PaperlayerView exact revision on arXiv

Source-linked AI summary

CALM-BP: Observation-Matched Physiological Semantic Grounding for Non-Contact Blood Pressure Estimation

Haiyang Sun, Boyuan Gu, Yongjie Liu

arXiv:2608.21744v1cs.HC

TL;DR

Non-contact BP estimation has measured rPPG evidence but limited semantic structure for judging window reliability and cue fusion. CALM-BP constructs bounded language priors from the same rPPG window and uses them as fusion guidance; across the evaluated results and controls, benefits are greatest when semantic guidance remains observation-matched.

  • Problem

    Numerical rPPG pipelines leave a window’s quality, reliability, waveform, and state-aware context largely implicit, making the prediction path harder to audit.

  • Method

    CALM-BP encodes measured rPPG features and promptized priors from the same forehead observation, then aligns them before SBP and DBP inference.

  • Results

    CALM-BP improves on the subject-disjoint FlowBP-Set split, transfers directly to external targets, and benefits most when semantic priors remain matched to the current observation.

  • Takeaways & Limitations

    The findings position CALM-BP as a semantic-guided physiological signal estimation framework, not an independent language-based BP predictor.

  • Takeaways & Limitations

    Effectiveness depends on reliable rPPG and predefined descriptors, which cannot recover physiological information absent or severely corrupted in the measured signal.

Abstract

from arXiv · show

Language grounding increasingly involves non-text observations whose structure is not naturally expressed as words or objects. We study this problem for physiological time series in non-contact blood pressure (BP) estimation: remote photoplethysmography (rPPG) provides measured evidence about bodily state, but numerical pipelines expose little semantic structure about why a window is reliable or how its cues should be fused. We introduce observation-matched physiological semantic grounding, where language-derived priors must be constructed from the same rPPG observation, remain bounded by an auditable prior contract, and avoid BP-label or identity leakage. CALM-BP does not treat language as new physiological evidence; instead, it verbalizes rPPG descriptors into a controlled semantic interface while rPPG remains the primary haemodynamic evidence source. FlowBP-Set pairs forehead observations, synchronized BP labels, and structured physiological prompts from 81 participants. Main BP results, direct cross-dataset evaluation, language-realization ablation, and observation-mismatch controls test whether language helps because it organizes the current physiological observation rather than because it is arbitrary auxiliary text. The FlowBP-Set dataset contains sensitive facial video and physiological recordings and is therefore not publicly available due to privacy and ethical restrictions. Data access may be considered upon reasonable request and subject to applicable ethical and institutional approval.

Highlights

CALM-BP grounds language in the same rPPG observation window while keeping rPPG as the primary evidence for non-contact BP estimation. FlowBP-Set and grounding controls test whether matched semantic guidance improves auditability and performance without leakage.

  • Dataset: FlowBP-Set pairs forehead video observations with synchronized BP references and structured physiological prompts.The resource supports subject-disjoint evaluation, direct transfer, and tests with matched or detached priors.
  • Evaluation: Grounding controls test whether gains come from matched priors rather than arbitrary text, identity leakage, target leakage, or coarse protocol shortcuts.The evaluation includes mismatched and observation-matched settings designed to make the grounding claim testable.
  • Motivation: The study targets unobtrusive BP observation because conventional cuff-based or contact measurements are intermittent, inconvenient, and difficult to use frequently.Non-contact rPPG extracts pulse-related traces from facial video as a complementary route.
  • Motivation: The approach addresses limited semantic structure in numerical rPPG pipelines for separating measured evidence from quality, reliability, waveform, and state-aware cues.This framing is intended to make the prediction path easier to audit and to test dependence on the current physiological observation.
  • Method: The physiological branch encodes measured rPPG evidence, while the semantic branch verbalizes observation-bounded descriptors before cross-modal fusion and SBP/DBP inference.Language contributes a structured fusion interface rather than a direct BP prediction pathway.
  • Method: CALM-BP uses numerical rPPG as primary BP evidence and aligns it with bounded semantic priors under an explicit prior contract.The semantic pathway organizes descriptors from the same observation without adding BP-label or identity information.

2 Related Work

Prior work motivates semantic guidance for physiological time series, while CALM-BP operationalizes it through observation-matched, bounded prompts derived from the same rPPG window. The numerical rPPG pathway remains the source of physiological evidence, and the semantic pathway supplies auditable fusion guidance.

  • Related physiological and language-guided time-series work: Prior rPPG-based BP pipelines model visual, rPPG, and physiological signals, but typically leave a window’s semantic context implicit.This limits explicit separation of measured evidence from quality, reliability, waveform, and recording-state cues.
  • Related physiological and language-guided time-series work: Language-guided time-series studies motivate pairing sequence observations with broader representation spaces, trend descriptions, or structured medical prompts.CALM-BP applies this semantic-guidance idea to a physiological setting with bounded observation-derived priors.
  • Dataset and observation processing: FlowBP-Set pairs forehead observations, synchronized BP references, and structured prompts across resting, deep-breathing, and post-exercise states.Its design supports comparisons between matched, detached, and subject-disjoint semantic priors.
  • Dataset and observation processing: The processing pipeline uses centered 2-second forehead intervals, uniformly samples 21 frames, and derives cleaned three-channel rPPG observations for both numerical and semantic branches.Detrending, normalization, motion-artifact suppression, VMD component selection, and FIR denoising reduce signal disturbances before descriptor construction.
  • Observation-matched semantic guidance: CALM-BP constructs a prompt for each rPPG window from observable amplitude, trend, stability, waveform-shape, and recording-state descriptors.The prompt is a structured interface for the same window used by the physiological branch, rather than an independent BP-evidence source.
  • Observation-matched semantic guidance: The semantic prior contract restricts priors to the current rPPG observation and excludes BP targets, participant identity, and dataset-specific shortcuts.This makes language a bounded reliability-aware fusion guide rather than a hidden label channel.

4 CALM-BP

CALM-BP combines a physiological rPPG branch with an observation-matched semantic branch, aligning bounded language priors with measured representations before blood-pressure inference.

  • Semantic guidance: CALM-BP extracts bounded physiological descriptors from each rPPG window and verbalizes them into a language prior.The prior is constructed from variation, temporal, shape, and state cues in the same observation window.
  • Physiological branch: The physiological branch encodes rPPG dynamics as the primary evidence pathway, preserving channel interactions and local temporal variation.An inverted embedding tokenizes the multivariate temporal sequence before lightweight Transformer processing.
  • Language prior encoder: A frozen Qwen3-0.6B encoder converts structured physiological prompts into semantic features designed to meet the physiological representation during fusion.The language representation is guidance from the encoded prior, not free-form blood-pressure inference.
  • Fusion and BP inference: Cross-modal fusion projects physiological and semantic representations into a comparable space, uses channel-wise similarity to retrieve guidance, and adds it residually.Softmax-normalized similarity weights aggregate semantic representations while the residual connection preserves the physiological path.
  • Fusion and BP inference: The fused representation is decoded with a multivariate Transformer and projected into predicted SBP and DBP values.Training minimizes mean squared error with L2 regularization using λ = 10^-4 in all experiments.

5 Experiments

The experiments evaluate CALM-BP in subject-disjoint FlowBP-Set settings, direct cross-dataset transfer, and controlled ablations of its semantic guidance. Results indicate that performance benefits depend on observation-matched priors rather than arbitrary language input.

  • 5.2 Main Results: FlowBP-Set evaluation covers resting, deep-breathing, and post-exercise states in an in-domain subject-disjoint setting.The dataset uses 81 participants, with a 65/16 participant-disjoint train/test split created before window sampling.
  • 5.2 Main Results: CALM-BP improves over rPPG and BP baselines while retaining rPPG as the primary physiological evidence source.
  • 5.3 Cross-Dataset Evaluation: Direct evaluation on BESTLab BP and Vital Videos tests transfer without external BP-label training or fine-tuning.External observations and semantic priors are reconstructed using the same preprocessing and descriptor-to-prompt pipeline.
  • 5.4 Ablation Studies: Removing the time-series encoder, tokenizer, or prompt encoder progressively degrades performance, whereas the full frozen-LLM configuration performs best.The ablation attributes the gain to combining dedicated physiological and language-derived prior encoders.
  • 5.4 Ablation Studies: Matched physiological priors outperform weakened, generic, random, or contradictory alternatives, showing that the prior must describe the actual observation.
  • 5.4 Ablation Studies: Paraphrases, reordered fields, synonyms, and multiple templates preserve performance, while mismatched priors degrade it across random, within-subject, and state-preserved conditions.These controls distinguish observation-level alignment from fixed wording, identity similarity, or coarse state similarity.
  • 5.5 Result Discussion: Overall, CALM-BP improves on the subject-disjoint FlowBP-Set split, transfers directly to external targets, and benefits most from matched semantic guidance.

6 Discussion

The discussion frames CALM-BP as a bounded semantic interface that organizes measured rPPG evidence rather than adding independent physiological information. Ablations and mismatch controls support observation-level correspondence, while effectiveness remains limited by signal quality and descriptor coverage.

  • 6 Discussion: CALM-BP verbalizes selected rPPG descriptors, including variation, temporal reliability, waveform regularity, and recording state, under an explicit prior contract.
  • 6 Discussion: Under matched information content, verbalized priors outperform the evaluated numerical-descriptor and categorical-embedding alternatives, without establishing universal superiority of language.
  • 6 Discussion: Observation-mismatch experiments show degradation when priors are detached from their corresponding rPPG windows, including within-subject and state-preserved mismatches.
  • 6 Discussion: CALM-BP remains dependent on reliable rPPG signals and predefined descriptors, so it cannot recover information absent or severely corrupted in the measured signal.
  • 6 Discussion: Prompt paraphrasing, field-order changes, synonym replacement, and multiple templates preserve performance close to the original prompt.

7 Conclusion

The paper presents FlowBP-Set and CALM-BP for observation-matched physiological semantic grounding in non-contact BP estimation. Its results attribute benefit to same-window semantic guidance rather than arbitrary text or label leakage.

  • 7 Conclusion: CALM-BP keeps rPPG as the primary evidence source and uses same-window semantic priors to guide fusion.
  • 7 Conclusion: FlowBP-Set results, direct external transfer, and grounding controls support matched semantic guidance over arbitrary text or label leakage.
  • 7 Conclusion: Future work should broaden population coverage, semantic perturbation tests, and downstream health-language validation.

Limitations

The paper’s scope is limited to bounded, observation-matched semantic guidance for BP estimation from rPPG windows. It is not a clinically validated continuous-waveform monitor or complete health-language system, and its sensitive data require controlled handling.

  • Scope: The study addresses one physiological target—BP estimation from rPPG windows—and does not instantiate a complete health-language system.The excluded capabilities include planning, memory update, and tool orchestration.
  • Representation Limits: Prompt semantics depend on predefined descriptor construction and may omit physiological cues not captured by the selected descriptors.
  • Evaluation Limits: The controls do not exhaustively perturb every individual semantic field or prompt surface form.
  • Clinical Scope: CALM-BP predicts SBP and DBP with interpretable cues rather than a clinically validated continuous waveform or deployment-ready monitor.Clinical use requires appropriate validation, calibration, and human oversight.
  • Ethical Considerations: Facial video, physiological traces, and BP labels can expose sensitive health information, requiring documented consent, privacy handling, access scope, and intended use.

A Additional FlowBP-Set Details

FlowBP-Set pairs synchronized forehead observations with BP references and structured physiological prompts under participant-disjoint splits. Its safeguards and mismatch controls test whether semantic priors must remain matched to the current observation.

  • FlowBP-Set contains 81 participants, approximately 145,800 synchronized SBP/DBP windows, and about 1.9 million retained forehead images.The main split uses 65 training participants and 16 held-out participants, with each participant confined to one split.
  • The camera and CNAP reference-BP system share an acquisition clock, enabling each retained BP timestamp to pair with a temporally aligned forehead observation window.Video was sampled at 30 FPS, while the 1000 Hz CNAP signal was downsampled to the video frame rate.
  • Recording states include resting, deep-breathing, and post-exercise, but are used only as bounded synchronized context rather than BP categories, identity cues, or diagnostic status.
  • Retained windows require a visible stable forehead ROI, synchronized streams, and continuous reference BP; severe motion, illumination instability, ROI drift, missing references, and boundary effects trigger exclusion.Normalization statistics are estimated from training participants and fixed for held-out evaluation.
  • Mismatch controls replace the semantic prior while preserving the original rPPG observation and BP target, testing whether benefits depend on observation-prior matching.Controls include random, within-subject, and state-preserved mismatches.

A.4 Difference Between Laboratory and In-the-Wild External Evaluation

External evaluation distinguishes controlled laboratory transfer from in-the-wild transfer. The reported gap is interpreted as sensitivity to capture conditions and physiological signal quality, while transfer occurs without target-label fine-tuning.

  • BESTLab Physio is a controlled laboratory benchmark, whereas VV is closer to an in-the-wild video benchmark with broader capture variability.
  • CALM-BP obtains lower errors on BESTLab Physio than on VV, especially for SBP estimation.
  • BESTLab Physio was evaluated from a local copy because no stable public download link was identified, so the dataset is not redistributed.The authors describe preprocessing and evaluation protocols for transparency.
  • Laboratory stability improves forehead ROI tracking, RGB/rPPG extraction, descriptor construction, and the likelihood that the matched prior describes the numerical branch’s physiological evidence.
  • In-the-wild variation in pose, lighting, motion, compression, visibility, and background can reduce rPPG quality and destabilize semantic descriptors; recording-state cues may also be unavailable.
  • Cross-dataset generalization remains sensitive to capture conditions and signal quality, despite transfer without target-label fine-tuning.The evaluation design keeps the physiological branch responsible for BP inference and restricts the semantic branch to bounded same-window descriptors.

C.1 Overview of Prior Construction

The prior-construction pipeline deterministically converts descriptors from the current rPPG window into bounded verbal fields and a fixed template. It excludes BP labels, identity, and other forbidden information, while allowing synchronized recording-state context when available.

  • Prior construction computes observation-level descriptors, quantizes them into bounded semantic fields, and populates a fixed verbalization template using allowed fields.
  • The resulting prompt is encoded with a frozen semantic encoder, while prior generation remains separate from model-side encoding for auditability.
  • The prompt generator excludes SBP, DBP, participant identity, diagnosis, demographics, split membership, file paths, session IDs, and target-derived BP categories.The text is therefore a semantic description of the current observation window rather than an independent BP measurement or hidden label channel.
  • Deterministic quantization uses thresholds fixed before evaluation and applied unchanged to held-out participants and external datasets, with training-participant normalization statistics.Each prior token is intended to remain traceable to a predefined descriptor.
  • Descriptors are mapped to signal-variation, temporal-reliability, waveform-regularity, and optionally synchronized recording-state fields.The recording-state cue is omitted when unavailable in external datasets rather than imputed.
  • Waveform regularity labels windows as “regular waveform” when NER < 0.2 and “irregular waveform” otherwise, describing reliability rather than directly inferring BP.
  • The recording-state field may encode resting, deep-breathing, or post-exercise only from synchronized logs and is audited with state-preserved mismatch controls.

C.4 Mismatch Prior Construction

Mismatch construction breaks exact observation-prior pairing while preserving the physiological observation and BP target. Variants separately test arbitrary detachment, subject-level similarity, and recording-state preservation.

  • Mismatch controls replace Pi with Pj generated from another window Xj while keeping Xi and Yi unchanged, so performance changes reflect broken matching rather than altered BP labels.
  • Within-subject mismatch preserves participant-level sampling similarity without exposing identity information, while state-preserved mismatch retains recording state and random mismatch preserves neither constraint.
Loading 2608.21744v1…