PaperlayerView exact revision on arXiv

Source-linked AI summary

PulseBound: Future-Beat State Forecasting Under an Explicit Information Boundary

Chenyang Xu, Donglin Xie, Xi Xiang, Xiaoyu Li, Yufan Lu, Jiqiun Gao, Yi Zhao, Xin-Yi Li, Guangpu Zhu, Zijian Wang, Xiwen Yang, Dezhen Wang, Lin Chen, Shenda Hong, Leilei Li

arXiv:2610.12010v1cs.AI

TL;DR

Predictive PPG learning can leave unclear whether a forecast context depends only on admissible observations, because normalization and derived views may use withheld samples. PulseBound combines an explicit stored-window boundary with future-beat prediction; it reports gains over persistence and strong results in a separate downstream comparison.

  • Problem

    Causal attention alone does not ensure that withheld samples cannot influence visible inputs through normalization, nonlocal transforms, or companion views.

    Source: §1 · p. 2

  • Method

    PulseBound constructs prefix-dependent encoder inputs under a stored-window information boundary and predicts future-beat descriptors with training-only supervision.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    PulseBound reduces nine-state transformed-space MAE relative to persistence by 28.06% on MIMIC and 22.22% on VitalDB, and achieves the best mean on 9/13 frozen linear-probe and 7/13 full-fine-tuning tasks.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The results frame information access, supervised future structure, and representation transfer as separately testable aspects of predictive physiological learning.

    Source: §6 · p. 10

    Forecasting tests hold out groups within MIMIC and VitalDB rather than entire sources, and deterministic targets remain signal proxies rather than clinical ground truth.

    Source: §5 · p. 9

Abstract

from arXiv · show

Predictive representation learning from photoplethysmography (PPG) can violate causal information access even with causal attention, as normalization, nonlocal transforms, or companion views may depend on withheld samples. We introduce PulseBound, a PPG representation learner combining physiologically structured future-beat prediction with an explicit stored-window information boundary. A content-independent cutoff separates the visible prefix from the prediction target. Prefix-only normalization, suffix replacement before derived-view construction, and aligned masking ensure that encoder inputs depend only on the visible prefix and cutoff. This yields stored-suffix invariance: with fixed model state, randomness, prefix, and cutoff, changing the stored suffix cannot change the forecast context. A shared horizon-conditioned head predicts nine rhythm and morphology descriptors for up to four extractor-valid future beats, using elementwise validity masks; optional ECG-derived pulse-arrival-time supervision is restricted to training. On MIMIC and VitalDB groups held out from PulseBound backbone pretraining, PulseBound reduces nine-state transformed-space MAE relative to last-visible-beat persistence by 28.06% and 22.22%, respectively, with gains in MAE, MAE-Skill, and Spearman correlation across all 40 source-cutoff-horizon cells. In a separate comparison of seven models on 13 downstream tasks, PulseBound achieves the best mean on nine frozen linear-probe and seven full-fine-tuning tasks. Stored-suffix interventions cause zero recorded changes in forecast contexts or predictions, with zero suffix-input gradients at audited precision under the stored-window interface. These findings separate three testable aspects of predictive physiological representation learning: information access, supervised future structure, and transfer.

1 INTRODUCTION

PPG’s redundant pulse cycles and sensitivity to acquisition conditions mean accurate reconstruction may not reveal whether a representation captures physiological change. PulseBound pairs future-beat prediction with an explicit boundary on which stored observations can shape the forecast context.

1 INTRODUCTION

  • PPG’s redundant cycles and sensitivity to motion, contact, illumination, and device effects mean accurate missing-sample reconstruction need not capture evolving physiological properties.PPG is widely available across bedside monitors, pulse oximeters, and wearables, motivating reusable representation learning.Source: §1 · p. 1
  • Causal attention alone does not prevent withheld samples from influencing visible inputs through full-window normalization, global transforms, companion views, or target-dependent cutoff selection.The relevant boundary covers the complete computation constructing the forecast context, not just its attention mask.Source: §1 · p. 2
  • PulseBound uses a content-independent cutoff, prefix-only normalization, suffix replacement before derived views, and aligned masking so encoder inputs depend only on the stored prefix and cutoff.With model state, randomness, prefix, and cutoff fixed, replacing the stored suffix cannot change the forecast context; the guarantee begins at the stored-window interface.Source: §1 · p. 2
  • A shared horizon-conditioned head predicts nine PPG-derived descriptors for up to four extractor-valid future beats, with component-wise validity masks and optional training-only ECG-derived PAT supervision.The targets are inspectable signal-derived proxies, not clinical ground truth; ECG does not enter the encoder or deployment path.Source: §1 · p. 2
  • The paper separately evaluates future-state prediction, audits the information boundary by intervention and input gradients, and compares the complete pretrained system across 13 downstream tasks.Checkpoint-compatible frontends are retained for released baselines, so downstream comparisons are system-level rather than causal attribution to one pretraining objective.Source: §1 · p. 2

2 RELATED WORK

Prior PPG representation work spans morphology, acquisition diversity, predictive and generative objectives, blood-pressure transfer, subject structure, signal quality, and cross-modal supervision. PulseBound’s distinction is coupling such objectives to an explicit contract governing which stored observations can influence prediction.

2 RELATED WORK

  • PPG representation learning has explored participant structure, morphology, acquisition diversity, predictive and generative supervision, and broader masked and contrastive approaches.The cited examples include PaPaGei, Pulse-PPG, DNA-PPG, GPT-PPG, SIGMA-PPG, TS2TC, and RhythmJEPA.Source: §2 · pp. 2–3
  • Prior work studies BP transfer, inter-subject structure, signal-quality assessment, and cross-modal physiological supervision, but these works address distinct representation properties and evaluation settings.The paper avoids numerical rankings across unmatched BP-specific protocols and does not claim its quality evaluation establishes device invariance.Source: §2 · p. 3
  • ECG can supervise a PPG representation without increasing deployment-time sensing requirements; in PulseBound, ECG and ECG-derived labels stay out of the encoder and forecast context.Deployment remains PPG-only.Source: §2 · p. 3
  • PulseBound’s distinction is coupling existing objectives and scaffolds to an explicit one-sided information-access contract, with fixed descriptor targets and an auditable conditioning path.Its downstream benchmark characterizes system-level transfer using checkpoint-compatible frontends rather than attributing gains causally to any single objective.Source: §2 · p. 3

3 METHOD

PulseBound separates the prefix-dependent forecast context from training-only targets and uses that boundary to learn future-beat representations. Its method combines boundary-preserving signal construction, validity-aware beat descriptors, a shared multi-horizon predictor, joint objectives, and a separate PPG-only transfer readout.

3 METHOD

  • PulseBound separates prefix-dependent encoding from training-only supervision, so target signals do not enter the forecast context.The separation preserves PPG-only inference.Source: §3 · p. 3, Figure 1 · p. 4

3.1 FORECASTING TASK AND INFORMATION-BOUNDARY CONTRACT

  • A waveform-independent cutoff defines the visible prefix and hidden suffix; labels and validity masks may depend on the suffix but cannot enter the encoder or forecast context.The forward-access contract begins at the stored-window interface and does not guarantee upstream raw-signal preprocessing or statistical independence between past and future physiology.Source: §3.1 · p. 4

3.2 BOUNDARY-PRESERVING SIGNAL AND TOKEN CONSTRUCTION

  • Prefix-only normalization and suffix replacement before nonlocal derived views, followed by aligned masking, keep encoder inputs suffix-independent.The derived views are prefix-global and therefore do not imply sample-wise streaming causality.Source: §3.2 · pp. 4–5

3.3 TEMPORAL CONNECTIVITY AND GLOBAL READOUTS

  • The backbone routes temporal queries to earlier temporal tokens and themselves, while registers aggregate the masked sequence without feeding information back into temporal states.This connectivity controls mixing after prefix-dependent inputs are constructed; the stored-window suffix-invariance guarantee instead follows from input construction.Source: §3.3 · p. 5

3.4 DETERMINISTIC FUTURE-BEAT SUPERVISION

  • A training-only extractor identifies up to four future beats and supplies nine PPG-derived descriptors plus optional PAT, with elementwise validity masks excluding inadmissible measurements.Horizons index accepted events rather than fixed elapsed times; neither ECG nor extracted targets enters the encoder or forecast head.Source: §3.4 · p. 5

3.5 HORIZON-CONDITIONED FORECAST AND CURRICULUM

  • A shared predictor conditions each forecast on the visible-prefix representation and future-event embedding, with a cutoff-fraction input and shared output head across horizons.Training progressively activates horizons H1–H4 and averages masked smooth-L1 loss over valid active state elements.Source: §3.5 · pp. 5–6

3.6 JOINT OPTIMIZATION AND PPG-ONLY INFERENCE

  • Future-state prediction is jointly optimized with token prediction, waveform reconstruction, available scaffold losses, subject contrastive supervision, and a ponder regularizer.Downstream inference encodes a complete 30-second PPG window without masking, a tokenizer, or auxiliary modalities; its 768-dimensional readout combines global and temporal features.Source: §3.6 · p. 6

4 EXPERIMENTS

Experiments assess transfer and future-beat forecasting under the stored-prefix contract; PulseBound improves forecasting against persistence and leads on most downstream tasks.

4.1 EVALUATION PROTOCOL

  • Transfer uses matched examples, labels, outer splits, and task metrics across 13 tasks; frozen linear probing is primary and fine-tuning is secondary.Epoch 29 was selected by validation-only linear probing; forecasting and the stored-suffix audit use terminal epoch 79, and MIMIC group-disjointness applies to backbone pretraining, not the full-corpus tokenizer.Source: §4.1 · p. 6

4.2 DOWNSTREAM TRANSFER AND ABLATION ANALYSIS

  • PulseBound has the best mean on 9/13 frozen linear-probe tasks and 7/13 full-fine-tuning tasks, with other systems leading several tasks.Table 1 compares seven systems across 13 tasks; PulseBound’s linear-probe leads span PPG-BP, WESAD stress and affect, DaLiA activity, Stanford quality, and BIDMC HR.Source: Table 1 · p. 7, §4.2 · p. 7
  • At seed 0, full PulseBound exceeds the capacity-matched no-state predictor by 0.0399 Skill, while the shuffled-target predictor scores 0.0032 Skill.The shuffled-target result is near the training-median reference; the controls support aligned state supervision’s importance without showing that no-state representations lack recoverable future information.Source: Table 2 · p. 7, §4.2 · pp. 7–8

4.3 FUTURE-BEAT STATE FORECASTING

  • 28.06% and 22.22%: PulseBound reduces nine-state transformed-space MAE versus last-visible-beat persistence on MIMIC and VitalDB, respectively.MAE, MAE-Skill, and Spearman correlation favor PulseBound across all 40 source–cutoff–horizon cells; those cells are not independent replications.Source: Table 3 · p. 8, §4.3 · p. 8
  • Across three seeds, no-state ridge and MLP probes achieve equal-source Skills of 0.2844 ± 0.0115 and 0.3488 ± 0.0027, while shuffled-target direct prediction reaches 0.0038 ± 0.0006.These available-case controls do not share Table 3’s persistence-paired support, so their absolute Skill values are not directly comparable to its primary result.Source: §4.3 · p. 9
  • Across the five fixed cutoffs, PulseBound outperforms persistence; absolute H4 MAE remains below persistence on both sources despite larger relative degradation across H1–H4.PAT is a conditional diagnostic because its persistence reference uses an observed ECG-derived quantity and its physical-unit results vary by source.Source: §4.3 · p. 9
  • Across 10,000 audited windows per source, five suffix interventions caused zero recorded changes in forecast contexts or predictions, and suffix-input gradients were zero.The audit supports the stored-window boundary, not upstream preprocessing; target-quality assessment is separate, and extractor coverage is not label correctness.Source: §4.3 · p. 9

5 DISCUSSION AND LIMITATIONS

PulseBound’s guarantees and results have specific boundaries: the access contract starts at stored windows, targets are signal proxies, and transfer varies by task.

5 DISCUSSION AND LIMITATIONS

  • The information-access guarantee starts at the stored, preprocessed-window interface and does not cover upstream filtering or resampling or imply raw-signal streaming equivalence.Boundary audits test information access; forecasting accuracy separately tests whether the admissible prefix contains useful information about future beat states.Source: §5 · p. 9
  • Future-state supervision is an alignment objective, not evidence that other pretraining objectives cannot encode predictive information; downstream transfer is task-dependent.No-state representations retain positive available-case Skill, and downstream transfer uses a different checkpoint and readout from intrinsic forecasting.Source: §5 · p. 9
  • Targets are measurement-dependent signal proxies, not clinical ground truth; extractor validity does not eliminate systematic detection or morphology error.Forecasting tests hold out groups within MIMIC and VitalDB rather than entire sources; the frozen Event-VQ tokenizer predates the MIMIC patient split, and the backbone remains large for resource-constrained deployment.Source: §5 · p. 9

6 CONCLUSION

PulseBound combines interpretable future-beat prediction with a stored-prefix information boundary. It improves forecasting and performs strongly across downstream tasks, while framing access, supervision, and transfer as distinct questions.

6 CONCLUSION

  • PulseBound reduces nine-state transformed-space MAE versus persistence by 28.06% on MIMIC and 22.22% on VitalDB, and has the best mean on 9/13 frozen linear-probe and 7/13 fine-tuning tasks.Boundary interventions and input-gradient audits support the stored-window contract.Source: §6 · p. 9
  • Three-seed controls indicate that state supervision aligns representations with forecastable physiology rather than uniquely creating predictive information.The paper frames information access, supervised future structure, and transfer as separately testable aspects of predictive physiological representation learning.Source: §6 · p. 10
  • Test labels were not used for checkpoint selection, tuning, or ablation choice; transfer uses epoch-29 EMA checkpoints, while forecasting and audits use epoch 79.Raw-record reconstruction reproduced three sampled VitalDB windows but shared the SDK temporal grid and was not an exhaustive native-timestamp or device-synchronization audit.Source: REPRODUCIBILITY STATEMENT · p. 11

A EXTENDED DISCUSSION AND LIMITATIONS

PulseBound’s stored-window guarantee is distinct from architectural causality and raw-signal streaming equivalence. The discussion separates forecasting from transfer and cautions that descriptor skill does not establish clinical accuracy or broad deployment readiness.

Information access versus architectural connectivity

  • The guarantee begins at the stored, preprocessed window: offline filtering, resampling, or full-window normalization can transmit raw-suffix information into stored prefixes.A raw-signal guarantee requires boundary-preserving preprocessing and raw-to-context interventions; stored-suffix audits support only the implementation-level contract.Source: A EXTENDED DISCUSSION AND LIMITATIONS · p. 14

Prediction and transfer are distinct outcomes

  • Forecasting and transfer need not favor the same configuration; explicit state supervision aligns evaluated future descriptors but does not show that other objectives cannot encode predictive information.Transfer uses an earlier checkpoint and a different readout, so intrinsic Skill is complementary to, not a sufficient explanation of, downstream utility.Source: A EXTENDED DISCUSSION AND LIMITATIONS · p. 14

Inspectable targets remain measurement-dependent

  • Descriptor validity masking excludes inadmissible targets but does not correct systematic detection errors; coverage is not target accuracy.Horizons index accepted beats rather than fixed durations, and lower transformed PAT MAE can coexist with worse millisecond MAE and rank correlation on MIMIC.Source: A EXTENDED DISCUSSION AND LIMITATIONS · p. 14

Generalization and deployment scope

  • Forecasting tests hold out groups within two pretraining sources, not entire sources; device and demographic shifts, calibration, energy use, and prospective clinical validation need separate assessment.PPG-only inference describes sensing requirements, not computational efficiency; the downstream backbone has 97.0M parameters and executes all 12 layers without early termination.Source: A EXTENDED DISCUSSION AND LIMITATIONS · p. 14

B EXACT FUTURE-STATE TARGETS

Future-state targets are deterministic PPG descriptors extracted from filtered, normalized windows, with invalid elements excluded from the objective. PAT is derived from synchronized ECG when available, independently of PPG-target validity.

B EXACT FUTURE-STATE TARGETS

  • The target extractor zero-fills nonfinite samples, uses visible-prefix statistics, clips normalized PPG to [−8, 8], applies a 0.5–8 Hz FFT band-pass, and detects refractory local maxima.At 50 Hz, the configured 0.30-s cutoff guard and 0.35-s onset lookback yield an effective 18-sample (0.36-s) future-peak guard; a 0.55-s post-peak window is reserved.Source: B EXACT FUTURE-STATE TARGETS · p. 14
  • PAT pairs each PPG onset with the most recent ECG R-peak 0.05–0.65 s earlier, while PPG-derived target validity remains independent of PAT validity.A per-window ECG-availability flag prevents placeholder ECG from becoming a target.Source: B EXACT FUTURE-STATE TARGETS · p. 14
  • Table 4 specifies deterministic state transforms, with clip truncating values to a range and U and D denoting pulse areas after and before the peak.Onset is the minimum filtered value in the preceding 0.35 s; invalid elements are zero-filled and excluded from Equation 13.Source: B EXACT FUTURE-STATE TARGETS · p. 14, Table 4 · p. 15

C ARCHITECTURE AND OBJECTIVE DETAILS

The appendix specifies PulseBound’s attention biases, joint training objectives, frozen tokenizer, and evaluation protocols. It also documents checkpoint selection, split handling, and a closed-set interpretation for Real-World PPG identification.

C ARCHITECTURE AND OBJECTIVE DETAILS

  • Each attention head uses learned dense type-pair and relative-time biases, while the causal mask—not the learned biases—determines permitted connections.The no-graph-bias ablation removes additive bias terms but leaves the causal mask unchanged.Source: C ARCHITECTURE AND OBJECTIVE DETAILS · p. 15

C.1 FORECAST CURRICULUM AND EXACT JOINT OBJECTIVE

  • Training progressively activates horizons H1–H4 and averages a masked smooth-L1 state loss over valid active elements, alongside token, waveform, scaffold, subject, and ponder objectives.No teacher is used, and all 12 layers execute during training and inference; the ponder regularizer does not enable early exit.Source: §C.1 · pp. 15–16

C.2 FROZEN EVENT-VQ TOKENIZER

  • PulseBound’s frozen Event-VQ tokenizer was trained on the full processed MIMIC corpus before the later patient-level split, so group-disjoint claims apply to backbone pretraining, not tokenizer training.The tokenizer is trained separately and frozen; its pretraining included PPG from 628 groups later assigned to MIMIC validation and test partitions.Source: §C.2 · pp. 16–17

D.1 CANONICAL FULL-MODEL PROTOCOL

  • Downstream LP/FT uses the validation-selected zero-based epoch-29 EMA checkpoint; future-state evaluation and stored-suffix audits use the terminal epoch-79 EMA checkpoint from the same three runs.Checkpoint selection, hyperparameter tuning, and ablation choice do not use test labels; the cutoff is sampled independently of waveform, metadata, beats, and targets.Source: §D.1 · p. 18

D EVALUATION PROTOCOL AND DOWNSTREAM ANALYSES

  • The downstream protocol uses the same evaluation examples, labels, splits, and task metrics across methods, while retaining baseline-specific preprocessing only for checkpoint compatibility.Real-World PPG is evaluated only as closed-set identification because subject identity is the class label and a subject-held-out split would change the class space.Source: §D.2 · p. 18

D.2 DOWNSTREAM TASKS AND SPLITS

  • The 13-task downstream evaluation shares examples, labels, outer splits, and primary metrics across methods, with LP designated primary and FT secondary.BIDMC HR/RR/SpO2 are excluded from the primary clean cross-model aggregate when a baseline cannot rule out pretraining overlap, though task-level values remain reported.Source: §D.2 · p. 18, Table 11 · p. 19

D.3 COMPLETE FORECASTING AND DOWNSTREAM ABLATION ANALYSIS

The ablations show a forecasting advantage for aligned state supervision, but downstream effects vary by task and intervention. The section also reports cross-model task-level outcomes and clarifies how checkpoints, variability, target audits, and corpus provenance constrain interpretation.

D.3 COMPLETE FORECASTING AND DOWNSTREAM ABLATION ANALYSIS

  • Across three pretraining seeds, full PulseBound improves some downstream means over no-state pretraining but is worse on activity macro-F1; differences are descriptive, not claims of statistical significance.Stress AUROC is nearly identical, and comparisons with shuffled-target pretraining are also mixed across tasks.Source: §D.3 · p. 19
  • 0.3926 macro Skill for full PulseBound versus 0.3527 for the capacity-matched no-state MLP and 0.0032 for the shuffled-target control at seed 0.This is a seed-0 comparison; capacity matching does not make it a shared-probe comparison of representation quality.Source: §D.3 · p. 19, Table 12 · p. 20
  • PAT, curriculum, physiological views, and source composition produce task-dependent results; full PulseBound exceeds MIMIC-only source Skill, while MIMIC-only has lower SBP MAE and higher activity macro-F1.Removing PAT gives seed-0 Skill 0.3940 versus 0.3926 for full PulseBound, without positive aggregate forecasting evidence for PAT supervision.Source: §D.3 · pp. 19–20
  • No-state ridge and MLP probes retain positive future-state Skill, while the MLP outperforms ridge; the shuffled-target predictor is near zero only after source averaging.The shuffled-target predictor remains positive on MIMIC and negative on VitalDB, and these results do not establish a representation information limit.Source: §D.3 · pp. 20–21, Table 13 · p. 21

D.4 STATISTICAL AGGREGATION

  • Aggregation averages downstream seeds within each pretraining seed before computing descriptive mean and sample SD; these SDs are not confidence intervals or independent-run estimates.Seed-0 ablations fix the pretrained checkpoint and summarize downstream seeds only.Source: §D.4 · p. 21

D.5 CROSS-MODEL DOWNSTREAM AGGREGATE

  • PulseBound has the best mean in 16 of 26 non-random-initialization LP/FT task–protocol cells, with a 1.45 mean rank across 20 non-BIDMC cells.These counts summarize task-level outcomes, not independent replications.Source: §D.5 · p. 21

D.6 SELECTED QUANTITATIVE SUMMARY

  • The cross-task averages are descriptive: they combine heterogeneous task metrics or regression units without scale normalization, and pooled paired effects are unreported.Automatic target audits use an ECG-derived reference rather than clinical ground truth; inference measurements are architecture-only encoder measurements.Source: §D.6 · p. 23, Table 17 · p. 23

E.1 MIMIC PREPROCESSING AND CORPUS FINGERPRINTS

  • The final runs draw 5,160,960 windows per epoch, split 80:20 between MIMIC and VitalDB within fixed patient-level training partitions.MIMIC preprocessing resamples to 50 Hz and forms 30-s windows with 1-s patches.Source: §E.1 · p. 24

E.2 VITALDB MULTIMODAL CONVERSION, PROVENANCE, AND AUDIT

The section documents how multimodal VitalDB windows were converted and audited, while distinguishing completed checks from provenance and reconstruction work still outstanding. It also reports training configuration and the scope of later stability and forecasting tables.

E.2 VITALDB MULTIMODAL CONVERSION, PROVENANCE, AND AUDIT

  • 6,388 conversion records were discovered; 6,153 completed and 235 failed, yielding 114 shards and 1,149,347 stored 30-s windows.The counts describe conversion records, not independently verified patients.Source: §E.2 · p. 24
  • The converter samples selected modalities on a shared 125-Hz grid, while stored waveforms contain 1,500 samples per 30-s window at 50 Hz.The stored start index refers to the 125-Hz source grid; invalid optional modalities remain PPG windows with availability flags unset.Source: §E.2 · p. 24
  • Finite scaffold targets do not uniformly indicate cross-modal supervision because HR and RMSSD can fall back to PPG peaks when ECG is unavailable.Rhythm does not use that fallback; SBP, DBP, and MAP require valid arterial pressure.Source: §E.2 · p. 25
  • Three source-grid-start windows were reconstructed with zero maximum absolute difference over finite elements, but the check does not cover nonzero offsets or all track and modality patterns.Scalar metric extractors were not independently reimplemented, and the waveform processing used the same SDK temporal grid and basic SciPy operators.Source: §E.2 · p. 26
  • The recovered manifest supports row-level traceability but is not a standalone one-row-per-window record of selected raw tracks and scaffold provenance.A converter-independent audit protocol is specified, including a complete row manifest, stratified sampling, reconstruction, and comparison of stored outputs.Source: §E.2 · p. 26

E.3 FINAL TRAINING AND SOFTWARE CONFIGURATION

  • Each 80-epoch pretraining run used 16 A100 GPUs and global batch 4096; the recorded seed-0 run totaled 260.616 GPU-hours.The cost excludes dataloader initialization and checkpoint I/O; peak training memory was unavailable.Source: §E.3 · p. 27

E.4 NUMERICAL-STABILITY CHECKS

  • The numerical-stability check covers 40,960 window–cutoff cases per source and measures objective stability, not trained-model accuracy.The state-wise forecasting table averages over five legal cutoffs and conditions PAT on ECG availability and valid extraction.Source: Table 21 · p. 27, §F.1 · p. 27

F.2 HORIZON, CUTOFF, AND ECG-CONDITIONAL PAT DIAGNOSTICS

Forecast diagnostics show lower error through four accepted beats and across fixed cutoffs, while horizon degradation and PAT performance retain important qualifications. Separate audits test stored-window suffix invariance and the quality and availability of extracted targets; low-label transfer varies by task.

F.2 HORIZON, CUTOFF, AND ECG-CONDITIONAL PAT DIAGNOSTICS

  • PulseBound retains lower absolute H4 MAE on both sources, but its H1–H4 relative error increase exceeds persistence on MIMIC and VitalDB.Its Skill falls from 0.3911 ±0.0011 to 0.3291 ± 0.0004 on MIMIC and from 0.4672 ± 0.0009 to 0.4207 ± 0.0004 on VitalDB; lower error does not imply slower horizon degradation.Source: §F.2 · p. 27, Table 23 · p. 28
  • PulseBound outperforms persistence across all five fixed cutoffs, although moving the cutoff also changes target events and potentially valid support.These comparisons assess sensitivity across forecasting conditions rather than isolating the causal effect of additional context.Source: §F.2 · p. 27, Table 24 · p. 28
  • PAT transformed-space error and Skill improve in all 40 source–cutoff–horizon cells, but correlation improves in only 20 and MIMIC physical MAE and ρ worsen.On MIMIC, physical MAE is 109.14 versus 107.88 ms and ρ is 0.3379 versus 0.4013; PulseBound receives PPG alone while PAT persistence conditions on observed ECG-derived information.Source: §F.2 · p. 29, Table 25 · p. 29

F.3 STORED-SUFFIX INTERVENTION AUDIT

  • Five stored-suffix interventions leave audited visible states, registers, forecast contexts, and predictions unchanged at recorded precision.The audit used 10,000 windows per source and found zero recorded intervention deviations; its finite-sample checks do not prove invariance for all perturbations or the full Jacobian.Source: §F.3 · pp. 29–30, Table 26 · p. 30

F.4 TARGET-QUALITY AUDIT

  • PPG-derived heart-rate MAE against the stored reference is 4.57 bpm on MIMIC and 5.27 bpm on VitalDB; interval MAE is 82.65 and 73.48 ms, respectively.Reference ECG peaks are automatically detected rather than manually annotated, and interval comparisons may span missed detections.Source: §F.4 · p. 30, Table 27 · p. 31
  • Mean PPG-state availability rises from 98.60% to 99.98% across MIMIC quality quartiles and from 97.17% to 100.00% on VitalDB, but extractor acceptance does not establish target correctness.The VitalDB Q4 value is rounded; the automated audit does not independently validate every target dimension or establish clinical accuracy.Source: §F.4 · pp. 30–31, Table 28 · p. 31

G.1 LABEL EFFICIENCY

  • At 25% nominal labels, PulseBound beats the no-state control on BIDMC HR MAE (2.4444 versus 4.3286) and WESAD stress AUROC (0.9541 versus 0.9403), but loses on PPG-BP SBP MAE and Stanford quality macro-AUROC.The no-state control scores 15.8941 versus 16.6867 on SBP MAE and 0.9210 versus 0.9075 on macro-AUROC; a trainable random backbone exceeds frozen PulseBound on DaLiA at 10% and 25%.Source: §G.1 · pp. 31, 33

G.2 EXPLORATORY STRUCTURAL ABLATIONS

The structural ablation screen is exploratory, the source-contribution comparison is compute-matched, and the efficiency measurements describe encoder architectures rather than full trained systems.

G.2 EXPLORATORY STRUCTURAL ABLATIONS

  • The no-register and no-graph-bias variants appear in a 5%-data, single-seed development screen, not the confirmatory comparisons.The reported values are comparable only within that screen.Source: §G.2 · p. 33, Table 35 · p. 33

G.3 COMPUTE-MATCHED SOURCE CONTRIBUTION

  • VitalDB contribution is assessed by comparing full PulseBound with a MIMIC-only control under matched windows, updates, capacity, schedule, tokenizer, seeds, and downstream hyperparameters.The corresponding forecasting and transfer contrasts are reported in Table 12, but their values are not supplied here.Source: §G.3 · p. 33

G.4 EFFICIENCY

  • Removing registers gives the highest measured batch-64 throughput and lowest batch-64 latency, but the ordering differs at batch size 1.These architecture-only measurements lack repeated-run uncertainty and do not establish a favorable accuracy–efficiency trade-off for the ablations.Source: §G.4 · p. 33, Table 36 · p. 34

Related papers

Loading 2610.12010v1…