acceptodds
Under review as a conference paper at ICLR 2027

MSMA-PPG: Channel-Aligned Expert Routing for Selected-Wavelength PPG Prediction

Abstract

Selecting a photoplethysmography (PPG) wavelength by signal quality does not necessarily minimize physiological prediction error. We introduce MSMA-PPG, a framework that learns from synchronized multi-wavelength recordings and predicts from one selected PPG stream and acceleration. Wavelength-specific projections and residual adapters share an optical temporal backbone, full-to-subset alignment links paired training views, and a Channel-Aligned Mixture of Experts combines optical, motion, and fused representations through availability-masked top-2 routing. We evaluate single-stream representation transfer across heart rate (HR), blood pressure (BP), and atrial fibrillation (AF), alongside paired HR diagnostics of wavelength selection and expert routing. Across the seven-model PPG-only test results, with routing disabled, MSMA-PPG has the lowest reported participant-macro HR MAE on EarSet and PPG-DaLiA (5.020 and 9.246 bpm), the lowest systolic BP MAE (15.456 mmHg), and the second-highest AF AUPRC/AUROC (0.5177/0.8909). The paired diagnostics reveal a different ordering after adaptation: although frozen affine probes favor aligned initialization on EarSet and PPG-DaLiA, adapted aligned models have higher window-pooled MAE than reconstruction-only controls by 0.797, 1.485, and 0.373 bpm on WildPPG, EarSet, and PPG-DaLiA. Retrospective selection from recorded wavelengths also produces model-dependent responses: lower composite policy scores sometimes reduce HR error. Frozen expert interventions reveal strong dependence on the trained motion/shared pathways and mixture weights, despite the aligned model's higher absolute error. These findings distinguish competitive optical representation transfer from the separate effects of acquisition choices and fitted expert routing on downstream accuracy.

Primary area:
learning on time series and dynamical systems
Keywords:
PPG, MoE, Smart Sensing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2TYdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduce MSMA-PPG, a framework that learns from synchronized multi-wavelength recordings and predicts from one selected PPG stream and acceleration. It combines wavelength-specific adapters, full-to-subset alignment, and availability-masked expert routing. Experiments show competitive optical representation transfer and predictor-dependent wavelength selection, yet cross-view alignment does not improve adapted heart-rate prediction over reconstruction-only training.

Critical strengths
  1. The paper studies a practical scenario. It uses abundant multi-wavelength data for training, while only limited input signals are available during inference.
  2. The proposed MSMA-PPG learns optical representations. It obtains competitive results on heart rate, blood pressure and atrial fibrillation prediction tasks.
  3. The paper uses matched time window diagnostics. These help separate predictor-dependent wavelength selection effects from impacts of the trained expert configuration.
Critical weaknesses
  1. The methodological and empirical advances need clearer positioning. The framework integrates existing techniques, including channel-specific adaptation, cross-view alignment, and sparse expert routing. The experiments are informative. However, the paper fails to clearly state its new methodological insights, beyond the observed model-dependent wavelength preferences and expert dependencies.
  2. The source of the optical representation gains remains unclear. The paper does not include proper ablation studies to separate the effects of paired alignment and wavelength-specific processing under the same PPG-only evaluation setup. Without such ablations, we cannot verify whether these core designs bring advantages over a simpler reconstruction-pretrained encoder.
  3. The architectural value of the complete routed predictor is insufficiently established. Most external baselines test representations with routing turned off, under the PPG-only setting. The full system using selected PPG and ACC only compares against its reconstruction-only version. Controlled comparisons against fully trained simple fusion methods are required to show the benefit of CA-MoE. Without these experiments, the architectural contribution is hard to justify.
Questions for authors
  1. To evaluate whether CA-MoE is necessary, can you include comparisons with simple fusion baselines? All compared models take selected PPG and ACC as input. They should be trained independently, with consistent initialization, data splits and comparable training resources, instead of relying only on interventions applied to the pre-trained model. If such results are already available, please provide them.
  2. Can you run ablation studies under the same PPG-only evaluation setup from Table 1? Maintain consistent training data and other settings. These ablations should examine whether paired alignment and wavelength-specific processing bring independent gains to the reported representation transfer performance.
  3. Please provide participant‑level and sample counts for the train, validation, and test splits of each prediction task. It would help to explicitly distinguish between the full original dataset, the cohorts used for Table 1, and those used in the paired HR experiments. Please also clarify how data was partitioned at participant level.
6 comments

Hi authors, I'm looking forward to discussing the paper. If my concerns are addressed, I will reconsider my score.

AuthorsOpenReview ↗

Matched Fusion Comparison, Optical Ablation Scope, and Data Splits

Our design uses synchronized optical views to supervise the structure that is predictable from one supplied waveform, while keeping wavelength identity available to the prediction module. The full view is used during training; at inference, experts are eligible only when supported by the supplied PPG and optional ACC. With green PPG + ACC fixed on EarSet and the backbone frozen, switching from the affine head to CA-MoE moves the alignment contrast from −0.758 to +0.057 bpm (Table 17). This shows why frozen affine-probe gains should be checked with the intended prediction module.

1. Independently trained fusion baseline. Table 3 adds a concatenation MLP trained independently of CA-MoE. Both receive the same available features, use the same participant split and task loss, and run for 5,000 EarSet adaptation updates. Within each alignment setting, they start from the same backbone initialization. With green PPG and ACC, CA-MoE and MLP yield participant-macro HR MAEs of 4.664 and 4.909 bpm with alignment and 4.733 and 4.714 bpm without it. Aligned CA-MoE is 0.245 bpm below the aligned MLP and 0.050 bpm below the best MLP setting; without alignment, MLP has the lower MAE. All other Table 3 settings use PPG alone, where CA-MoE weights both available experts. Overall, CA-MoE shows a modest advantage when paired with alignment.

2. Alignment versus wavelength-specific processing in the frozen benchmark. In the PPG-only comparison (Table 3), we hold wavelength-specific processing fixed and toggle alignment in the adapted models. Because wavelength-specific information was already integrated during training, this experiment cannot isolate the effect of wavelength-specific processing. On EarSet with green PPG, alignment lowers CA-MoE MAE from 4.806 to 4.613 bpm, whereas it raises MLP MAE from 4.680 to 4.927 bpm. These results suggest that alignment works particularly well in combination with CA-MoE in these adapted runs, rather than providing a consistent benefit across fusion modules; Revised Table 1 compares the full routed predictor with ACC masked against pretrained baselines with frozen encoders and affine heads, alongside end-to-end ResNet1D.

3. Participant and sample counts. Within each task, every model in Table 1 uses the same participant-level training and validation split, which Table 3 also reuses. For each model and metric, we score each of three training seeds on that fixed validation split and report the mean. We split participants (patients for MIMIC) before indexing windows, and each window is 30s. The counts are in Table 15:

SettingTrain: people / windowsValidation: people / windows
EarSet HR12 / 6,1924 / 2,187
WildPPG HR11 / 261,4842 / 44,223
PPG-DaLiA HR10 / 43,7072 / 8,136
MIMIC BP397 / 4,726,832103 / 295,123
MIMIC AF1,020 / 11,861,141210 / 636,569

For AF, we aggregate window logits within each record before computing AUPRC and AUROC. The AF counts above are windows.

Table 14 distinguishes the source datasets from the analysis cohorts. WildPPG contains 16 participants (216 h), EarSet 30 (about 17 h), and PPG-DaLiA 15 (about 36 h). The MIMIC-III-Ext-PPG v1.1.0 release contains 6,189 patients and 6,399,754 30-second segments. Separately, the historical paired HR validation analysis in Table 2 uses 2/4/2 participants and 44,223/2,187/8,136 windows for WildPPG/EarSet/PPG-DaLiA, respectively.

My concerns have not been resolved. I still have some questions:

  1. The improvement is only 0.05 bpm, which is too small. This improvement may be from the random seed you chose or the hyperparameter settings.

  2. The experiments still do not separate the contribution of wavelength-specific processing.

  3. Is there a test set? You only report the sample counts of the training and validation sets.

I maintain my current score.

Reviewer_Ro5bOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces MSMA-PPG which is framework for physiological prediction from a single selected PPG wavelength combined with an accel, trained on synchronized multi-wavelength recordings. The architecture combines wavelength-specific projections and residual adapters around a shared optical temporal Transformer, a shared-weight full-to-subset alignment branch (stop-gradient full-view target supervises a masked subset view), and a Channel-Aligned Mixture of Experts with availability-masked top-2 routing over optical, motion, and fused experts. The core contribution of this work is enabling the optimization of a single channel PPG to perform nearly as much as a multi channel solution.

Critical strengths

This paper attempts to address an interesting problem which is enabling higher performance of single channel PPG systems (when adding accel) without the the need of a full multi channel system, which has cost and power advantages. Thus, I think the intent of the work is compelling and is solving an important problem.

As I discuss below, the difficultly in the writing made it hard to extract additional strengths.

Critical weaknesses

Honestly, I had a lot of trouble understanding and reading this paper. The writing is very difficult to parse.

The paper motivates the work by highlighting that some current approaches for PPG-based sensing utilize signal conditions to select an optical path for inference, but claims that a high signal quality score may not necessarily identify the channel with the closest error for a given downstream target. This comes across as a rather niche problem to solve since in theory the constraint of only using one channel for inference when multiple are available is self-imposed. More importantly, the rest of the paper does not align with the problem setup. For example, the MSMA-PPG method described doesn’t actually propose an alternative method for optimal channel selection to address the motivating problem. Instead, it proposes a ‘channel alignment and ‘expert routing’ approach to select from channel-specific vs shared representations for predictions, given a pre-selected PPG channel. More confusingly, the results in Table 2 demonstrate that the performance without the proposed channel alignment / mixture of experts (i.e. reconstruction only) outperformed the proposed strategy across all tasks.

The datasets used for evals of AFib and BP do not contain multi-path or multi-wavelength PPG, which would be needed to demonstrate gains in better channel selection. Thus, this is again very inconsistent with the motivation of the work.

Questions for authors

How is the upstream policy, π_ref formulated? How is signal quality computed, what are the energy weights, how is motion used?

What is the advantage of the proposed channel alignment / expert routing method if it cannot outperform reconstruction-only baselines?

3 comments
AuthorsOpenReview ↗

Multi-Wavelength Training, Single-Waveform Prediction, and HR Replay

Thank you for the feedback on clarity. We revised the Introduction to lead with the prediction task and use a green PPG example to explain what the model sees during training and inference. The central question is: Can we use synchronized multi-wavelength recordings during training to improve physiological prediction when the device supplies just one PPG waveform? Some records, such as the MIMIC data used for blood pressure (BP) and atrial fibrillation (AF), contain only a single PLETH waveform. A wearable acquisition mode can impose the same constraint. In either case, MSMA-PPG predicts from the supplied PPG and uses acceleration (ACC) when available.

How training uses the additional wavelengths. Consider a synchronized training window containing red, infrared, and green PPG. We encode the full recorded view, then hide red and infrared from a second pass that sees green PPG and, when available, ACC. The full-view embedding serves as a training target for this subset pass, encouraging it to preserve structure that can be inferred from its own input. Both passes share the encoder; wavelength-specific projections and adapters retain the identity of each optical input. The other alignment losses link synchronized optical views and link PPG to ACC.

At inference, when only green PPG is supplied, CA-MoE weights the green-specific and shared representations computed from that waveform. With green PPG and ACC, the motion expert is also eligible, and the router weights two of the three paths. Neither case requires the other optical waveforms or their quality scores. The dashed zone in revised Fig. 1(a) marks the replay policy before the waveform enters the predictor (§4).

1. Why replay wavelength choices, and how is the reference policy defined? The separate heart rate (HR) replay asks how sensitive a fixed predictor is to the waveform it receives. On the same EarSet windows with ACC held fixed, switching from green to infrared raises pooled HR MAE from 8.566 to 11.119 bpm for the aligned predictor and lowers it from 6.909 to 6.772 bpm for the reconstruction-only predictor (Table 7). This illustrates why the preferred wavelength can depend on the fitted predictor. The replay supplies one waveform at a time using an external policy. The BP and AF experiments instead keep the single recorded PLETH input fixed.

For each recorded wavelength cc, let fcf_c be the fraction of valid finite samples, scs_c the RMS of the centered valid samples, and rcr_c the RMS of consecutive differences where both samples are valid. The quality score is qc=fc/(1+rc/max⁡(sc,10−6))q_c=f_c/(1+r_c/\max(s_c,10^{-6})). Let aa be the RMS over valid samples on the three ACC axes; the motion state is u=log⁡(1+a)/(1+log⁡(1+a))u=\log(1+a)/(1+\log(1+a)). Each RMS adds 10−610^{-6} inside the square root. The reference policy πref\pi_{\mathrm{ref}} selects the observed wavelength maximizing qc−0.25eCc−0.25uMcq_c-0.25eC_c-0.25uM_c, where ee defaults to 0.50.5, C=(1.0,1.1,0.8)C=(1.0,1.1,0.8) are the energy costs, and M=(0.8,0.7,1.0)M=(0.8,0.7,1.0) are the motion penalties in red/IR/green order (Appendix B.3). The predictor receives the selected PPG, its identity and quality, ACC, and the energy and motion states. The replay uses assigned costs; it does not measure hardware power or compare one waveform with simultaneous multi-wavelength prediction.

2. What do alignment and routing add? In Table 2, the historical aligned recipe has higher pooled HR MAE after adaptation on all three datasets. Both adapted models use CA-MoE, and the reconstruction-only recipe turns off Stage II gate regularization as well as alignment. Table 3 separates these comparisons: we toggle the alignment losses with the same training data, fixed inputs, and gate regularization, and independently train a concatenation MLP and CA-MoE within each alignment setting, matching their inputs, backbone initialization, and update counts. With the CA-MoE architecture fixed, alignment reduces EarSet participant-macro validation MAE from 4.806 to 4.613 bpm with green PPG alone and from 4.733 to 4.664 bpm with green PPG and ACC.

With alignment enabled on EarSet, CA-MoE vs. MLP gives 4.613 vs. 4.927 bpm with green PPG alone and 4.664 vs. 4.909 bpm with ACC. Without alignment, the MLP has the lower MAE for both input settings. Comparing the best setting of each module narrows the differences to 0.067 bpm with PPG alone and 0.050 bpm with ACC, in favor of CA-MoE.

On the same fixed PLETH input, aligned CA-MoE also lowers MIMIC SBP/DBP MAE by 0.458/0.619 mmHg and raises AF AUPRC from 0.625 to 0.649 relative to the aligned MLP (Table 3).

Thnanks for the updates and some of the clarifications. The most I can update my score is a 2. Note that the writing of this paper needs to still be dramatically improved.

Thank you for your time, constructive feedback, and reassessment of our submission. We acknowledge that the writing still needs substantial improvement. We will continue to refine the manuscript’s clarity and organization and strengthen the experimental section to better communicate our methodology and findings.

Reviewer_HnCNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MSMA-PPG, a framework for physiological prediction from a single selected PPG wavelength and accelerometer measurements. The method combines wavelength-specific adapters, a shared temporal Transformer, cross-channel representation alignment, and an availability-aware mixture-of-experts router. The authors evaluate PPG-only representation transfer on heart rate (HR), blood pressure (BP), and atrial fibrillation (AF), along with retrospective wavelength-selection and routing experiments.

MSMA-PPG achieves competitive results in the PPG-only transfer benchmark, including the lowest reported HR MAE on EarSet and PPG-DaLiA and the lowest systolic BP MAE. However, in the adapted HR experiments, the aligned model consistently underperforms its reconstruction-only counterpart. Additional experiments investigate sensitivity to wavelength selection and expert routing.

Critical strengths
  • Interesting and practically relevant problem. The paper distinguishes wavelength selection based on signal quality from selection based on downstream prediction accuracy. The matched-window experiments provide useful evidence that the relationship between sensing quality and prediction error depends on the trained model.

  • Broad evaluation and informative diagnostics. The PPG-only comparison spans multiple physiological tasks and includes several recent foundation-model baselines. The routing and sensor-removal experiments also provide useful insights into the dependence of trained models on optical and motion representations.

Critical weaknesses
  • The proposed alignment framework does not demonstrate improvements in the main adapted setting. Table 2 shows that the aligned model performs worse than reconstruction-only training on all three HR datasets, with MAE increases of 0.797, 1.485, and 0.373 bpm on WildPPG, EarSet, and PPG-DaLiA, respectively. Although the aligned representation performs better in some frozen-probe evaluations, this improvement does not translate to the adapted predictor. Moreover, the PPG-only foundation-model comparison does not isolate the contribution of alignment or routing. Consequently, the experimental evidence does not convincingly establish the benefit of the proposed components.

  • Limited validation of generalization and deployment benefits. The paired HR experiments evaluate only eight held-out participants across three datasets, and the paper does not report participant-level uncertainty estimates or multiple training-seed results. Some improvements over the strongest PPG-only baselines are also small, making their robustness difficult to assess. Furthermore, wavelength selection is simulated retrospectively using simultaneously recorded signals, without demonstrating an actual energy-efficient acquisition system.

Questions for authors
  • Can the authors clarify why alignment improves frozen representations on EarSet and PPG-DaLiA but consistently hurts performance after task adaptation? An ablation separating the effects of cross-wavelength alignment, PPG-motion alignment, full-to-subset consistency, and expert routing would help establish which components are beneficial.

  • Can the authors provide uncertainty estimates across participants and training seeds, particularly for the relatively small improvements over the strongest PPG-only baselines? This would help assess whether the reported gains are robust.

  • Does the learned top-2 router outperform a comparably trained non-MoE fusion baseline? The current frozen routing interventions demonstrate dependence on the learned mixture but do not establish whether the proposed routing architecture is preferable to simpler alternatives.

1 comment
AuthorsOpenReview ↗

Alignment Across Readouts, Loss Ablations, and Three-Seed Results

All models in Table 1 use the same participant-level training and validation splits (Table 15), and each reported metric is the mean of three training seeds scored on the same validation split. In Table 1, pretrained baselines such as PaPaGei-S use frozen encoders with affine heads, ResNet1D is trained end to end, and MSMA-PPG retains CA-MoE with ACC masked. This compares prediction pipelines; the matched comparisons below address alignment, readout, and task adaptation.

1. Why might the frozen alignment benefit fail to carry over? On EarSet, replacing the complete prediction module while keeping the backbone frozen moves the alignment contrast from a 0.758 bpm benefit with the affine head to near zero with CA-MoE (+0.057 bpm in MAE). All three Table 17 rows use the same fixed green PPG + ACC inputs, validation split, and definitions of alignment on/off. Errors are participant-macro validation MAE (bpm):

ReadoutAlignment offAlignment onOn − off
Frozen backbone + affine head11.49810.740−0.758
Frozen backbone + CA-MoE5.8965.953+0.057
Adapted backbone + CA-MoE4.7334.664−0.069

Table 17 offers a possible explanation for why the historical frozen-probe benefit did not carry over: changing only the prediction module is enough to bring the alignment contrast close to zero in the matched point estimates. Updating the backbone lowers MAE in both CA-MoE arms, and the alignment contrast stays near zero (−0.069 bpm). The routed contrasts are point estimates from four validation participants.

2. What do the three alignment losses contribute? Tables 3 and 16 give the same all-off to all-on comparison for CA-MoE: participant-macro validation MAE falls from 4.806 to 4.613 bpm with green PPG alone and from 4.733 to 4.664 bpm with ACC. Table 16 removes one loss at a time, leaving the other two on and keeping reconstruction and gate regularization fixed. Each removal is compared with all three losses on, using the same 2,187 validation windows from four EarSet participants.

Loss removedΔMAE, green PPG (bpm)ΔMAE, PPG + ACC (bpm)95% interval, PPG + ACC
Spectral consistency+0.055+0.029[−0.355, +0.293]
PPG–ACC alignment+0.115+0.067[−0.042, +0.173]
Full-to-subset consistency+0.146+0.203[−0.110, +0.573]

Every removal raises the point estimate, so the experiment establish a reliable gain for those individual loss.

3. How much uncertainty can we quantify? Table 1 reports three-seed means for every model on the same validation split. Our improvements over the strongest baseline are statistically significant on several metrics. The Table 16 intervals quantify variation across four participants conditional on the fitted models for the individual loss removals. In the separate historical adapted comparison, A has higher MAE for seven of eight validation participants (Appendix A).

4. How does CA-MoE compare with independently trained simple fusion? Table 3 compares CA-MoE with an independently trained concatenation MLP. The comparison matches available features, participant split, task loss, and 5,000 EarSet adaptation updates, with the same backbone initialization within each alignment arm. With green PPG and ACC, CA-MoE minus MLP is −0.245 bpm with alignment (4.664 vs. 4.909) and +0.019 bpm without it (4.733 vs. 4.714). Comparing the best setting for each gives −0.050 bpm. Overall, CA-MoE shows a modest advantage when paired with alignment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

MSMA-PPG Revision: Single-Waveform Prediction and Matched Evaluation

Thank you for the reviews. We study how synchronized multi-wavelength recordings can supervise prediction when the device supplies just one PPG waveform or worse PPG waveform. A full-view target supervises the subset during training. At inference, CA-MoE weights representations supported by the supplied waveform and any ACC. The single recorded MIMIC PLETH waveform tests this setting for BP and AF; wavelength choice is evaluated separately in retrospective HR replay.

Changes for clarity. The revised Introduction uses a green PPG example to explain what the model sees during training and inference. Figure 1 marks the external wavelength policy as part of the HR replay. Section 6 explains which question each comparison answers: Table 1 benchmarks single-waveform prediction, Table 2 reports the historical HR comparison, and Table 3 tests alignment and fusion under matched conditions.

Evaluation scope. All models in Table 1 use the same participant-level training and validation splits (Table 15). Every reported metric is the mean of three training seeds scored on the same validation split. Pretrained baselines, including PaPaGei-S, use frozen encoders with affine heads; ResNet1D trains end to end. MSMA-PPG retains CA-MoE with ACC masked. Table 1 compares these complete pipelines. Table 3 reuses the same splits to test alignment objectives and fusion modules on common training data.

Alignment and fusion. With CA-MoE fixed on EarSet, alignment lowers participant-macro validation MAE from 4.806 to 4.613 bpm with green PPG alone and from 4.733 to 4.664 bpm with ACC (Tables 3 and 16). The independently trained MLP with ACC gives 4.909 bpm with alignment and 4.714 without it, so the best CA-MoE result is 0.050 bpm below the best MLP result. Without alignment, the MLP has the better validation point estimate in every Table 3 row. The matched results support using alignment and CA-MoE together in these settings.

Why the frozen result may not carry over. In the matched EarSet comparison with green PPG + ACC, switching the frozen readout moves the alignment contrast from −0.758 to +0.057 bpm (Table 17). Updating the backbone lowers both CA-MoE errors and yields a contrast of −0.069 bpm. These point estimates from four validation participants suggest that the frozen-probe gain can shrink under a routed readout. In the historical comparison (Table 2), both adapted models use CA-MoE and policy-selected PPG + ACC, and the aligned recipe yields higher pooled MAE on all three HR datasets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.