{"id":"eec20b21-75d0-46b1-8c8c-8cd4bb0b374a","arxiv_id":"2502.01108","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A field-trained, open-source PPG foundation model outperforms a clinical-data-trained model on 10 of 11 downstream health tasks across wearable and clinical settings.","lead":"Pulse-PPG is a new open-source artificial intelligence model for analyzing pulse signals from smartwatches, trained on messy real-world data from 120 people over 100 days. The authors report it outperforms a leading clinical-data-trained model on 10 of 11 health tasks, suggesting real-world training data can help models work in labs and hospitals too.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experiment 3's field-vs-clinical pretraining comparison is confounded by corpus scale: the field MOODS corpus has roughly 4x more segments per epoch than the MIMIC-III clinical corpus, so the claimed advantage of field data may be an artifact of dataset size rather than data domain.","rationale":"This is a solid empirical contribution: the model weights are released, the evaluation spans 11 tasks across five datasets, and the Light Pulse-PPG experiment controls for architecture and input window size. Those are real strengths and support the paper's practical value. The load-bearing soft spot is the causal claim about data domain. Experiment 3 is designed to compare field versus clinical pre-training, but the two corpora differ in size by about a factor of four in segments per epoch. Since both models use identical hyperparameters, the field model simply sees more data. This is not an internal inconsistency; it is an uncontrolled variable that prevents the specific claim 'pre-training on field data outperforms pre-training on clinical data' from being attributed to the field setting. The reader's weakest_assumption about the learned motif-based distance function is also legitimate: if the distance function is not semantically grounded, the pretraining signal could encode noise rather than physiology. However, I do not treat that as the primary threat to the central claim, because the downstream transfer results provide indirect evidence that the representations carry useful signal, and the distance-function concern does not explain why field data would beat clinical data at matched architecture. The scale confound, by contrast, directly explains the headline result without invoking any field-data advantage. The proposed matched-size experiment would settle this: if the field advantage persists when scale and steps are equalized, the paper's claim is much stronger; if it disappears, the conclusion should be tempered to a data-scale effect. The verdict remains conditional, as the reader concluded, pending this control.","tokens_in":40487,"tokens_out":6413,"duration_ms":71851,"concrete_test":"Train Pulse-PPG on a matched-size clinical corpus: either subsample the MOODS field pretraining set to 151,738 segments (matching the MIMIC-III count) and compare against the full clinical run, or train the clinical MIMIC-III model for enough epochs/steps to see the same total number of segments as the field model. Use the same optimizer, batch size, and hyperparameters as Section 4.3, with at least 3 random seeds, and report mean and standard deviation for the 11 downstream linear-probe metrics. If field pretraining no longer beats clinical pretraining when scale is matched, the headline claim must be weakened to 'field data at larger scale helps' rather than 'field data helps.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim, that pre-training on field PPG outperforms pre-training on clinical PPG, rests mainly on Section 6.3 (Experiment 3, Table 7). The field corpus is MOODS, with 606,833 unique 4-minute segments per training epoch (Section 4.3.4), while the clinical corpus is the MIMIC-III PPG subset with 151,738 5-minute segments (Section 6.3.2). Both models use the same training procedures and hyperparameters, so the field model sees roughly 4x more pre-training segments and more training steps than the clinical model. The comparison therefore changes two variables at once: data domain (field vs clinical) and pretraining corpus scale. Appendix A.2's Light Pulse-PPG comparison controls model parameters and input window size against PaPaGei, but it does not match pre-training data scale, so it does not resolve this confound. No significance tests or seed variance are reported for this comparison. Unless the clinical model is trained with a correspondingly longer schedule, or the field corpus is subsampled to the same size, the conclusion that field data itself drives the improvement is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Pulse-PPG, a 28.5M-parameter PPG foundation model pre-trained with a relative contrastive learning objective (RelCon) on uncurated 4-minute windows of wrist-worn field PPG from the MOODS study (120 participants, up to 100 days). The authors evaluate frozen embeddings with linear probing and fine-tuning across 11 tasks in five datasets spanning wearable field, wearable lab, and clinical PPG. They report that Pulse-PPG outperforms the open-source clinical PPG foundation model PaPaGei on 10/11 tasks, that field pre-training beats clinical pre-training on the same architecture, and that the released model is a useful general-purpose backbone. The paper also includes ablations on window length, normalization, and a parameter-matched 'Light Pulse-PPG' model.","tokens_in":40775,"tokens_out":4267,"duration_ms":47681,"significance":"If the claims hold, this is a useful contribution to PPG foundation modeling: it provides an open-source model trained on realistic field data, a relatively clean linear-probing benchmark against PaPaGei, and a set of downstream evaluations spanning multiple domains. The authors are also transparent about limitations (label noise in field stress, lack of skin-tone data, single-study pretraining). The central scientific claim—that field-data pretraining itself, rather than scale or architecture, drives the gains—is not yet established because Experiment 3 confounds data domain with corpus size. The paper ships code and weights, which is a concrete strength that should be credited.","major_comments":[{"comment":"The conclusion that pre-training on field PPG outperforms pre-training on clinical PPG is confounded by corpus scale. The field model is trained on 606,833 unique 4-minute segments per epoch for 6 epochs (Section 4.3.4), while the clinical model is trained on 151,738 5-minute segments with 'the same training procedures and hyperparameters' (Section 6.3.2), i.e., the same number of epochs. The field model therefore sees roughly 4x more segments and 4x more gradient updates. The Light Pulse-PPG comparison in Appendix A.2 controls model parameters and input window length, but not pretraining data scale, so it does not resolve this confound. I ask the authors to either match the number of segments or gradient steps across the two pretraining conditions (e.g., subsample MOODS to the MIMIC-III size or extend the clinical training schedule), or to provide an analysis that explicitly separates domain from scale. As written, the paper's headline claim that 'pre-training on field data outperforms its pre-training on clinical data' is not supported by Experiment 3.","section":"§6.3, Table 7"},{"comment":"The pretraining pipeline assumes that the unsupervised masked-reconstruction distance function produces semantically meaningful relative orderings of PPG windows, but this assumption is never validated independently of the downstream tasks. Because this distance function defines all positive and negative relationships in RelCon (Eq. 5), a distance function that primarily encodes reconstruction artifacts or sensor noise would still be consistent with the reported downstream gains if those artifacts correlate with task labels (e.g., motion artifacts correlating with activity). I request an explicit validation or ablation: for example, compare RelCon pretraining using the learned distance against (a) random relative orderings, (b) a fixed hand-crafted PPG similarity such as normalized correlation or beat-level morphology distance, and (c) on a small labeled benchmark, an oracle label-based distance. This would test whether the motif-based ordering, rather than the contrastive framework or the field-data scale, drives the improvements.","section":"§4.1.1, Eq. (2), Eq. (5)"},{"comment":"No statistical significance tests, confidence intervals, or seed variance are reported for any of the headline comparisons. The '10/11 tasks' claims are based on a single run per model, and the per-task differences in Tables 6 and 7 are often small relative to the metric variability across tasks. I ask the authors to report results across at least a few random seeds (or a paired bootstrap across tasks) and to state which differences are statistically reliable. This is especially important for the field-versus-clinical comparison in Table 7, where the number of tasks is small and the effect sizes are modest for some clinical tasks.","section":"§6.1–§6.4, Tables 6–7"}],"minor_comments":[{"comment":"There is a numerical inconsistency: Section 4.2.1 says the MOODS dataset is composed of 822,247 unique 4-minute 50 Hz PPG segments from 122 participants, while Section 4.3.4 says each epoch is composed of 606,833 unique 4-minute PPG segments and that 120 participants were used. Please clarify whether the 606,833 figure excludes the validation/test splits or reflects a different exclusion criterion, and reconcile the participant counts.","section":"§4.2.1 vs. §4.3.4"},{"comment":"The notation in Equation 2 is garbled: 'where X∈R^{T×D} and x∈R^D with T as the time length, ∈S as a∈ but with a subsampling of stride s' is unreadable. Please define the set S, the subsampling operator with stride s, and the dimensions of the query/key/value features precisely.","section":"Eq. (2), §3.2.2"},{"comment":"There are several typos and formatting issues: 'comminmunity' in Section 1, 'Feasability' in the Section 8.5 heading, and the table header 'Quality Type Field Clean Clean Clean Clean' in Table 1, which is malformed.","section":"Throughout"},{"comment":"The axes of Figures 3 and 4 are not labeled. Please label the horizontal and vertical axes with the metric names (e.g., F1 score, MAPE) and add units where applicable so the plots are self-contained.","section":"Figures 3 and 4"},{"comment":"The clinical pretraining corpus is not identical to PaPaGei's original pretraining corpus; the paper notes that PaPaGei did not release its curation code. This limitation should be stated more prominently in the main text of Section 6.3, since it is another potential confound in the cross-model comparison.","section":"§6.3.2 footnote"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the scale confound in Experiment 3, which is fixable with additional experiments or a reanalysis. The comparison against PaPaGei is reasonably fair for a linear-probing evaluation, and the open release of weights is a strong point. I did not find evidence of misconduct; the paper's limitations section is candid. The main revision should focus on isolating domain from scale and on validating the learned distance function."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-documented empirical contribution and the open-source model is genuinely useful, but the paper's central causal claim—field pretraining beats clinical pretraining because the data is field—is not yet established. The comparison in Experiment 3 changes two variables at once.\n\nWhat's new: first open-source PPG foundation model trained exclusively on field data; the field-versus-clinical pretraining comparison is new; the model weights and code are released. The evaluation is careful: linear probing on frozen embeddings, one set of pretrained weights per model, 11 tasks across 5 datasets spanning wearable field, wearable lab, and clinical settings. The benchmark against PaPaGei is fair, and the ablations on window size and normalization are helpful. The paper also benchmarks against general time-series foundation models (Chronos, MOMENT) and supervised baselines. The limitations section is honest about label noise and deployment issues.\n\nSoft spots: the stress-test confound is real. MOODS has roughly 606k 4-minute segments per epoch; the MIMIC-III subset has about 151k 5-minute segments. The field model trains on roughly 4x more data and more steps. Same hyperparameters were used, but corpus scale and domain are entangled. Appendix A.2's Light Pulse-PPG controls for parameter count and input window size against PaPaGei, but not for pretraining corpus size, so it doesn't resolve the confound. No significance tests or repeated-seed variance are reported. The learned distance function is also unvalidated: it's trained by masked reconstruction on field PPG and used to define relative orderings, but there is no independent check that those orderings correspond to physiological semantics beyond the downstream task results. That said, the downstream evaluation is not circular—the tasks are external benchmarks.\n\nRecommendation: worth a serious referee. The artifact and benchmark are valuable, and the field-versus-clinical question is important. A referee should ask for a matched-size retest—subsample MOODS to MIMIC's size, or train the clinical model for more epochs—and report seed variance. The distance function deserves validation or at least an ablation.","headline":"A solid, well-documented open-source PPG encoder with a fair benchmark against PaPaGei, but the headline claim that field data itself beats clinical data is confounded by corpus scale and needs a matched-size retest.","tokens_in":41285,"tokens_out":1936,"would_cite":true,"duration_ms":21161,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that a PPG foundation model trained only on raw, uncurated field data from 120 smartwatch users outperforms a clinical-data-trained model on 10 of 11 health tasks.","keywords":["photoplethysmography","foundation model","wearable health","contrastive learning","self-supervised learning","field PPG","relative contrastive learning","motif-based distance"],"falsifier":"A decisive check would be to take clean clinical PPG beats, compute the learned motif distances between all pairs of windows, and see whether the closest-ranked pairs share independent beat-morphology labels (for example, systolic rise time or dicrotic notch position) significantly more often than random pairs. If the rankings carry no such physiological signal, the pre-training objective is measuring artifacts, not pulse shape, and the reported transfer gains would not be expected to generalize.","tokens_in":40299,"feed_emoji":"🫀","tokens_out":10618,"duration_ms":100042,"temperature":0.7,"pith_summary":"The paper sets out to establish that a PPG (photoplethysmography) foundation model can be built from raw, uncurated smartwatch data collected in daily life, rather than from clean clinical recordings, and that this field-trained model transfers better to both wearable and clinical tasks. It introduces Pulse-PPG, an open-source encoder pre-trained on about 200 million seconds of wrist PPG from 120 participants over 100 days. Across 11 downstream tasks on five datasets, Pulse-PPG outperforms PaPaGei, a state-of-the-art open-source PPG foundation model trained on clinical data, on 10 of 11 tasks; the only exception is sleep disturbance, a domain absent from the field pre-training data. The paper also shows that pre-training the same architecture on a large clinical PPG dataset is worse on 10 of 11 tasks, suggesting that realistic field variability, not signal cleanliness, is what makes the representations transfer.","feed_headline":"Field-trained PPG model wins 10 of 11 health tasks","feed_subtitle":"Pulse-PPG beats clinical-only models after training on raw field PPG.","key_machinery":"The load-bearing object is the learnable motif-based distance function used to compare PPG windows without segmenting beats. A motif is a short temporal shape within the pulsative waveform, such as a systolic rise; the distance function uses a cross-attention reconstruction error in which each motif of the anchor window retrieves the closest motif in the candidate window through a softmax kernel regression and tries to reconstruct itself. Because it is trained on masked field PPG, it learns to match motifs even in noisy, unsegmented signals. The frozen distance function then drives the relative contrastive loss: for every anchor, candidates are ranked by distance, the closest becomes the positive pair, and farther candidates become negative pairs, so the encoder learns fine-grained relative similarity instead of coarse binary similarity. The encoder is a 1D ResNet-26 with instance normalization and global pooling that maps variable-length PPG into a 512-dimensional embedding, and its 127k-parameter distance model keeps the pre-training tractable.","core_discovery":"On the paper's own terms, the central discovery is that noise is information. Pulse-PPG is trained in two self-supervised stages: first, a lightweight dilated-convolution network learns an unsupervised motif-based distance function by reconstructing a masked two-second span of a PPG window from the most similar motifs in another window; second, a 28.5-million-parameter 1D ResNet encoder is trained with a relative contrastive loss that uses this frozen distance to order a candidate set of same-subject and cross-subject windows by their distance from an anchor, pulling the embedding toward relatively close windows and away from relatively distant ones. Frozen embeddings from this encoder, evaluated with linear probes, beat PaPaGei's clinical-data embeddings on 10 of 11 tasks spanning wearable field stress and activity, wearable lab stress and instantaneous heart rate, and clinical blood pressure, hypertension, and sleep disturbance. Re-training the same model on a curated MIMIC-III clinical PPG corpus produces worse results on 10 of 11 tasks, including several clinical tasks. The paper reads this as evidence that exposure to real-world motion artifacts, ambient light, and skin-contact variability teaches a PPG encoder fine-grained, transferable structure that clean clinical data does not provide.","pith_inferences":["Beyond the paper, the same relative-contrastive recipe could transfer to other biosignals (ECG, EMG, respiration) that lack geometric invariances, as long as a domain-specific motif distance can be learned from masked reconstruction.","Beyond the paper, the results imply a testable ranking rule for pre-training data: match deployment noise conditions first, match deployment labels second; a model trained on wrist PPG from one device generation should generalize better to a new wrist device than one trained on finger clinical PPG.","Beyond the paper, the motif-distance function could be validated directly against physiological ground truth by checking whether windows ranked 'closest' share beat-morphology features (systolic rise time, dicrotic notch position) on clean clinical data; the paper does not perform that check.","Beyond the paper, adding overnight PPG from a sleep study to the field pre-training corpus would be a natural extension, and the paper's own field-versus-clinical result predicts that the mixed corpus would beat clinical-only pre-training even on sleep tasks."],"forward_implications":["An open-source, field-trained PPG encoder gives researchers a general-purpose backbone; tasks with small labeled datasets can be solved with a linear probe on frozen embeddings instead of training from scratch.","Pre-training data selection matters more than data cleanliness: field PPG beats clinical PPG even for clinical downstream tasks, so future PPG foundation models should be trained on realistic wearable recordings rather than only curated hospital waveforms.","The one clear failure, sleep disturbance, coincides with a domain absent from pre-training, implying that coverage of target physiology matters at least as much as the pre-training objective.","Because Pulse-PPG has 28.5M parameters and still ranks in the top two against Chronos (200M) and MOMENT (385M) on average metrics, a modest PPG-specific model can compete with much larger general time-series foundation models on physiological tasks.","Fine-tuning Pulse-PPG improves average F1 by 9.6% and average MAE by 19.78% over linear probing, providing a cheap path to task-specific performance."],"supporting_citations":[{"why":"Provides PaPaGei, the open-source clinical-data PPG foundation model used as the main baseline across all downstream comparisons.","marker":"[74]"},{"why":"Supplies the MOODS wearable field PPG dataset, the sole pre-training corpus of roughly 200 million seconds of raw wrist PPG.","marker":"[70]"},{"why":"Introduces the retrieval-based reconstruction idea behind the learnable motif-based distance function used to rank PPG windows.","marker":"[107]"},{"why":"Introduces RelCon, the relative contrastive learning framework that turns ranked distances into positive and negative pairs for the encoder.","marker":"[108]"},{"why":"Provides the curated MIMIC-III clinical PPG dataset used to re-train Pulse-PPG for the field-versus-clinical pre-training comparison.","marker":"[105]"},{"why":"Supplies the PPG-BP dataset used for systolic and diastolic blood pressure, average heart rate, and hypertension downstream tasks.","marker":"[29]"},{"why":"Supplies the SDB sleep-disordered-breathing dataset used for the sleep disturbance task, the one task where Pulse-PPG trails PaPaGei.","marker":"[34]"},{"why":"Supplies PPG-DaLiA, the wearable lab dataset for nine-class activity and instantaneous heart rate regression tasks.","marker":"[84]"},{"why":"Supplies WESAD, the wearable lab dataset for binary and four-class stress classification tasks.","marker":"[91]"}],"fun_headline_variants":["Field PPG beats clinical in 10 of 11 health tasks","Open-source Pulse-PPG: field data wins 10 of 11 tasks","Raw field PPG outperforms clinical training in health AI","Pulse-PPG: trained on noisy field data, tops clinical on 10/11","Real-world PPG training beats clean clinical data for wearables"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the unsupervised motif-based distance function orders PPG windows by real physiological similarity rather than by noise patterns or reconstruction artifacts; if that ranking is semantically empty, the relative contrastive loss trains the encoder to sort noise, and the downstream gains would not survive contact with new datasets.","fun_headline_variants_meta":{"raw":{"variants":["Field PPG beats clinical in 10 of 11 health tasks","Open-source Pulse-PPG: field data wins 10 of 11 tasks","Raw field PPG outperforms clinical training in health AI","Pulse-PPG: trained on noisy field data, tops clinical on 10/11","Real-world PPG training beats clean clinical data for wearables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1482,"prompt_tokens":1014,"completion_tokens":468,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":376}},"tokens_in":630,"tokens_out":468,"duration_ms":4871,"temperature":1.0,"reasoning_tokens":376,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T16:33:03.448918+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check would be to take clean clinical PPG beats, compute the learned motif distances between all pairs of windows, and see whether the closest-ranked pairs share independent beat-morphology labels (for example, systolic rise time or dicrotic notch position) significantly more often than random pairs. If the rankings carry no such physiological signal, the pre-training objective is measuring artifacts, not pulse shape, and the reported transfer gains would not be expected to generalize.","supporting_citations":[{"cited_title":"Momentary Stressor Logging and Reflective Visualizations: Implications for Stress Management with Wearables","cited_arxiv_id":"2401.16307","evidence_quote":"Supplies the MOODS wearable field PPG dataset, the sole pre-training corpus of roughly 200 million seconds of raw wrist PPG."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the retrieval-based reconstruction idea behind the learnable motif-based distance function used to rank PPG windows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the curated MIMIC-III clinical PPG dataset used to re-train Pulse-PPG for the field-versus-clinical pre-training comparison."}],"review_version":1}