acceptodds
Under review as a conference paper at ICLR 2027

DPIKAN: Decomposed Physics-Informed KANs for Physiological Signal Representation Learning

Abstract

Learning representations from physiological signals is challenging because heterogeneous sensing modalities provide indirect and modality dependent observations of underlying cardiovascular dynamics. In this work, we propose the Decomposed Physics-Informed Kolmogorov–Arnold Network (DPIKAN), a compact framework that incorporates physiological knowledge through two mechanisms: multi-scale waveform modeling and cardiovascular descriptor guidance. For multi-scale waveform modeling, DPIKAN decomposes physiological signals into fast- and slow-varying components and assigns them component-specific functional capacities through a dual-order KAN. For cardiovascular descriptor guidance, it integrates cardiovascular descriptors through a physics-informed KAN and regularizes how predictions respond to variations in these descriptors. During experiments, we evaluate DPIKAN for blood pressure estimation on four datasets spanning photoplethysmography (PPG), bioimpedance (BioZ), and millimeter-wave (mmWave) sensing. We further introduce Ring-HGCPT, a smart ring PPG dataset collected during controlled cardiovascular perturbations, and evaluate DPIKAN using leave-one-subject-out cross-validation to assess its generalization to unseen participants. Across these heterogeneous sensing settings, DPIKAN achieves competitive performance for BP estimation. On Ring-HGCPT, it achieves mean errors of 1.08 and 0.09 mmHg for systolic blood pressure (SBP) and diastolic blood pressure (DBP), respectively, with corresponding error standard deviations of 5.41 and 4.81 mmHg. Overall, DPIKAN offers a compact, physiology-informed approach to representation learning from heterogeneous physiological signals. We will release our source code and data.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physiological signal representation learning, physics-informed learning, Kolmogorov--Arnold networks, cuffless blood pressure estimation, wearable sensing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yLjfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DPIKAN, a physics informed KAN approach to estimate blood pressure from physiological signals. The approach consists of decomposing the signal into slow-varying and fast-varying components, integrating biological information such as heart rate in the learning process, and regularizing using a physics informed loss. The approach is evaluated on public datasets and a newly collected Ring PPG dataset which is also released. DPIKAN outperforms different categories of baselines while providing efficiency gains.

Critical strengths

Integration physiological priors in the learning process: Adding physiological information during the training through physics informed KAN is interesting.

Contribution of a new dataset: The new Ring PPG dataset which will be released is a useful resource for the community. Moreover, the efficiency analysis on a Raspberry Pi is a strong evaluation moving us closer toward on-device deployments.

Critical weaknesses

Significance of the contribution to ICLR: The proposed approach is evaluated only for blood pressure estimation. Therefore, the paper's approach is a KAN-based physics informed framework for blood pressure estimation. This is quite narrow and as such it is not a sufficient contribution for ICLR. Moreover, framing the approach as a physiological representation learning implies more tasks, thus it is an overclaim. If the model learns cardiovascular dynamics, then it should be applicable to targets beyond blood pressure estimation

Results: (1) Many results presented Table 1 are only different by a narrow margin. It is unclear if these differences are significant or meaningful. For example, Ring-CPT Corr difference between PITN and DPIKAN is 0.01. Please provide stat. tests or CIs for these results.

Questions for authors
  • DPIKAN seems like a general approach for physiological modelling. What is the rationale behind not choosing large datasets such as MIMIC-III? This could open up larger number of people and different tasks.
3 comments

W1: Thank you for highlighting this distinction. We clarify our scope as follows:

DPIKAN is intended as an application-oriented representation learning framework for physiological BP sensing, rather than a task-agnostic physiological foundation model. Physiological signals contain domain-specific structures, including slow cardiovascular regulation, fast beat-level dynamics, and measurable cardiovascular descriptors. DPIKAN is designed around these properties through physiology-guided decomposition, component-specific KAN capacity, and explicit physiological regularization. BP estimation serves as the target application for evaluating whether the learned representations capture these domain-relevant structures.

W2:

For on-device physiological estimation, computational cost and inference latency are as important as predictive accuracy. Compared with PITN, DPIKAN requires approximately 1/12621/1262 of the FLOPs and achieves a 58×58\times inference speedup on a Raspberry Pi 4B, while maintaining competitive predictive performance. Thank you for requesting uncertainty estimates. We performed paired subject-level bootstrap analysis, treating each participant as the independent unit. We define Δr=rDPIKAN−rPITN,ΔRMSE=RMSEPITN−RMSEDPIKAN\Delta r=r_{\mathrm{DPIKAN}}-r_{\mathrm{PITN}},\quad \Delta RMSE=RMSE_{\mathrm{PITN}}-RMSE_{\mathrm{DPIKAN}}, such that positive values favor DPIKAN.

DatasetTargetΔr\Delta r (95% CI)ΔRMSE\Delta\mathrm{RMSE} (95% CI), mmHg
Graphene-HGCPTSBP0.012 [−0.005, 0.030][-0.005,\,0.030]0.300 [−0.117, 0.717][-0.117,\,0.717]
Graphene-HGCPTDBP0.008 [−0.038, 0.035][-0.038,\,0.035]0.383 [−0.367, 0.950][-0.367,\,0.950]
Ring-CPTSBP0.010 [−0.014, 0.038][-0.014,\,0.038]0.200 [−0.480, 0.740][-0.480,\,0.740]
Ring-CPTDBP-0.032 [−0.096, 0.052][-0.096,\,0.052]0.220 [−0.340, 0.920][-0.340,\,0.920]
Blumio-mmWaveSBP0.006 [0.003, 0.153][0.003,\,0.153]0.432 [0.160, 0.769][0.160,\,0.769]
Blumio-mmWaveDBP0.009 [−0.018, 0.124][-0.018,\,0.124]0.220 [−0.123, 0.487][-0.123,\,0.487]
Blumio-PPGSBP0.003 [−0.043, 0.145][-0.043,\,0.145]0.721 [0.214, 1.334][0.214,\,1.334]
Blumio-PPGDBP0.012 [0.003, 0.188][0.003,\,0.188]0.533 [0.117, 0.871][0.117,\,0.871]

Q1: Additional evaluation on MIMIC-III

Our work targets lightweight cuffless BP estimation using wearable and contactless sensors in non-ICU settings, whereas MIMIC-III primarily contains bedside physiological recordings collected in ICUs. We therefore prioritized datasets that better match our intended sensing modalities and on-device deployment scenario.

As an additional cross-setting evaluation, we use MIMIC-BP, a curated BP-estimation dataset derived from the MIMIC-III waveform database. MIMIC-BP provides synchronized physiological waveforms and reference BP signals extracted from ICU bedside recordings, enabling direct evaluation of waveform-based BP estimation. We have clarified the relationship between MIMIC-BP and MIMIC-III in the revised manuscript. We evaluate DPIKAN on 85-subject MIMIC-BP subset using LOSO.

ModelSBP MAE ↓SBP RMSE ↓SBP ME/SDEDBP MAE ↓DBP RMSE ↓DBP ME/SDE
PINN16.1820.57−0.50/20.58-0.50/20.589.9612.43−1.06/12.39-1.06/12.39
PITN15.5919.891.51/19.851.51/19.859.6712.121.93/11.971.93/11.97
DPIKAN13.5017.050.28/17.060.28/17.068.5510.640.31/10.640.31/10.64

All errors are reported in mmHg. These results extend the evaluation to ICU PPG recordings, while our primary target remains lightweight wearable and contactless sensing.

Thank you again for your feedback. As the discussion period draws to a close, please let us know if you have any remaining questions. We would be happy to address them.

Thank you for the clarifications and additional experiments. However, my concerns are unresolved: (1) framing the method as application-oriented representation learning conflates end-to-end single-task regression with transferable representation learning. If the model genuinely captures underlying cardiovascular dynamics, it should demonstrate transferability to at least one additional downstream task beyond BP estimation. (2) the predictive performance gains over PITN on the primary wearable/contactless datasets remain marginal. Accordingly, I will maintain my current score.

Reviewer_3UE1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DPIKAN for cuffless blood pressure estimation from physiological signals, including PPG, BioZ, and mmWave. The model decomposes the input into fast-varying and slow-varying components and processes them using Chebyshev-KAN branches with polynomial orders b+1b+1 and bb, respectively. The learned representations are combined with three cardiovascular descriptors: pulse amplitude, a morphology-derived PWV proxy, and heart rate. Training incorporates a descriptor-based Taylor consistency loss following PINN/PITN.

The authors also introduce Ring-HGCPT, a smart-ring PPG dataset containing 400 PPG–BP pairs from 20 healthy participants undergoing handgrip and cold pressor tests. Experiments use subject-specific evaluation on Graphene-HGCPT, Ring-CPT, and Blumio, and leave-one-subject-out (LOSO) evaluation on Ring-HGCPT. Additional analyses include component ablations, efficiency measurements on a Raspberry Pi 4, and visualizations of prediction agreement and learned representations.

Critical strengths
  1. Compact architecture across sensing modalities. The same architecture is evaluated on PPG, BioZ, and mmWave without modality-specific architectural changes. The reported parameter reduction relative to PITN, together with comparable or slightly better prediction errors, is practically relevant. Measuring inference latency on a Raspberry Pi 4 also strengthens the deployment-oriented evaluation.
  2. A useful cross-subject benchmark. Ring-HGCPT combines wearable ring sensing with controlled cardiovascular perturbations and LOSO evaluation. This complements the subject-specific evaluations used in prior physics-informed BP studies. Releasing the dataset and collection protocol would support reproducible comparisons, and the authors acknowledge the cohort’s limitations.
Critical weaknesses
  1. Limited novelty. The waveform backbone closely follows TimeKAN, while descriptor conditioning and Taylor regularization follow PINN/PITN. The methodological advance needs clearer justification, and the base polynomial order and sensitivity to alternative order gaps are not reported.
  2. Unclear contribution of waveform representations. Table 3 removes descriptors and the physics loss together, so this ablation does not isolate their contributions. A descriptor-only baseline and ablations on Ring-HGCPT are missing.
  3. Insufficient justification of the physics loss. Eqs. (6)–(7) constrain prediction changes using only descriptor variations, although the waveform input also changes. The assumption behind this constraint needs clarification.
  4. Ambiguous heart-rate definition. Appendix C.1 does not establish why the interval between the derivative landmarks in Eq. (9c) represents a complete cardiac cycle.
  5. Limited generalization evidence and modest gains. The public datasets use subject-specific evaluation, with RMSE gains of 0.2–0.7 mmHg over PITN. LOSO evidence comes from 20 young, healthy participants. The SBP improvement over TimesNet is small (RMSE 5.5 vs. 5.7 mmHg), although the DBP improvement is clearer.
  6. Incomplete dataset documentation. Ring-HGCPT lacks clear window-to-cuff alignment, windows-per-label counts, and input-channel definitions. The quality-control criteria also need explanation, particularly given that only approximately 4 of 20 recorded hours were retained.
  7. Insufficient baseline tuning details. The large gap between PITN and PINN under LOSO, and TimeKAN’s near-zero correlations on Graphene-HGCPT, are difficult to interpret without dataset-specific tuning and model-selection details.
Questions for authors
  1. Physics loss. Why is a descriptor-only Taylor constraint appropriate when the waveform also changes? How are the within-subject/session training pairs formed, what are their temporal gaps on Ring-HGCPT, and are unlabeled training segments included?
  2. Component contributions. Could you add a descriptor-only MLP/KAN and repeat the main ablations on Ring-HGCPT, including a retrained waveform-only model, using the same LOSO splits and comparable tuning budgets?
  3. Descriptor definitions. How are the landmarks in Eqs. (9b)–(9c) paired within or across beats? Why does the interval in Eq. (9c) represent a full cardiac period, and how do the resulting values compare with reference heart rate?
  4. Ring-HGCPT protocol. Please specify window length, cuff alignment, windows per label, the four input channels, and quality-control criteria. Is Fig. 5(b)’s correlation pooled or macro-averaged by subject?
  5. Baseline comparisons. How were baselines tuned and selected on each dataset? Please address the PITN/PINN and TimeKAN discrepancies and report uncertainty for the small differences between leading methods.
  6. Order allocation and efficiency. What base order is used, and how sensitive are results to the branch-order gap? Please provide exact parameter counts, FLOPs, and latency for DPIKAN, PINN, and PITN.
4 comments

R1: Methodological contribution, order allocation, and efficiency (W1 and Q6)

  • Methodological contribution. The architectural differences are summarized below:
MethodPhysiology-guided decompositionComponent-specific KAN capacityTaylor regularizationKAN-based physiological predictor
TimeKAN✗✓✗✗
PINN✗✗✓✗
PITN✗✗✓✗
DPIKAN✓✓✓✓

Unlike TimeKAN’s general-purpose frequency modeling, DPIKAN interprets the decomposed components as slow hemodynamic trends and fast pulsatile variations. Unlike PINN and PITN, it further models these components using capacity-specific KAN branches and a KAN-based physiological predictor. DPIKAN therefore jointly incorporates physiological structure, component-specific functional capacity, and cardiovascular sensitivity constraints. We have clarified these distinctions in the Related Work.

  • Order allocation. We use the base Chebyshev degree b=1b=1, giving Dslow=1D_{\mathrm{slow}}=1 and Dfast=2D_{\mathrm{fast}}=2. Table 3 already contains the following order-allocation ablation:
Slow/Fast ordersGraphene-HGCPT Corr. (SBP/DBP)Graphene-HGCPT RMSE (SBP/DBP)Ring-CPT Corr. (SBP/DBP)Ring-CPT RMSE (SBP/DBP)
1/11/1 (fixed low order)0.39/0.3413.3/11.40.56/0.499.3/6.6
2/22/2 (fixed high order)0.42/0.3912.8/11.40.58/0.499.1/6.7
1/21/2 (DPIKAN)0.61/0.5711.1/9.90.75/0.666.8/5.3

We have revised the labels and discussion of Table 3 to explicitly identify the corresponding order pairs and its role as an order-allocation ablation.

  • Exact efficiency statistics.
ModelParameters (K)FLOPs (K)Raspberry Pi 4B latency (s)
PINN109.30277.300.004
PITN (base)299,980.00299,970.002.902
DPIKAN8.22237.600.050

Relative to PITN, DPIKAN uses 99.997% fewer parameters, approximately 1/12621/1262 of the FLOPs a 58×58\times inference speedup We have corrected the previously reported parameter-reduction percentage in the revised manuscript.

R2: Component contributions (W2 and Q2)

We implement a descriptor-only baseline following the variants using identical LOSO splits and training settings:

VariantWaveformDescriptorsSBP Corr./RMSEDBP Corr./RMSE
Descriptor-only KAN✗✓0.57/12.600.49/9.66
Waveform-only✓✗0.67/13.560.32/10.76
Full DPIKAN✓✓0.90/5.500.88/4.80

These results have been added to Appendix.

R3: Physics-loss assumption and pair construction (W3 and Q1)

The physics loss is a descriptor-conditioned soft regularizer, rather than a model of the complete waveform transition. The waveform representation is learned through the supervised prediction loss and serves as the conditioning context of fθ(X,u)f_\theta(\mathbf X,\mathbf u), whereas the Taylor term constrains only the local sensitivity of the prediction with respect to the explicit cardiovascular descriptors u\mathbf u. The underlying assumption is that, over adjacent cardiac cycles, these descriptors capture the dominant BP-relevant variation, while waveform changes not summarized by u\mathbf u are handled by the supervised waveform pathway and treated as residual variation rather than an exact physical constraint. We have clarified this partial-derivative interpretation and local-smoothness assumption around Eqs. (6)–(7).

On Ring-HGCPT, pairs are formed between adjacent cardiac cycles within the same cuff-aligned recording. Their median interval is 0.64 s (IQR: 0.52–0.80 s), and no pair crosses recording, participant, session, or protocol-stage boundaries. All pairs are extracted from cuff-aligned training recordings; no additional recordings without cuff-based labels are introduced.

R4: Descriptor landmarks and heart-rate definition (W4 and Q3)

All landmarks are extracted from the same segmented cardiac cycle; AA and JJ denote its start and end boundaries rather than two arbitrary derivative landmarks.

Cardiac cycles are segmented using the waveform derivative: the maximum-slope point identifies the systolic upstroke, and the preceding zero crossing defines the cycle boundary. Within each cycle, AA is the waveform foot, CC is the systolic maximum, and their amplitude difference defines ΔA\Delta A. Points BB and FF mark the characteristic arrivals of the forward systolic and reflected waves, respectively; therefore, 1/(tF−tB)1/(t_F-t_B) is used as a single-site PWV proxy. Finally, AA and JJ are the left and right boundaries of the same complete beat, so tJ−tAt_J-t_A is one cardiac period and HR=60/(tJ−tA)HR=60/(t_J-t_A) when time is measured in seconds. An independent ECG-based HR reference was not collected; therefore, HR is used only as a waveform-derived physiological descriptor.

We have clarified these landmark definitions and the within-beat pairing in Appendix C.1 and the corresponding figure.

R5: Ring-HGCPT protocol and evaluation scope (W5, W6, and Q4)

Ring-HGCPT is designed to evaluate wearable BP monitoring under controlled cardiovascular variations and unseen-user settings, rather than clinical diagnosis or medical-device validation.

  • Input channels and reference labels. Each input contains 2,9502{,}950 time samples per channel across four PPG channels, acquired using two sampling slots for red light (660 nm) and two for green light (527 nm). Each cuff-based blood pressure measurement is paired with one non-overlapping PPG window, yielding 20 labeled pairs per participant and 400 pairs in total.
  • Acquisition time and quality control. The approximately 20 hours refer to the cumulative experimental-session duration, whereas PPG was acquired only during predefined cuff-aligned windows totaling approximately four hours. No PPG was recorded during the remaining task, rest, and preparation periods; thus, these periods were not excluded for signal-quality reasons. All 400 acquired windows were retained without additional post hoc screening. We have clarified this acquisition procedure consistently in Section 4 and the Appendix.
  • Correlation in Fig. 5(b). We computed the reported correlation by pooling all held-out predictions and their corresponding cuff-based reference measurements across the LOSO folds.

These protocol and evaluation details have been added to Section 4 and Appendix.

R6: Baseline tuning and statistical uncertainty (W5, W7, and Q5)

For on-device physiological estimation, computational cost and inference latency are as important as predictive accuracy. Compared with PITN, DPIKAN requires approximately 1/12621/1262 of the FLOPs and achieves a 58×58\times inference speedup on a Raspberry Pi 4B, while maintaining competitive predictive performance. Thank you for requesting uncertainty estimates. We performed paired subject-level bootstrap analysis, treating each participant as the independent unit. We define Δr=rDPIKAN−rPITN\Delta r=r_{\mathrm{DPIKAN}}-r_{\mathrm{PITN}} and $\Delta\mathrm{RMSE}=\mathrm{RMSE}{\mathrm{PITN}} -\mathrm{RMSE}{\mathrm{DPIKAN}}$, such that positive values favor DPIKAN.

DatasetTargetΔr\Delta r (95% CI)ΔRMSE\Delta\mathrm{RMSE} (95% CI), mmHg
Graphene-HGCPTSBP0.012 [−0.005, 0.030][-0.005,\,0.030]0.300 [−0.117, 0.717][-0.117,\,0.717]
Graphene-HGCPTDBP0.008 [−0.038, 0.035][-0.038,\,0.035]0.383 [−0.367, 0.950][-0.367,\,0.950]
Ring-CPTSBP0.010 [−0.014, 0.038][-0.014,\,0.038]0.200 [−0.480, 0.740][-0.480,\,0.740]
Ring-CPTDBP-0.032 [−0.096, 0.052][-0.096,\,0.052]0.220 [−0.340, 0.920][-0.340,\,0.920]
Blumio-mmWaveSBP0.006 [0.003, 0.153][0.003,\,0.153]0.432 [0.160, 0.769][0.160,\,0.769]
Blumio-mmWaveDBP0.009 [−0.018, 0.124][-0.018,\,0.124]0.220 [−0.123, 0.487][-0.123,\,0.487]
Blumio-PPGSBP0.003 [−0.043, 0.145][-0.043,\,0.145]0.721 [0.214, 1.334][0.214,\,1.334]
Blumio-PPGDBP0.012 [0.003, 0.188][0.003,\,0.188]0.533 [0.117, 0.871][0.117,\,0.871]
  • Baseline configurations. All baselines used the same preprocessing and subject splits within each dataset. For PITN, PINN, and TimeKAN, we used the hyperparameter settings reported in their respective papers. On Graphene-HGCPT and Ring-CPT, training for DPIKAN, PITN, and PINN stopped when the physics loss reached 0.010.01 or below or after 2,0002{,}000 epochs, whichever came first.
  • PITN and PINN on Ring-HGCPT. We observed difficulty training PITN to convergence under leave-one-subject-out evaluation. PITN achieved SBP and DBP correlations of 0.230.23 and 0.210.21, with RMSEs of 11.911.9 and 10.110.1 mmHg. Under the same subject splits, PINN achieved correlations of 0.830.83 and 0.800.80, with RMSEs of 5.65.6 and 5.75.7 mmHg.
  • TimeKAN on Graphene-HGCPT. We verified the near-zero correlations reported in Table 1. TimeKAN achieved correlations of 0.010.01 for SBP and −0.02-0.02 for DBP, with RMSEs of 14.514.5 and 12.612.6 mmHg, respectively.

Thank you for the detailed rebuttal and additional experiments.

The new Ring-HGCPT ablation addresses my concern about whether the learned waveform representation adds value beyond the descriptors alone. I also appreciate the clarification of the adjacent-beat pairing and the pooled correlation in Fig. 5(b), as well as the additional LOSO evaluation on MIMIC-BP.

For my first remaining concern, Eq. (7) compares predictions across adjacent beats with changing waveform inputs, while the Taylor expansion varies only the descriptors. Could you repeat the “w/o Physics Loss” ablation on Ring-HGCPT, retaining both inputs and setting γ = 0 under the same LOSO splits? This would isolate the contribution of the regularizer on Ring-HGCPT.

For my second concern, please specify where each input window starts and ends relative to cuff inflation and the reference reading. If 200 Hz is the effective per-channel sampling rate, 2,950 samples correspond to 14.75 seconds per window, or about 1.64 hours across 400 windows. Could you confirm the effective per-channel sampling rate and clarify how these input windows relate to the approximately four hours of acquired PPG?

For my third concern, could you report participant-level paired confidence intervals for the SBP and DBP RMSE differences between DPIKAN and PINN and between DPIKAN and TimesNet? The reported SBP RMSE improvements are 0.1 and 0.2 mmHg, respectively, and the intervals would help assess the precision of these comparisons.

Overall, while the rebuttal partly addresses my original concerns, I still have questions about the contribution of the regularizer and the Ring-HGCPT evaluation. Given these remaining issues, I am inclined to maintain my original rating of 2 (Weak Rejection).

Best, Reviewer 3UE1

Q1: Effect of the physics loss on Ring-HGCPT

With both inputs retained, removing the physics loss increased RMSE and reduced Pearson correlation for both SBP and DBP. We retrained DPIKAN without the physics loss while retaining both waveform inputs and cardiovascular descriptors. Both variants used the same 20 LOSO splits and the same metric aggregation as Table 2.

VariantSBP Corr. ↑SBP RMSE ↓ (mmHg)DBP Corr. ↑DBP RMSE ↓ (mmHg)
DPIKAN w/o physics loss0.2110.370.169.77
Full DPIKAN0.905.50.884.8

Q2: PPG sampling rate and alignment with cuff measurements

PPG recording and cuff inflation started simultaneously. We downsampled the PPG from 200 Hz to 100 Hz per channel, so each 2,950-sample input window starts at inflation onset and spans the next 29.5 seconds. The SBP and DBP readings from that same cuff measurement serve as the window’s reference labels. The exact time at which the cuff displayed the readings is unavailable.

The 400 input windows cover 3.28 hours of PPG in total. We will correct the earlier estimate of approximately four hours to 3.28 hours in the manuscript.

Q3: Participant-level paired confidence intervals on Ring-HGCPT

Thank you for requesting these uncertainty estimates. We computed paired 95% bootstrap confidence intervals for the SBP and DBP RMSE differences between DPIKAN and PINN and between DPIKAN and TimesNet. We resampled the 20 held-out participants with replacement, kept each participant’s results paired across models, and recomputed the RMSE difference using the same aggregation as in Table 2. Positive differences indicate lower RMSE for DPIKAN.

BaselineTargetRMSE difference (mmHg)95% CI (mmHg)
PINNSBP+0.10[0.012, 0.188]
PINNDBP+0.90[0.020, 1.780]
TimesNetSBP+0.20[0.024, 0.376]
TimesNetDBP+1.70[0.020, 3.380]
Reviewer_cSWgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Decomposed Physics-Informed KANs (DPIKAN) for cuffless blood pressure estimation from PPG, BioZ, and mmWave signals. The method separates physiological signals into fast and slow components, processes them using dual-order Chebyshev KANs, and incorporates cardiovascular constraints through a physics-informed loss. The authors also introduce Ring-HGCPT, a new dataset collected from 20 subjects under handgrip and cold pressor tests. Experiments on three public datasets and Ring-HGCPT show that DPIKAN achieves competitive or state-of-the-art performance while using 97.3% fewer parameters than PITN. The authors further demonstrate efficient inference on a Raspberry Pi 4B.

Critical strengths
  1. Well-Motivated Capacity Allocation and Lightweight Efficiency: Decoupling physiological signals into fast morphological residuals and slow trends and matching them with asymmetric Chebyshev polynomial degrees is intuitive and highly parameter-efficient. DPIKAN reduces parameter count by 97.3% over the runner-up PITN and demonstrates practical real-time latency on a Raspberry Pi 4B.

  2. Strong cross-subject generalization: Tested across PPG, BioZ, and mmWave, DPIKAN consistently beats 10 baselines. More importantly, under the LOSO split on Ring-HGCPT (Table 2), previous physics-informed baselines like PITN fail completely (Corr. 0.23/0.21), whereas DPIKAN maintains 0.90/0.88 correlation.

  3. Valuable New Benchmark Dataset: The newly collected Ring-HGCPT dataset provides synchronized multi-wavelength ring PPG, tri-axial accelerometry, and controlled hemodynamic perturbations (handgrip and cold pressor tests) across 20 subjects, offering a useful testbed for wearable BP estimation.

Critical weaknesses
  1. Missing key hyperparameters: Although line 648-656 claim all hyperparameters are in Appendix C, the paper never reports the base degree bb, number of layers LL, hidden size DmodelD_{\text{model}}, physics loss weight γ\gamma, or window length TT.

  2. Questionable Taylor locality: First-order Taylor approximation (Eq. 6–7) assumes adjacent segments ii and i+1i+1 are close in time. But Ring-HGCPT only collects cuff BP at the end of the stages separated by 1 to 3.5 minutes across abrupt cold-pressor/handgrip transitions (line 299-301). Applying local linear Taylor expansion across multi-minute gaps with sharp hemodynamic jumps is questionable, please clarify how Lphysics\mathcal{L}_{\text{physics}} is actually computed on Ring-HGCPT.

Questions for authors

Non-existent figure reference: Appendix D.2 (lines 966–971) spends a paragraph analyzing "Fig. 6(c)", but Fig. 6 only has subfigures (a) and (b).

2 comments

W1: Missing hyperparameters

Thank you for identifying this omission. The DPIKAN hyperparameters used in our experiments are summarized below:

HyperparameterSetting
Base Chebyshev degree bb11
Slow-trend / fast-residual degreesDslow=b=1D_{\mathrm{slow}}=b=1; Dfast=b+1=2D_{\mathrm{fast}}=b+1=2
Number of representation layers LL22
Hidden dimension DmodelD_{\mathrm{model}}3232
Physics-loss weight γ\gamma11
Input length TT, public datasets3232 data points per beat
Input length TT, Ring-HGCPT2,9502{,}950 time data points with four PPG channels

For Ring-HGCPT, each fixed-length segment is aligned with one cuff-based BP reference; we do not generate overlapping sliding windows. We have added these settings, together with the dataset-specific input configurations, to Appendix C of the revised manuscript.

W2: Taylor locality on Ring-HGCPT

The Taylor pairs are formed between adjacent cardiac cycles within the same continuous recording. Across the evaluated datasets, the median cardiac-cycle duration ranges from 0.64 to 0.87 s; thus, ii and i+1i+1 are separated by only one heartbeat and constitute a local beat-to-beat transition, rather than cuff measurements separated by 1–3.5 min. We additionally quantified this temporal scale:

DatasetNumber of cyclesMedian duration (s)Interquartile range (s)
Graphene-HGCPT24,8290.730.65–0.82
Ring-CPT6,5440.680.64–0.72
Blumio-PPG22,9070.840.73–0.95
Blumio-mmWave18,5060.870.77–0.98
Ring-HGCPT16,9210.640.52–0.80

For the public datasets, cycle duration is calculated from the extracted heart-rate descriptor as T=60/u3T=60/u_3. For Ring-HGCPT, cardiac cycles are automatically detected within each cuff-aligned 2,950-sample PPG recording. Taylor pairs are formed only between adjacent cycles from the same recording; no pair crosses recording, participant, session, or protocol-stage boundaries.

The resulting Taylor step is approximately 0.6–0.9 s. Each i→i+1i\rightarrow i+1 pair preserves temporal locality within a continuous physiological recording. The cuff measurement supplies the reference label for the supervised BP loss but does not define the Taylor interval. We have clarified this construction around Eqs. (6)–(7) and in Appendix.

Q1

Thank you for identifying this typographical error. We have corrected “Fig. 6(c)” to “the right panel of Fig. 6(b)” and updated the associated description accordingly.

Thank you again for your feedback. As the discussion period draws to a close, please let us know if you have any remaining questions. We would be happy to address them.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Author Responses and Revisions

Thank you to the Area Chair and reviewers for the discussion. We summarize the main findings and clarifications below.

  • Scope of the contribution. DPIKAN learns task-specific representations for cuffless blood pressure (BP) estimation by combining slow–fast waveform modeling with explicit cardiovascular descriptors and a physiology-informed regularizer. The additional MIMIC-BP evaluation extends BP estimation to ICU PPG recordings.

  • Accuracy–efficiency trade-off. For cuffless BP estimation on resource-constrained wearable and edge devices, computational cost and inference latency are core evaluation criteria alongside predictive accuracy. On the primary wearable and contactless datasets, DPIKAN achieves competitive accuracy with modest RMSE improvements over PITN. On a Raspberry Pi 4B, DPIKAN reduces measured inference latency from 2.902 s to 0.050 s, a 58× speedup over PITN.

  • Effect of the physics loss on Ring-HGCPT. We retrained DPIKAN without the physics loss, retaining both waveform inputs and cardiovascular descriptors and using the same LOSO splits. Removing the loss increased SBP/DBP RMSE from 5.5/4.8 to 10.37/9.77 mmHg and reduced Pearson correlation from 0.90/0.88 to 0.21/0.16.

  • Ring-HGCPT acquisition protocol. Each input window starts at cuff inflation onset and spans 29.5 s at the effective input rate of 100 Hz per channel. The BP reading from the same cuff measurement serves as the reference. The 400 windows total 3.28 hours, which corrects our earlier estimate of approximately four hours.

We have revised accordingly in our manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.