A Pulse Wave Is Not a Patient History

TL;DR for operators

A physiological monitor can reveal how a patient’s pulse behaves, but that does not mean it can reconstruct the patient’s medical history. Recognizing pulse regularity, identifying observations associated with medication administration, and inferring past diagnoses require different evidence.

PPG-LM¹ makes this distinction measurable. The model leads 10 of 11 reported zero-shot physiological recognition comparisons, yet achieves only 0.520–0.570 balanced accuracy on eight visit-level clinical-fact tasks. Its multi-level architecture improves physiological representation learning, but broader temporal context does not automatically make clinical history observable.

For developers of clinical monitoring and wearable analytics, the strongest near-term research pathway is connecting waveform observations to searchable physiological descriptions and reusable predictive features. Medication-event association, large-scale retrieval, and clinical decision support require separate validation. The study does not establish readiness for independent clinical diagnosis.

What a Pulse Wave Can and Cannot Reveal

Consider a hospital monitor recording a patient’s pulse throughout an encounter. A short recording may reveal pulse rate, rhythm regularity, and signal quality. Observations collected around a medication administration may contain patterns associated with that clinical event. But determining whether the patient takes antihypertensive medication at home requires information that may not be directly expressed in the pulse waveform.

These questions differ in how directly their answers are represented in the available observations.

Photoplethysmography (PPG) measures blood-volume changes using light, producing a waveform whose timing and shape carry physiological information. The signal is widely available from clinical monitors and wearable devices, making it attractive for reusable AI representations.

The difficulty is that physiological measurements and clinical records describe different aspects of the patient. More monitoring time can improve the representation of physiological trajectories without necessarily revealing diagnoses, medication histories, or laboratory measurements.

PPG-LM tests how far a shared waveform-language architecture can extend across those different information levels.

The results establish a considerable performance gap between recognizing local physiology and identifying broader clinical facts. This gap matters whenever a product combines language-generated descriptions with physiological monitoring: a fluent statement about a waveform does not carry the same evidentiary basis as an assertion about the patient’s clinical history.

Three Levels of Language Supervision

PPG-LM’s central contribution is an architecture that connects physiological signals with language at three temporal levels: individual waveform segments, medication events, and clinical visits.

Its training corpus comes from MC-MED, a Stanford emergency-department dataset containing linked physiological monitoring and structured electronic health records (EHRs). The processed cohort includes 36,430 patients and 53,473 visits. Approximately 73,700 hours of PPG from the patient-disjoint training partition support pretraining.

Rather than relying on a pretrained general-purpose language model to interpret numerical signal samples, the researchers jointly train waveform and text encoders from scratch.

The language supervision is constructed through deterministic templates.

At the segment level, 30-second PPG windows are converted into descriptions using derived measurements such as pulse rate, rhythm regularity, amplitude, signal quality, and waveform morphology. Available charted vital signs can supplement those descriptions.

At the event level, structured medication-administration records identify clinical events. The model combines up to three preceding waveform windows and five following windows, while encoding their signed temporal distances from the recorded administration time. Comparable no-administration intervals serve as controls.

At the visit level, descriptions incorporate recorded demographics, presentation characteristics, triage measurements, selected diagnoses, and home-medication information. The architecture aggregates as many as 256 selected windows and represents elapsed time, observation gaps, and monitoring coverage.

Training proceeds in two stages. The first combines contrastive waveform-text alignment with waveform-conditioned caption generation, adapting the CoCa training approach. The second adds event- and visit-level representations, including explicit matching between physiological trajectories and temporal statements.

This design matters because language supervision supplies different information at each level. A segment description concerns observable waveform attributes; an event description supplies administrative context; a visit description introduces clinical facts that the waveform alone may not reveal.

Local Recognition Improves More Than Clinical-History Inference

The researchers evaluate PPG-LM on MC-MED and two external datasets: MIMIC-III intensive-care recordings and VitalDB intraoperative monitoring data.

The main results differ substantially across the three levels.

Sources: Section 5.1, Table 2; Section 5.2, Tables 5–6. The MC-MED event results display the five highest-scoring PPG-LM medication classes rather than all ten classes.

Local physiological recognition is the clearest success. PPG-LM leads comparisons covering rhythm, signal quality, perfusion, morphology, and respiratory-rate classes. Its only displayed exception is VitalDB respiratory-rate classification, where macro-F1 is 0.329 against 0.330 for the strongest comparator.

Generated physiological descriptions also achieve deterministic fact accuracy of 0.923 on MC-MED, 0.920 on MIMIC-III, and 0.820 on VitalDB. Heart-rate estimates extracted from valid generated captions have mean absolute errors of 3.90, 4.27, and 3.98 beats per minute, respectively.

Yet performance becomes considerably weaker when the question shifts from local physiology to patient history.

For example, recognizing recorded antihypertensive use from visit representations produces balanced accuracy of 0.570 on MC-MED. That exceeds the evaluated alternatives but remains modest in absolute terms. Most visit-level results are only slightly above the 0.500 chance benchmark for balanced accuracy.

Medication-event results occupy an intermediate position. Their AUROC values indicate meaningful discrimination of observations associated with documented administrations. However, clinical indication, concurrent interventions, and timing differences may explain part of those associations. The benchmark does not identify the physiological causal effects of medication.

The evidence therefore supports three different claims: waveform attributes can be recognized relatively well; documented medication context is partly distinguishable; broader clinical history remains difficult to infer reliably.

Temporal Learning Requires More Than Additional Windows

The architecture also addresses a problem that arises when a model moves from short segments to longer observation sequences: learning broader representations can damage earlier waveform-language correspondence.

PPG-LM introduces two complementary mechanisms.

First, it explicitly represents temporal relationships. Medication-event representations incorporate signed time offsets, while visit representations use continuous-time attention and learned latent queries. Temporal statement matching trains the model to distinguish descriptions consistent with observed trajectories from alternatives containing deliberately corrupted temporal facts.

Second, Stage 2 continues segment-level training and penalizes deviation from frozen Stage-1 waveform embeddings. This constrains changes to previously learned local representations while event- and visit-level objectives are optimized.

Controlled component ablations support the contribution of both mechanisms.

Removing temporal statement matching reduces temporal discrimination accuracy from 0.925 to 0.511. Removing signed event offsets reduces event AUROC from 0.707 to 0.617.

The representation-preservation result is particularly consequential. Segment retrieval recall at rank one falls from 0.447 to 0.202 without continued segment supervision, and to 0.015 when both segment supervision and distillation are removed.

These are controlled comparisons from Table 8, conducted on MC-MED validation using the CLIP configuration and matched Stage-2 training budgets. They establish component contributions within that experimental setting, rather than a universal rule for all hierarchical models.

Still, the design implication is concrete: extending a foundation model to longer temporal contexts requires both learning the relationships among observations and protecting valuable capabilities acquired at shorter temporal scales.

Stronger Representations Still Face Absolute Accuracy and Retrieval Limits

PPG-LM’s Stage-2 representations outperform six evaluated PPG and time-series foundation-model baselines across the reported 32 downstream clinical prediction tasks. Stage 2 also improves on Stage 1 in 30 tasks.

These evaluations use frozen model representations with supervised logistic- or ridge-regression probes fitted using labeled patients from each target dataset. They demonstrate representation transfer, not fully label-free clinical prediction.

Absolute performance varies markedly. On MC-MED, the Stage-2 representation achieves AUROC 0.870 for atrial fibrillation classification and $R^2=0.782$ for age regression. Glucose regression, however, achieves only $R^2=0.083$.

The contrast shows why relative benchmark leadership cannot substitute for task-specific adequacy.

A second limitation appears in waveform-text retrieval. With 100 candidates, signal-to-text recall at rank one reaches 0.975 on MC-MED, 0.805 on MIMIC-III, and 0.555 on VitalDB.

Increasing the candidate pool to 2,000 lowers those values to 0.780, 0.258, and 0.108, respectively (Appendix D.3, Table 18).

For a developer building natural-language search over archived monitoring recordings, candidate-pool size is therefore part of the evaluation requirement. Strong retrieval over a small benchmark pool does not establish reliable search over a larger clinical archive.

Caption output strategy presents another practical tradeoff. A supplementary comparison finds that direct caption generation achieves higher deterministic fact accuracy, while nearest-caption retrieval produces fewer unsupported claims and higher language-model-judge factuality scores. On MC-MED, nearest-caption retrieval reduces unsupported claims per caption from 2.94 to 1.51, but lowers fact accuracy from 0.923 to 0.877.

That result favors evaluating generation and retrieval as distinct interface choices rather than assuming one output method dominates every factuality measure.

The Product Boundary Is the Evidence Boundary

The strongest Cognaptus inference is that PPG-LM provides a plausible architecture for waveform-aware clinical information systems, especially where users need to search, organize, or describe monitoring observations.

A monitoring developer could explore waveform-to-text retrieval for locating segments associated with particular physiological characteristics. A wearable analytics developer could evaluate structured descriptions of pulse rate, rhythm regularity, and signal quality as an intermediate representation for downstream applications. A clinical AI developer could test whether frozen waveform features improve specified predictive tasks relative to existing encoders.

These are application pathways suggested by the architecture and benchmark evidence, not demonstrated production benefits.

The boundary becomes sharper when outputs involve a patient’s medical history or treatment response.

If a clinical application needs to establish whether a medication was administered, the administration record remains a distinct evidence source. If it needs a diagnosis or laboratory value, the corresponding clinical information cannot safely be replaced with a waveform-derived inference simply because the model can produce plausible language.

The training process itself reinforces this provenance requirement: some clinical information is supplied through EHR-derived caption supervision rather than measured directly from the waveform. A deployed interface should distinguish observed physiological attributes, retrieved chart facts, and statistical predictions.

Several further conditions remain unresolved. Pretraining comes from one adult emergency-department center; external tests do not establish broad multi-center generalization. Generated training descriptions inherit errors and missingness from measurements and clinical records. Some event and history timestamps are uncertain, and the language evaluations rely partly on templates and automated annotations.

The study also lacks prospective clinical validation, comprehensive subgroup evaluation, and sufficiently established uncertainty calibration for deployment.

Operational evaluation would therefore need realistic retrieval workloads, site-specific validation, explicit data provenance, calibrated uncertainty, and clinician review for consequential outputs.

A Language Interface Does Not Expand What the Sensor Observes

PPG-LM establishes that physiological waveforms can be aligned with clinically grounded language across multiple temporal scopes. Its strongest evidence concerns local physiological recognition, retrieval, and the usefulness of learned representations for downstream prediction.

Its more limited results are equally informative. Medication-event associations are not treatment effects, and increasingly sophisticated visit representations remain weak at recovering many patient-history facts.

For healthcare AI developers, the relevant distinction is between making observed physiology more accessible and treating physiological data as a substitute for the clinical record. PPG-LM advances the former. The latter remains an inference and validation problem that additional language capabilities alone do not resolve.

Cognaptus: Automate the Present, Incubate the Future.