acceptodds
Under review as a conference paper at ICLR 2027

HARP: Learning Heterogeneity-Aware Representations for Physiological Health Sensing

Abstract

Health-sensing data supporting classification of clinical and everyday health states are naturally heterogeneous, where they differ in sensor sets, temporal resolutions, and missingness patterns across datasets. Despite advances in time-series pretraining, existing models mainly expect that target data is complete and regularly sampled. The transfer across real-world heterogeneous sensing configurations and the adaptation without target outcome labels remain challenging. We introduce HARP, a pretrained framework for health classification across heterogeneous sensing configurations. HARP uses masked-reconstruction pretraining over 50k participants with diverse sensor sets and observation length to enable learning from incomplete patterns. By combining physical duration patch embedding with canonical variable embedding, consistent temporal and variable meaning across datasets are preserved. For downstream classification, HARP supports both supervised adaptation and zero-shot transfer by a lightweight sensor-language bridge when target labels are unavailable. Across six clinical and wearable heterogeneous sensing dataests, HARP consistently outperforms state-of-the-art pretrained time-series models, improving average performance over the best pretrained model by 12% under linear probing and 10.3% under zero-shot transfer, while achieving competitive performance on three regularly sampled benchmarks. Beyond predictive performance, interpretability and scaling analyses show that HARP preserves temporal and variable meaning across various configurations and benefits from additional unlabeled pretraining data, supporting the learning of reusable heterogeneity-aware representations for real-world physiological sensing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Physiological health modeling, Representation learning

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_AVZROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a pretraining framework for cross-dataset multimodal irregular health sensing, which supports both supervised finetuning and zero-shot transfer for downstream tasks. The pretraining is done through masked reconstruction, where time series signals from different sensor modalities are encoded through a shared encoder. The authors use a uniform binning to account for irregularities in the different sensor modalities, and augment the embeddings with physical-duration patch embedding and temporal and variable identity embedding. After pretraining, zero-shot transfer is achieved through a sensor–language bridge, which aligns embeddings with template captions. The proposed framework performs better than the baseline methods in 5 out of 6 downstream datasets (two of which are used for self-supervised pretraining).

However, as the paper stands, there are critical flaws in the experiment design that casts doubt on the validity of the results and the conclusions that rest upon them. I will expand on this in the weakness section. That being said, I do credit the authors for their significant effort in providing the current results and I see promise in this work should the flaws be addressed.

Critical strengths
  1. The paper provides reasonably well controlled evidence that heterogeneity-aware design offers real gain. Comparing with MIRA in Table 1, which is the only baseline pretrained on the same datasets as HARP, we can clearly see the advantage offered by the authors' heterogeneity-aware design. The fact on regularly sampled, complete benchmarks (Table 7) the gains mostly vanish also helps sharpen this point.
  2. The authors provide direct evidence that their proposed pretraining recipe outperform other strategies. Table 10 shows that under identical data access, HARP beats the two baseline methods on all datasets and on both metrics.
Critical weaknesses
  1. The baseline choices in Table 1 are questionable.
  • The supervised reference (ATENet) is weak and excluded without proper justification. MTM (2509.17809) reports 88.0/58.6 on P12 and 90.3/58.3 on P19 (AUROC/AUPRC). From Table 3 in MTM, other methods such as ViTST also exceed ATENet on P19. The authors should include stronger baselines or provide reasonable justifications of why ATENet was the only supervised reference.With ATENet trained on the same full set of target labels used in linear probing, the author's excluding it from the ranking seems not sufficiently justified. One potential argument is that ATENet was trained end-to-end while HARP's main weights were frozen and only a linear projection head was allowed to move. To sustain this argument, the authors should provide experiments comparing supervised baselines with full-weight finetuned HARP.
  • Pretraining data is confounded with method. According to Appendix A.3, MIRA was the only baseline in Table 1 that was pretrained on the same datasets as HARP. Even so, since MIRA was significantly larger than HARP (113.4M vs 2.0M in params), it could be argued that MIRA was undertrained on such small datasets. Table 11 suggests that pretraining data plays a significant role in downstream performance, further necessitating removing this confound in fair comparisons with other methods.
  • Several closely related pretraining methods are missing from the baselines. STraTS (2107.14293) and DuETT (2304.13017) are close relatives to HARP's design and both are absent from comparisons. Moirai (2402.02592) is cited but not evaluated.
  • Some basic controls are absent. 2602.19531 shows that summary statistics + XGBoost achieves 85.7 AUROC on P12 and 90.0 on P19. Less critical omissions include HARP trained from scratch, HARP full-weight (end-to-end) finetuned, and a random-init frozen encoder linear probing.
  • The paper does not clarify how irregular sensor signals are fed into regular-grid models. In particular, NormWear's 4.45 AUPRC on P19 suggests a potential pipeline failure instead of model limitation.
  1. The zero-shot transfer performance is weak and baselines are near-chance. Table 2 shows that most LLM and TSLM baselines have near-chance (in some cases, worse than chance) performance. Beating these baselines shows little.
Questions for authors
  1. Add the SOTA supervised models (e.g. MTM) under the same protocol, and include them in the ranking or provide more justifications for exclusion from ranking.
  2. Add dataset-matched and capacity matched baselines and the missing baselines from the earlier weakness section.
  3. Please consider narrowing the scope of your claims about zero-shot transfer given the absolute performance and poor baseline performances.
7 comments
AuthorsOpenReview ↗

Response to supervised baseline and basic controls (W1.1, W1.4, Q1)

General response:

We thank the reviewer for providing insightful comments about well-controlled evidence and baseline selection. Our primary question is how to learn transferable representations across heterogeneous sensing configurations and adapt them to health classification when target outcome labels are limited or unavailable. In response to the reviewer’s concerns about

  • 1). supervised baseline selection and full-weight fine-tuning evaluation,
  • 2). the selection of other baselines (such as DuETT and Moirai), and
  • 3). a controlled pretraining-data comparison, we provide additional analyses and experiments to clarify our baseline choices and better address our research question.

The details are as follows:


W1.1 & Q1. Baseline selections

Thank you for these suggestions. We have added MTM (KDD ‘25) as a stronger supervised baseline, following its official implementation under our splits and metrics. Our reproduced results (close to those reported) are in the following table. From the newly added results, we observe that MTM is a stronger supervised reference than ATENet. HARP with linear probing still achieves competitive performance.

We exclude supervised models from the Table 1 ranking and use them as references mainly because of a difference in protocol. Supervised models are trained from scratch and update all model parameters using labels. In contrast, Table 1 evaluates frozen pretrained encoders, training only a linear classifier. Since our main question is whether a pretrained encoder can be used on unseen heterogeneous datasets, the fair comparison is among pretrained foundation models under the same frozen-encoder protocol. Under this protocol, HARP outperforms existing foundation models by 5.4 points on average. We will add this clarification underlying Table 1.

W1.4 Basic controls and statistical baselines

Thank you for pointing the baseline and control setting. We have added the requested controls, all evaluated with the same splits, label access, and metrics:

  • Summary statistics + XGBoost, following its official implementation and using our validation splits and metrics.

  • HARP from scratch vs. full fine-tuning. With the same architecture, fine-tuning from the pretrained encoder outperforms training from scratch on all six datasets. This isolates the benefit of pretraining. With full fine-tuning, HARP achieves the best AUROC on five of the six datasets, including against MTM.

ModelMIMIC-III AUROCAUPRCMIMIC-IV AUROCAUPRCP12 AUROCAUPRCP19 AUROCAUPRCPMData AUROCMF1PAM AUROCMF1
statistics+XGBoost86.21 ± 0.8742.76 ± 2.2987.78 ± 0.7151.61 ± 1.7486.32 ± 1.1253.71 ± 5.0186.88 ± 2.4347.71 ± 4.9752.98 ± 2.5119.97 ± 0.7999.81 ± 0.0296.14 ± 0.38
Supervised, trained end to end (reference)
ATENet (NeurIPS'25)85.14 ± 0.9939.37 ± 3.5987.18 ± 0.5249.40 ± 1.2885.92 ± 0.6853.83 ± 1.8584.02 ± 1.3841.16 ± 3.0257.31 ± 3.2117.54 ± 3.6299.07 ± 0.4894.80 ± 1.60
MTM (KDD'25)86.87 ± 1.0543.92 ± 2.4687.64 ± 0.5351.62 ± 1.1887.62 ± 1.7657.10 ± 4.1089.89 ± 1.8057.32 ± 3.2755.53 ± 6.7416.50 ± 1.3099.86 ± 0.0596.90 ± 0.46
HARP, same architecture from scratch85.47 ± 1.2139.09 ± 2.7087.84 ± 0.4850.28 ± 1.4182.93 ± 1.4347.44 ± 2.6688.95 ± 1.0955.79 ± 2.9355.15 ± 2.5316.88 ± 0.7099.73 ± 0.0796.50 ± 0.57
HARP pretrained on MIMIC-IV + P12
HARP, frozen encoder + linear probe (Table 1)87.27 ± 0.6642.97 ± 1.8788.26 ± 0.2451.94 ± 0.9085.23 ± 1.8151.64 ± 2.9687.80 ± 2.6050.55 ± 3.5360.04 ± 4.3821.06 ± 1.7599.66 ± 0.0996.03 ± 0.74
HARP, full fine-tune88.43 ± 1.2145.29 ± 3.1488.70 ± 0.4254.65 ± 1.2386.32 ± 1.3655.14 ± 3.1389.95 ± 2.0759.71 ± 5.2658.37 ± 1.8525.80 ± 2.5599.89 ± 0.0696.95 ± 0.63

Bold: best; italic: second-best.

AuthorsOpenReview ↗

Response to pretraining data confound (W1.2 & Q2)

W1.2 and Q2. Pretraining data confound

We thank the reviewer for this point. Table 1 supports two distinct claims, and only the second requires controlling for pretraining data:

  1. Existing pretrained models transfer poorly to unseen heterogeneous datasets. Table 1 evaluates each model’s released checkpoint with linear probing, testing whether existing models can be reused as they are. Here, each model’s own pretraining data is part of what is being evaluated, so matching the pretraining data is not required.
  2. HARP’s pretraining mechanism is more effective than that of alternative irregular time-series foundation models. We agree that this claim is confounded by pretraining data and model capacity.

To address the concern about the second claim, we additionally pretrained the baseline MOMENT from random initialization on the same data as HARP (MIMIC-IV + P12), using its official implementation, and evaluated it under the same linear-probing protocol. Results are shown below (AUROC / AUPRC).

With the same pretraining data, HARP achieves the highest mean AUROC across all six datasets in this comparison. These results indicate that the performance gain of HARP comes from its heterogeneity-aware pretraining design, which embeds temporal and channel identity, rather than from pretraining data or model size. We have revised Section 4.1 to distinguish this controlled comparison from the comparison using released checkpoints.

ModelMIMIC-III AUROCAUPRCMIMIC-IV AUROCAUPRCP12 AUROCAUPRCP19 AUROCAUPRCPMData AUROCMF1PAM AUROCMF1
Pretrained on MIMIC-IV + P12, linear probing
MOMENT76.86 ± 0.9522.84 ± 1.6078.88 ± 0.8032.27 ± 1.2271.86 ± 2.4429.12 ± 2.8081.23 ± 2.0741.14 ± 5.1753.81 ± 6.2420.10 ± 1.4296.80 ± 0.7082.73 ± 2.90
HARP87.27 ± 0.6642.97 ± 1.8788.26 ± 0.2451.94 ± 0.9085.23 ± 1.8151.64 ± 2.9687.80 ± 2.6050.55 ± 3.5360.04 ± 4.3821.06 ± 1.7599.66 ± 0.0996.03 ± 0.74

Regarding the MIRA pretraining concern, we observe that MIRA's pretraining had converged: on every retained checkpoint, its forecasting loss is constant on training windows (0.00053) and on held-out patients (0.00049 on MIMIC-III, 0.00056 on PhysioNet-2019), with no train–validation gap, while the learning rate decays sevenfold.

AuthorsOpenReview ↗

Response to related pretraining method (W1.3) and baseline data processing (W1.5)

W1.3 other related pretraining methods

DuETT: Since DuETT is designed for self-supervised pretraining on a single dataset, we compare in a matched in-domain setting: DuETT and HARP are each pretrained and evaluated on the same target dataset, using DuETT’s official code and the same variable set as all other baselines (e.g., 12 variables for MIMIC-III and MIMIC-IV, following [1]). We did not add STraTS separately, as DuETT reports outperforming it in its original paper [2]. Results are shown below.

In the matched in-domain setting, HARP outperforms DuETT on four of the six datasets, and the differences on MIMIC-III and P12 are within one standard deviation. More importantly, the single HARP encoder from Table 1, pretrained only on MIMIC-IV and P12, outperforms DuETT pretrained on each target dataset on all six datasets, including four never seen during pretraining. DuETT, by design, requires a fixed input variable set and must be pretrained separately for each dataset, whereas HARP transfers one encoder across datasets with different variables and sampling patterns.

DatasetDuETT, in-domainHARP, in-domainHARP (Table 1)
MIMIC-III85.92 ± 0.6585.46 ± 1.2687.27 ± 0.66
MIMIC-IV87.31 ± 0.6088.06 ± 0.2688.26 ± 0.24
P1285.01 ± 1.4984.61 ± 1.2585.23 ± 1.81
P1986.31 ± 2.5186.86 ± 2.3187.80 ± 2.60
PMData56.89 ± 2.8060.55 ± 3.6660.04 ± 4.38
PAM (macro-F1)95.09 ± 1.3997.74 ± 0.4296.03 ± 0.74

Bold: best; italic: second-best.

Bold: best; italic: second-best.

Moirai: we added Moirai as another time-series foundation model since it pretrained on different datasets and not fixed to specific data inputs.The performance is as shown. HARP achieves higher mean AUROC than Moirai on all six datasets, with larger gains on the clinical benchmarks. Same as other time-series foundation models trained on regular datasets, it cannot transfer to heterogeneous irregular datasets, further supporting the effectiveness of HARP’s heterogeneity-aware representations. We will include Moirai in the revised Table 1.

ModelMIMIC-III AUROCMIMIC-III AUPRCMIMIC-IV AUROCMIMIC-IV AUPRCP12 AUROCP12 AUPRCP19 AUROCP19 AUPRCPMData AUROCPMData MF1PAM AUROCPAM MF1
Chronos-278.27 ± 0.8026.91 ± 1.1580.49 ± 0.3635.46 ± 1.7572.83 ± 1.8337.33 ± 1.5981.31 ± 2.9640.67 ± 4.0756.87 ± 3.5819.61 ± 0.7190.67 ± 0.6489.79 ± 0.52
(new)Moirai78.64 ± 0.7126.34 ± 1.8679.68 ± 0.5633.90 ± 1.2768.78 ± 1.0226.66 ± 0.9179.69 ± 1.7537.59 ± 4.9954.88 ± 2.6823.31 ± 0.5498.91 ± 0.1791.98 ± 0.71
HARP87.27 ± 0.6642.97 ± 1.8788.26 ± 0.2451.94 ± 0.9085.23 ± 1.8151.64 ± 2.9687.80 ± 2.6050.55 ± 3.5360.04 ± 4.3821.06 ± 1.7599.66 ± 0.0996.03 ± 0.74

[1.] Shukla and Marlin, 2021. “Multi-Time Attention Networks for Irregularly Sampled Time Series”, ICLR.

[2.] Labach et al., 2023. “DuETT: Dual Event Time Transformer for Electronic Health Records”, MLHC.


W1.5 Clarification of regular-grid processing.

  1. We have added details to the Appendix A.3 on how irregular records are presented to regular-grid baselines. These baselines take fixed-length inputs split into fixed-step patches. Following mTAN, each variable is represented as a sequence of length 512: its observed values are placed in temporal order, and the remaining positions are zero-padded. For example, a MIMIC-III patient with 216 heart-rate observations has these values in the first 216 positions and zeros in the remaining 296. We then apply each baseline’s own preprocessing and patching: MOMENT uses patches of 8 steps, and SLIP applies mean-absolute normalization, uses patches of 8 or 16 steps, and excludes empty patches when pooling.

  2. We thank the reviewer for flagging the NormWear result on P19. We found that its evaluation used only the first 12 hours of each record instead of the full 60-hour window used by all other methods. With the full window, NormWear achieves 79.38 ± 1.88 AUROC and 32.34 ± 4.31 AUPRC, in line with other TSFMs, and HARP remains higher. We will replace the erroneous result and update the affected aggregate gains. | Model and dataset | AUROC | AUPRC | |---|---|---| | NormWear, P19 (60 h) | 79.38 ± 1.88 | 32.34 ± 4.31 | | HARP, P19 (60 h) | 87.80 ± 2.60 | 50.55 ± 3.53 |

Reviewer_ud6POpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes HARP, a pretrained encoder for health time series, robust to heterogeneous variable sets, sampling intervals, missingness, and record lengths across datasets. It was pretrained with masked reconstruction on MIMIC-IV and P12, followed by a sensor-language bridge for label-free classification. The authors report linear-probe and zero-shot gains over time-series foundation models and time-series language models on six benchmarks.

Learning from heterogeneous health records (time series) is a practically meaningful problem. However, I am not convinced that the current comparisons support the claimed gains, as well as a bit concerned about the key design choices.

Critical strengths
  • The problem is important, and the tokenisation that combines value, mask, time, and variable identity is simple and easy to follow
Critical weaknesses

Clarity and Presentation

  • the authors would be encouraged to discuss the main differences between these genearl time series foundation models (TSFM) compared to the models typically develped for healthcare sensing. I believe one main differences is that generalt TSFM is trained for forecasting only, and some are therefore designed particularly for forecasting task (like using decoder only architecture), but it remains open how to derive good representatons from a decoder-only model.
  • I don't think Figure 1c is a fair list, since there are many differences also discussions between the model on these numeric health values and physiological waveforms. I believe the authors would like to focus on the former, but some methods like NormWear listed is focused on the latter. There are many differences on how to model them.
  • Relatedly, the pretraining corpus is MIMIC-IV and P12 only, with no wearable or waveform data, although Figure 1a shows HRV and step signals and the abstract refers to "50k participants with diverse sensor sets" (L29).

Methodology

  • here are several things that are unjustfied, for instance the choice of r in L197. Looking further into Figure 9, I am even more concerned, as 60ish min is okay for some less frequently sampled data (for instance vital signs), but is that okay for very higher-frequency ones? and average them within the bin will have any potential pitfalls?
  • Also, I am just a bit curious about the approach of averaging values within a bin, as the measurement frequency is treated as a very good predictor in ICU, but averaging it will remove such information such measurement freuqency right?

Experiment and Results

  • I don't think the comparison so far can lead to the conclusion confident enought to say these Gains (like in Table 1) are valid.
  • I am a bit concerned about the presentation of Table 11. HARP pretrained on P12 even performed worse on P12, compared to zero-shot of Chornos-2?

Minor Issues and Suggestions

  • Figure 1(b) could be enhanced if the number in the bracket () 's meaning could be clarified.
  • WESAD shows in Figure 1(b) but is not used in the experiments.
Questions for authors

I would expect some clarification on:

  1. How were sparse, irregular ICU records put to each checkpoint baseline (imputation, masking, timestamps, pooling)? Do the gains hold against baselines fine-tuned or continue-pretrained on the same ICU corpus?
  2. How should r and the token duration be chosen for high-frequency signals?
5 comments
AuthorsOpenReview ↗

Response to clarity and presentation comments W1&W2

General response

We thank the reviewer for providing insightful comments. HARP focuses on building a robust encoder that learns heterogeneity-aware representations that can transfer across heterogeneous sensing configurations with different variables, sampling patterns, and missingness, and on adapting these representations to health classification when target outcome labels are limited or unavailable. In response to the reviewer’s concerns, we add clarifications and additional analyses to:

  • 1). explain and clarify the differences and similarities between general TSFMs and health-sensing foundation models;
  • 2). clarify our baseline selection, which is diverse and valid for supporting our claims;
  • 3). clarify the details of patch processing
  • 4). add a full-weight fine-tuning comparison on the ICU datasets.

The details are as follows.


Clarity and Presentation

W1.

We thank the reviewer for the suggestion and have added a discussion to Section 2 and a taxonomy of all baselines (architecture, pretraining objective, data domain) to Appendix A.3 (see the table below).

The table shows that pretraining objectives and architectures vary within both general TSFMs and healthcare models, rather than separating the two groups. Among general TSFMs, MOMENT is pretrained with masked-patch reconstruction, while Chronos-2 and Moirai are pretrained for forecasting, and all three are encoder-based. Among healthcare models, NormWear and HARP use masked reconstruction, while MIRA is a decoder-only model pretrained with autoregressive forecasting. The most consistent difference between the two groups is therefore the pretraining domain: general TSFMs are pretrained on multi-domain corpora (e.g., energy, weather, economics), whereas healthcare models are pretrained on health data. Several general TSFMs nonetheless report evaluations on healthcare data, which is why we include them as baselines. Both groups also operate on the same data type, multivariate time series, and can be evaluated under the same downstream protocol.

We agree that deriving good representations from decoder-only models remains an open question. Among our state-of-the-art baselines, only MIRA is decoder-only. To give it a fair evaluation, we follow the representation-extraction protocol of the original MIRA paper [1]. We have added this detail to Appendix A.3. More generally, finding the best way to extract sequence-level representations from decoder-only models is beyond the scope of this work, and we note it as a direction for future study.

ModelPretraining objectivePretraining architecturePretraining time-series domainEvaluation time-series domain
MOMENTmasked-patch reconstructionencoder-basedHealthcare, engineering, finance, etc.Electricity, weather, health, etc.
NormWearmasked reconstructionencoder-basedHealthcareHealthcare
Chronos-2forecastingencoder-basedWeather, energy, economics, etc.Healthcare, energy, economics, etc.
Moiraiforecastingencoder-basedHealthcare, energy, economics, etc.Healthcare, energy, economics, etc.
MIRAautoregressive forecastingdecoder-onlyHealthcareHealthcare
HARPmasked reconstructionencoder-basedHealthcareHealthcare

[1.] Li et al., 2025. “MIRA: Medical Time Series Foundation Model for Real-World Health Data”, NeurIPS.


W2.

We agree that numeric health values and physiological waveforms raise different modeling challenges, and this heterogeneity is precisely what motivates our work. HARP does not focus only on numeric records; it aims to learn general representations that transfer across heterogeneous health time series, whose variable sets, sampling intervals, and missingness differ across datasets, from irregularly sampled ICU records to signals from wearable and clinical sensors (as stated in the Introduction). Our evaluation reflects this scope: in addition to numeric records (P12, P19, MIMIC-III, MIMIC-IV), PAM consists of IMU signals, and Table 7 further evaluates on IMU (HAR), EEG (Sleep-EDF), and ECG (PTB-XL) data. Waveform models such as NormWear are therefore relevant baselines for our setting.

We also clarify that Figure 1c is not a performance ranking. It visualizes the general question our paper asks in the Introduction: can a pretrained time-series encoder be transferred or reused on a target dataset whose variables, sampling intervals, and missingness differ from those it was pretrained on, and which may be unlabeled? Accordingly, Figure 1c marks the three capabilities this question requires: the cross-dataset capability, the kind of data the model was pretrained on, and zero-shot capability.

AuthorsOpenReview ↗

Response to clarity and presentation comments W3 and methodology W1&W2

W3.

Pretraining: We would like to clarify that HARP is pretrained on clinical records (MIMIC-IV and P12) but evaluated on both clinical and wearable/waveform data, reflecting our goal of learning representations that transfer across heterogeneous health-sensing settings.

We chose MIMIC-IV and P12 because they provide large cohorts (59,748 and 11,988 patients, respectively) with diverse physiological variables (e.g., heart rate, respiratory rate), irregular observations, and missingness. Although they differ from wearable and waveform data in source and sensor, they share the same nature: multivariate physiological time series whose variable sets, sampling intervals, and missingness vary across settings. HARP is specifically designed to learn representations that are not tied to a specific sensor configuration.

The downstream results support this design: although wearable and waveform data are never seen during pretraining, HARP transfers effectively to PMData, PAM, and the HAR, Sleep-EDF, and PTB-XL datasets (Table 7).

We understand that “diverse sensor sets” (L29) and the wearable signals in Figure 1a may suggest otherwise. We have revised the wording in the abstract to “diverse clinical variable sets” and clarified in the Figure 1a caption that the wearable signals illustrate target settings rather than pretraining data.


Methodology

W1.

As described in Section 3.1 (Data Preparation), a sensing record contains measurements of different variables taken at different times and rates, while a masked-reconstruction encoder needs inputs of fixed length. We therefore place each record on uniform bins of width r, a standard step for irregular clinical series [1].

How r is chosen. We set r per dataset according to its native recording interval: we use the finest width, no finer than the recording interval, that keeps each record within a few hundred bins. This gives 10 min for MIMIC-III, MIMIC-IV, P12, and PMData, and 60 min for P19. For PAM, each record contains 600 observations at 100 Hz. We retain one raw step per bin on its pseudo-time axis, which corresponds nominally to 10 ms, so adjacent samples are not averaged.

Averaging within a bin. Some within-bin aggregation is unavoidable when variables with different sampling rates are aligned on a shared time grid, which is standard practice for irregular time series. We therefore choose r such that aggregation is rare: it occurs in 1.8% of non-empty bins on P12, 2.3–2.5% on MIMIC-III and MIMIC-IV, and never on P19. Empty bins are not imputed but are marked by the observation mask. We agree that averaging can smooth short transients and lose within-bin timing. We have clarified the dataset-specific settings in Appendix A.5.1 and noted this limitation in Section 5.

[1] Harutyunyan et al., 2019. “Multitask learning and benchmarking with clinical time series data”, Scientific Data.


Q2.

How should r and the token duration be chosen for high-frequency signals?

The general principle for choosing r and the token duration is based mainly on two dataset properties: the missing rate and the sequence length. We aim to choose r such that, after averaging, little fine-grained temporal information is lost.

For high-frequency signals, r should stay close to the sampling interval, so that averaging within a bin loses little fine-grained temporal information. The token duration then follows from the sequence length, so that a record is covered by a manageable number of patches.

Therefore, in HARP, r and the token duration are parameters set per dataset rather than learned. For PAM, the high-frequency dataset in our evaluation (600 observations per record at 100 Hz [1]), r is one raw step (Table 4). As for duration, unlike other datasets such as MIMIC-III, PAM has no physical clock, so its samples are placed on a pseudo-time axis with one sample per unit (Appendix A.5.1), and patches are formed on that axis. The patch embedder accommodates the resulting patch width by resizing its input weights (Section 3.1).

We agree these parameters are currently chosen manually. Learning r and the token duration adaptively is a promising extension in future work.

[1] Zhang et al., 2022. “Graph-Guided Network for Irregularly Sampled Multivariate Time Series”, ICLR.


W2.

We agree that measurement frequency is a strong predictor in the ICU, and this is one of the reasons our encoder keeps it. As discussed above, averaging happens only in a small number of bins, so most bins keep their original measurement frequency. Averaging is applied only to the values that fall in the same bin. More importantly, which positions hold a measurement is stored in the observation mask, so how often and when each variable was measured is preserved at resolution r. The encoder uses this pattern directly: a token with no measurement is replaced by a learned “mask” token, and a partly observed token carries its per-bin mask.

AuthorsOpenReview ↗

Response to experiment and results and questions

W1.

We understand that this concern builds on the points above. In summary, we clarify the following protocol details to support the fairness of the comparison in Table 1 and confidence in the gains:

  • Baseline choice (W1, W2 of clarity and presentation): we compare against general TSFMs (e.g., Chronos-2, MOMENT, Moirai), health-sensing foundation models (e.g., NormWear), models designed for irregular observations (e.g., MIRA), and TSLMs (e.g., SLIP, OpenTSLM). They are evaluated under the same frozen-encoder protocol, splits, and metrics, and our evaluation includes waveform data, so waveform models are relevant baselines.

  • Pretraining data (W3 clarity and presentation)): HARP never sees wearable or waveform data during pretraining, so its gains on these datasets are a stricter test of generalization to unseen tasks.

  • Preprocessing (W1 of methodology): r is set by a fixed rule, not tuned on downstream results to favor our method, and averaging affects at most 2.5% of non-empty bins.

  • Measurement frequency (W2 of methodology): measurement frequency is preserved and used directly by the encoder.

Moreover, the gains are consistent across five splits, matched target-domain adaptation (Table 8), zero-shot evaluation (Table 2), and held-out waveform datasets (Table 7), which supports the validity of the results in Table 1.

We hope these clarifications address the reviewer’s concern, and we would welcome any further comments on specific aspects of the comparison that remain unclear.


W2.

Table 11 examines how different pretraining datasets affect HARP’s downstream performance under the linear-probing protocol. It does not report zero-shot results.

When pretrained on P12 alone and evaluated on P12, HARP achieves an AUROC of 84.61 ± 1.25 (Table 10), higher than Chronos-2’s linear-probing performance (AUROC 72.83 ± 1.83).

We assume the reviewer is referring to the first row, with 70.74 AUROC on P12, where HARP is pretrained on P19 only and is slightly worse than Chronos-2. This may be because HARP is pretrained on a single dataset in this row, whereas Chronos-2 is pretrained on multi-domain datasets.


Minor issues

  1. Thank you for the suggestion. The number in brackets is the number of variables (sensor channels) in the dataset: 13 for WESAD and 36 for PhysioNet-2012. Each row of the panel is one variable. We have added this clarification to the Figure 1 caption.

  2. WESAD appeared in Figure 1(b) only as an example of a regularly sampled dataset, to contrast with the varying intervals in P12. To keep the figure consistent with the experiments, we will updated Figure 1(b) to use PTB-XL instead.


Q1.

How were sparse, irregular ICU records put to each checkpoint baseline (imputation, masking, timestamps, pooling)? Do the gains hold against baselines fine-tuned or continue-pretrained on the same ICU corpus?

  1. We have added details to the Appendix A.3 on how irregular records are presented to regular-grid baselines. These baselines take fixed-length inputs split into fixed-step patches. Following mTAN, each variable is represented as a sequence of length 512: its observed values are placed in temporal order, and the remaining positions are zero-padded. For example, a MIMIC-III patient with 216 heart-rate observations has these values in the first 216 positions and zeros in the remaining 296. We then apply each baseline’s own preprocessing and patching: MOMENT uses patches of 8 steps, and SLIP applies mean-absolute normalization, uses patches of 8 or 16 steps, and excludes empty patches when pooling.
  2. We additionally compare full-weight fine-tuning of MOMENT, SLIP, and HARP on P12 under the same evaluation setting, and we find that the performance gain holds when baselines are fully fine-tuned. The results are shown below. HARP has the highest mean AUROC and AUPRC among the three models.
ModelAUROCAUPRC
MOMENT (full-weight fine-tuning)78.76 ± 0.6332.34 ± 2.06
SLIP (full-weight fine-tuning)73.6 ± 0.732.25 ± 1.22
HARP (full-weight fine-tuning)86.32 ± 1.3655.14 ± 3.13
Reviewer_bGU5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HARP learns representations for heterogeneous physiological time series using physical duration patches, explicit timing information and shared variable identities. The encoder is pretrained with masked reconstruction. For classification without target labels, a lightweight bridge aligns the frozen sensor encoder with a frozen language model using automatically generated record descriptions.

Critical strengths

The problem is relevant, and the proposed design is sensible for sparse clinical records. I also appreciate the relatively small encoder, at about 2M parameters.

The clinical results are encouraging. In Table 1, HARP improves AUROC over the strongest pretrained baseline on average across the clinical datasets. More importantly, the paper includes useful controls: MIRA is pretrained on the same corpus, and Table 8 applies the same bridge and unlabeled target-data access to alternative encoders. HARP still leads in that setting. These comparisons make the empirical contribution more convincing.

Critical weaknesses
  1. The novelty is mostly in the representation design, and its most distinctive part needs a more direct ablation. Masked reconstruction, flexible patch embeddings and sensor–language contrastive alignment all have existing precedents. I see the main contribution as combining physical duration patches with shared variable identities and missingness information for clinical transfer. The time and variable-identity ablations show that these inputs help. However, they do not directly establish the benefit of defining patches by physical duration. A matched comparison between fixed-step and fixed-duration patching, keeping the remaining inputs and training setup unchanged, would strengthen this argument.

  2. In Table 2, HARP trains its bridge on unlabeled target-training records and uses these records for standardization and calibration. The original LLM/TSLM baselines do not access that split or receive comparable adaptation. This is a reasonable target label-free setting, but the performance gap combines differences in representations and target-domain adaptation.

  3. The paper reports five folds and train/validation/test proportions, but it is unclear whether splitting is grouped by patient or participant. Pleass clarify whether encoder pretraining is performed separately for each fold, excluding the corresponding evaluation records.

Questions for authors

see above

4 comments
AuthorsOpenReview ↗

Response to W1&W2

General response:

We thank the reviewer for recognizing the relevance of the problem, the compact encoder, and the value of the controlled comparisons. HARP aims to learn transferable representations across heterogeneous sensing configurations with different variables, sampling patterns, and missingness, and to adapt these representations to health classification when target outcome labels are limited or unavailable. In response to the reviewer’s concerns, we provide:

  • 1). ablation studies of patch processing using fixed-step versus fixed-duration patches;
  • 2). a controlled ablation comparison in the zero-shot setting to demonstrate the effectiveness of the adaptation technique;
  • 3). clarification of fold-specific pretraining and data splits.

We provide extra analysis and experiments to strengthen the evidence for our representation design. The details are as follows.


W1. Fixed-duration patching ablation

".. A matched comparison between fixed-step and fixed-duration patching, keeping the remaining inputs and training setup unchanged, would strengthen this argument."

As the reviewer suggested, we compare fixed-step and fixed-duration patching to support our representation design using physical-duration patches. We therefore added two variants with the same pretraining and evaluation protocols. This comparison holds the remaining configuration fixed between the two reduced variants, which differ only in only how patches are formed:

  • Fixed-step patches: Each patch groups a fixed number of consecutive observed values[1]. With irregular observations, these patches can span different amounts of elapsed time.
  • Fixed-duration patches (ours): Each patch covers a fixed physical-time window.

We observed performance drop when replacing fixed duration (proposed) to fixed step, reduces performance by 1.32 points on average across the eight metrics. Fixed-duration patching achieves higher mean AUROC. These results support the benefit of physical-duration patching in this setting. Together with the original time and channel-identity ablations, they provide further evidence for HARP’s representation design. We have added the results to Table 3.

ModelAvg ΔMIMIC-III AUROCMIMIC-III AUPRCMIMIC-IV AUROCMIMIC-IV AUPRCP12 AUROCP12 AUPRCP19 AUROCP19 AUPRC
HARP (full)–87.27 ± 0.6642.97 ± 1.8788.26 ± 0.2451.94 ± 0.9085.23 ± 1.8151.64 ± 2.9687.80 ± 2.6050.55 ± 3.53
(new)Fixed-step patches; w/o explicit time and channel identity−6.3083.94 ± 0.4435.53 ± 1.9783.65 ± 0.4443.96 ± 0.5776.95 ± 2.0441.02 ± 2.9983.53 ± 2.5046.71 ± 3.58
(new)Fixed-duration patches; w/o explicit time and channel identity−4.9884.88 ± 1.1437.92 ± 2.6184.38 ± 0.9843.44 ± 2.3778.06 ± 1.8742.47 ± 2.2785.43 ± 3.1649.26 ± 5.28

[1]. Nie et al., 2023. “A Time Series is Worth 64 Words: Long-term Forecasting with Transformers”, ICLR.


W2. Target-adaptation control

We added an ablation that applies HARP’s bridge protocol to the SLIP and OpenTSLM encoders, so that all three models receive the same target-domain adaptation. Specifically, each time-series encoder remains frozen, and only the bridge is trained. All three models use the same unlabeled target-training records, standardization procedures, and test splits, with no target labels. Any remaining gap therefore comes from the representations alone.

The results are shown below. We observe that bridge adaptation improves OpenTSLM on all six datasets and SLIP on four. Moreover, HARP still achieves the best average performance among all methods and the best result on five of the six datasets, with an average gain of 4.53 points over the second-best baseline. Together, these results demonstrate that learning heterogeneity-aware representations drives HARP’s generalization to unseen target data. We updated the manuscript and have included the expanded comparison in Table 8 to explicit the fair adaption.

CategoryModelMIMIC-IIIMIMIC-IVP12P19PMDataPAM
TSLMs as released (Table 2)SLIP (Gemma3-270M)55.59 ± 0.4953.52 ± 0.6655.34 ± 1.5253.20 ± 2.8549.14 ± 3.4556.89 ± 0.66
OpenTSLM (Llama-1B)55.00 ± 1.5950.81 ± 2.3551.64 ± 3.0352.90 ± 2.2548.31 ± 0.8149.82 ± 0.89
Frozen encoder + shared zero-shot bridgeSLIP encoder + bridge65.76 ± 1.4858.71 ± 1.1954.89 ± 3.2955.70 ± 2.0452.01 ± 5.1852.28 ± 1.15
OpenTSLM encoder + bridge59.64 ± 1.7361.04 ± 2.2356.02 ± 2.8358.60 ± 1.4251.24 ± 1.4553.80 ± 1.39
HARP encoder + bridge67.88 ± 1.6464.03 ± 2.3962.09 ± 3.7363.11 ± 1.6554.58 ± 4.9455.80 ± 2.14
AuthorsOpenReview ↗

Response to W3

W3. Clarification of data split

Data splits.

  • For P12, P19, and PAM, we use the five train/validation/test splits released with Raindrop [1]. In P12 and P19, each sample is one patient, so these splits are grouped by patient, which means no patient appears in more than one of the training, validation, and test sets.
  • For MIMIC-III and MIMIC-IV, we follow the 5-fold stratified protocol of mTAN [2] for fair comparison, in which each sample is one hospital admission. Since some patients are admitted more than once, these splits are not grouped by patient.
  • PAM’s released splits are drawn at random over windows. We keep them unchanged so that our numbers are directly comparable with prior results on this benchmark. For PMData, we follow the same window-level splitting.

We acknowledge that admission- and window-level splits allow the same subject to appear in both training and test data. However, all methods are evaluated on identical splits, so the comparisons remain fair. While we follow prior work, we have added this clarification and the following table to Appendix A.2, highlighted in blue, and will acknowledge this limitation.

Pretraining.

For each of the five folds, the encoder is pretrained on the unlabeled training data of MIMIC-IV and P12 for that fold. We select the checkpoint with the lowest pretraining loss on the fold’s held-out validation set. Test samples are never used during pretraining. We then evaluate the selected encoder on the fold’s test set with either linear probing or zero-shot adaptation, and report the mean ± standard deviation over the five folds. We have added this clarification to Appendix A.4.

DatasetSample unitSample sizePatient sizeSplit protocolTrain / Val / TestGrouped by patient / participantIn pretraining
P12patient11,98811,988 patientsRaindrop, 5 published splits9,590 / 1,199 / 1,199YesTrain split, unlabelled
P19patient38,80338,803 patientsRaindrop, 5 published splits31,042 / 3,880 / 3,881YesNo
MIMIC-IIIhospital admission53,21142,441 patientsmTAN, 5-fold stratified34,054 / 8,514 / 10,643No (split over admissions)No
MIMIC-IVhospital admission77,17259,748 patientsmTAN, 5-fold stratified49,389 / 12,348 / 15,435No (split over admissions)Train split, unlabelled
PAM600-sample window5,3338 subjectsRaindrop, 5 published splits4,266 / 533 / 534No (split over windows)No
PMDataday1,72216 participants5-fold over days1,032 / 345 / 345No (split over days)No
AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your comments. We hope our responses have addressed your questions. If any concerns remain, we’d appreciate the chance to discuss them before the discussion period closes.

I appreciate the authors' rebuttal during the discussion period. The new analysis help clarify and strengthen the presentation. However, taking into account the overall novelty and contribution, I feel my initial rating already represents the upper bound for this submission. Therefore, I prefer to maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.