery, we retained the original architectures and checkpoints for all five models. We separately centered and standardized the corrupted input, output, and reference for each beat and channel, then measured the reduction in shape error. $$\begin{array}{lr} \text{Model}&\begin{array}{c}\text{Mean of condition-wise median}\\\\ \text{shape gains (dB)}\uparrow\end{array}\\\\ \hline \text{BEATFORMER}&\mathbf{6.23}\\\\ \text{Small Transformer}&5.98\\\\ \text{VCT-M2AE}&2.58\\\\ \text{MOMENT}&1.89\\\\ \text{CSFM}&1.53\\\\ \end{array}$$ After removing offset and amplitude effects, the overall shape gain is positive for all five models, with BEATFORMER achieving the highest average gain. This analysis demonstrates shape-recovery ability in CSFM despite its negative gains under the original waveform metric. It also supports BEATFORMER's average performance advantage under a common shape-scoring procedure. This supplementary measure characterizes shape recovery alongside the original evaluation of overall waveform recovery, which includes amplitude and offset errors.
AuthorsOpenReview ↗

Response to Reviewer xX74 (2/3): Morphology supervision and HR prediction

Morphology supervision and HR prediction

We thank the reviewer for raising this question. We examined why morphology supervision affects HR prediction using physiological evidence, heart-rate-stratified ablations, and an expanded evaluation with repeated measurements.

PPG morphology and heart rate are physiologically related. In an in-silico study, increasing HR independently reduced the amplitude of the PPG diastolic peak and shifted it closer to the systolic peak within the cardiac cycle (Mathieu et al., 2024, Figure 2). Thus, an increase in HR changes the structure of individual pulses as well as their spacing. PPG morphology can therefore provide HR-related information that supports prediction.

The original comparison used a limited evaluation set and a single experimental comparison. To examine how the ablation effect varied with HR, we grouped evaluation windows into low (<60< 60 bpm), intermediate (60--100 bpm), and high (>100> 100 bpm) ranges. Intermediate-HR windows formed the majority in both datasets, while high-HR windows accounted for 38.9% of PTT-PPG and 42.1% of Simultaneous. Removing morphology supervision produced the largest absolute MAE increase in the high-HR group in both datasets. These results show that the effect varies across HR ranges and that the high-HR group contributes substantially to the overall error increase.

Table 1. HR composition and morphology ablation by HR range in a single evaluation.

\text{Dataset}&\begin{array}{c}\text{HR range}\\\\ \text{(bpm)}\end{array}&\begin{array}{c}\text{Windows}\\\\ \text{(\\%)}\end{array}&\begin{array}{c}\text{Full}\\\\ \text{MAE}\end{array}&\begin{array}{c}\mathrm{no_morph}\\\\ \text{MAE}\end{array}&\begin{array}{c}\text{MAE}\\\\ \text{change}\end{array}&\begin{array}{c}\text{Relative}\\\\ \text{change}\end{array}\\\\ \hline \text{PTT-PPG}&< 60&\text{10 (1.0\\%)}&10.615&12.844&+2.229&+21.0\\%\\\\ \text{PTT-PPG}&\text{60--100}&\text{601 (60.1\\%)}&3.435&3.902&+0.467&+13.6\\%\\\\ \text{PTT-PPG}&> 100&\text{389 (38.9\\%)}&5.865&11.485&+5.620&+95.8\\%\\\\ \hline \text{Simultaneous}&< 60&\text{14 (1.4\\%)}&19.971&18.387&-1.584&-7.9\\%\\\\ \text{Simultaneous}&\text{60--100}&\text{565 (56.5\\%)}&9.551&12.753&+3.202&+33.5\\%\\\\ \text{Simultaneous}&> 100&\text{421 (42.1\\%)}&11.792&18.921&+7.129&+60.5\\%\\\\ \end{array}$$ MAE and its change are reported in bpm. MAE change is calculated as $\mathrm{MAE}_{\mathrm{no_morph}}-\mathrm{MAE}_{\mathrm{Full}}$, and relative change is this difference divided by $\mathrm{MAE}_{\mathrm{Full}}$, multiplied by 100%. We expanded the evaluation from 1,000 windows per dataset to 3,207 windows for PTT-PPG and 2,411 windows for Simultaneous, using subject-grouped five-fold and leave-one-subject-out cross-validation, respectively. Full and $\mathrm{no_morph}$ used identical windows, splits, and prediction-head configurations, with each backbone frozen and only the prediction head trained. Table 2 reports the mean $\pm$ standard deviation of MAE across outer test folds. The standard deviation reflects between-fold variability, and the relative increase is calculated from the two models' fold-averaged MAEs. **Table 2. Original and expanded HR evaluation results.** $$\begin{array}{llrrrr} \text{Dataset}&\text{Evaluation setting}&\begin{array}{c}\text{Full}\\\\ \text{MAE}\end{array}&\begin{array}{c}\mathrm{no_morph}\\\\ \text{MAE}\end{array}&\begin{array}{c}\text{MAE}\\\\ \text{change}\end{array}&\begin{array}{c}\text{Relative}\\\\ \text{change}\end{array}\\\\ \hline \text{PTT-PPG}&\text{Original evaluation}&4.452&6.941&+2.489&+55.9\\%\\\\ \text{PTT-PPG}&\text{Expanded, repeated}&4.032\pm3.036&4.326\pm2.934&+0.294&+7.3\\%\\\\ \hline \text{Simultaneous}&\text{Original evaluation}&10.640&15.428&+4.788&+45.0\\%\\\\ \text{Simultaneous}&\text{Expanded, repeated}&9.848\pm4.206&10.393\pm4.327&+0.545&+5.5\\%\\\\ \end{array}$$ In the expanded evaluation, removing morphology supervision still increases HR prediction error, but the relative increase is smaller than in the original evaluation: 7.3% versus 55.9% on PTT-PPG and 5.5% versus 45.0% on Simultaneous. Together, the physiological evidence, HR-stratified results, and repeated evaluation support a consistent auxiliary role for morphology supervision in HR prediction, with a more moderate overall effect in the expanded evaluation.
AuthorsOpenReview ↗

Response to Reviewer xX74 (3/3): Timing interventions and compact-model performance

Use of timing across beats. We fixed the original Full model and its prediction heads. For the timing intervention, we kept the waveforms unchanged and scaled beat timestamps to increase or decrease the corresponding HR by 8–25 bpm. For the morphology intervention, we kept ECG and timestamps unchanged and replaced three interior PPG beats with donor beats from the same subject, with a reference HR difference of 8–25 bpm and a PPG quality-score difference of at most 0.15. Donor beats underwent pulse-peak alignment and per-beat standardization. The table reports the prediction difference between the higher- and lower-HR conditions, averaged within each subject and then equally across subjects.

\text{Dataset}&\begin{array}{c}\text{Paired windows}\\\\ \text{/ subjects}\end{array}&\begin{array}{c}\text{Prediction change (bpm)}\\\\ \text{Fixed waveforms, changed RR}\end{array}&\begin{array}{c}\text{Prediction change (bpm)}\\\\ \text{Fixed RR, replaced PPG shape}\end{array}\\\\ \hline \text{PTT-PPG}&\text{116 / 10}&+7.169&+0.289\\\\ \text{Simultaneous}&\text{96 / 6}&+5.604&+0.482\\\\ \end{array}$$ On PTT-PPG, changing RR-related timing produces a positive shift in predicted HR even with beat waveforms held fixed. All 10 subjects show a positive mean response. This demonstrates that the model uses timing information across beats under the tested intervention, allowing benefits from morphology supervision to coexist with the use of rhythm information. Taken together, the target definitions and experiments suggest that morphology supervision supplies beat-level waveform and timing targets to the shared representation. The magnitude of each ablation reflects these training effects and does not directly rank the physiological importance of morphology and rhythm. The RR intervention in the original model provides direct evidence of timing use on PTT-PPG. **Performance at approximately 5M parameters** We appreciate the concern about model capacity. CSFM-Base is reported to have more than 100M parameters, compared with 73.8M for BEATFORMER. To further examine performance at a smaller scale, we evaluated a 4.95M-parameter BEATFORMER variant against the original 73.8M model and similarly sized baselines, AnyPPG and PaPaGei-S. The first two result columns compare with the original model; the last three compare compact models. $$\begin{array}{l|rr|rrr} \text{Dataset / task / metric}&\begin{array}{c}\text{BEATFORMER}\\\\ \text{73.8M}\end{array}&\begin{array}{c}\text{Relative}\\\\ \text{degradation (\\%)}\end{array}&\begin{array}{c}\text{BEATFORMER}\\\\ \text{4.95M}\end{array}&\begin{array}{c}\text{AnyPPG}\\\\ \text{5.8M}\end{array}&\begin{array}{c}\text{PaPaGei-S}\\\\ \sim\text{5M}\end{array}\\\\ \hline \text{PTT PPG / HR MAE (bpm)}\downarrow&4.025\pm3.177&38.41&\mathbf{5.571\pm3.345}&5.623\pm4.652&9.702\pm3.602\\\\ \text{Simultaneous / HR MAE (bpm)}\downarrow&9.861\pm4.114&15.14&\mathbf{11.354\pm4.511}&15.711\pm5.848&20.735\pm5.226\\\\ \text{Li ECG--PPG / PAT MAE (ms)}\downarrow&11.475\pm2.185&6.55&\mathbf{12.227\pm2.602}&16.418\pm2.503&20.334\pm2.388\\\\ \text{BUT PPG / Quality AUROC}\uparrow&0.839\pm0.033&1.79&\mathbf{0.824\pm0.031}&0.801\pm0.040&0.642\pm0.053\\\\ \text{WESAD / Stress AUROC}\uparrow&0.919\pm0.027&2.29&0.898\pm0.028&\mathbf{0.922\pm0.026}&0.768\pm0.048\\\\ \text{PTT PPG / Activity accuracy}\uparrow&0.680\pm0.145&7.21&\mathbf{0.631\pm0.192}&0.550\pm0.129&0.510\pm0.127\\\\ \text{PPG-DaLiA / Activity accuracy}\uparrow&0.517\pm0.108&7.93&\mathbf{0.476\pm0.084}&0.457\pm0.067&0.356\pm0.096\\\\ \text{ScientISST MOVE / Activity accuracy}\uparrow&0.656\pm0.245&-0.15&0.657\pm0.258&\mathbf{0.674\pm0.207}&0.668\pm0.206\\\\ \text{Simultaneous / Load accuracy}\uparrow&0.569\pm0.098&-1.23&\mathbf{0.576\pm0.116}&0.452\pm0.082&0.454\pm0.060\\\\ \text{SensSmartTech / Before--after AUROC}\uparrow&0.871\pm0.049&3.44&\mathbf{0.841\pm0.064}&0.787\pm0.085&0.620\pm0.055\\\\ \text{VTaC / False-alarm detection AUROC}\uparrow&0.843\pm0.035&4.27&\mathbf{0.807\pm0.008}&0.780\pm0.043&0.665\pm0.037\\\\ \text{PPG-DaLiA / Age MAE (years)}\downarrow&6.449\pm3.096&5.07&\mathbf{6.776\pm2.211}&6.810\pm2.177&8.936\pm3.686\\\\ \text{PPG-DaLiA / BMI MAE (kg/m} ^2\text{)}\downarrow&1.459\pm0.523&-5.48&1.379\pm0.452&\mathbf{1.309\pm0.522}&1.972\pm0.425\\\\ \text{Li ECG--PPG / SBP MAE (mmHg)}\downarrow&10.752\pm1.474&-0.81&\mathbf{10.665\pm1.700}&10.975\pm1.528&11.071\pm1.908\\\\ \text{Li ECG--PPG / DBP MAE (mmHg)}\downarrow&7.114\pm0.505&5.36&7.495\pm0.459&\mathbf{7.206\pm0.896}&7.802\pm1.021\\\\ \end{array}$$ Although reducing the model to 4.95M parameters decreases performance on most tasks, it still achieves the best mean performance on 11 of the 15 tasks among models of similar parameter scale.
AuthorsOpenReview ↗

Response to Reviewer xX74: Updated CSFM denoising adaptation

To further address your concern about the negative SNR improvements of CSFM, we conducted an additional denoising adaptation experiment and repeated the comparison using the same noisy inputs, reference waveforms, and scoring regions.

We retained the pretrained CSFM Base encoder and adapted it with a DPT-style recovery head that combines features from multiple encoder layers. The head predicts a correction that is added to the input waveform. The denoising adaptation used 411,060 fine-tuning updates, matching BEATFORMER's update count for denoising.

We evaluated all five models on the same 2,654 VTaC windows. Joint RMSE retains amplitude and offset errors. The table averages results across 15 noise conditions. For ΔSNR\Delta\mathrm{SNR}, we first take the median across windows within each condition and then average the condition medians with equal weight.

ModelJoint denoising RMSE ↓Joint ΔSNR\Delta\mathrm{SNR} (dB) ↑
BEATFORMER0.142330+3.94
Small Transformer0.155508+3.62
CSFM (residual DPT adaptation)0.168942+1.14
VCT-M2AE0.164278+1.58
MOMENT0.172938+0.81

Compared with the previous CSFM adaptation, the residual version reduced joint denoising RMSE from 0.380698 to 0.168942 (55.6%) and clean-input RMSE from 0.324893 to 0.028969 (91.1%). Median joint ΔSNR\Delta\mathrm{SNR} was positive in all 15 noise conditions, and pooled RMSE was lower than that of the unprocessed input in 12 conditions. Across the five light-noise conditions, the mean of the median joint ΔSNR\Delta\mathrm{SNR} values was +0.39 dB, and pooled RMSE improved in two conditions.

Against the re-adapted CSFM, BEATFORMER achieved a 15.8% lower joint denoising RMSE across the 15 VTaC noise conditions (0.142330 versus 0.168942) and a higher average joint ΔSNR\Delta\mathrm{SNR} (+3.94 versus +1.14 dB).

Thank you for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.