RECAST: Recent & Context-Aware Sampling for Test-Time Adaptation in Streaming Biosignals

Yong-Yeon Jo; Junho Song; Joon-myoung Kwon

Source-linked AI summary

RECAST: Recent & Context-Aware Sampling for Test-Time Adaptation in Streaming Biosignals

Yong-Yeon Jo, Junho Song, Joon-myoung Kwon

arXiv:2608.28271v1Aug 28, 2026cs.AI

TL;DR

Streaming biosignals drift across patients and over time, leaving buffered TTA without a clear rule for which samples should drive each update. RECAST selects samples using recency, contextual similarity, and predictive reliability while leaving the model and training objective unchanged. Across two blood-pressure datasets, it improves accuracy and tracking with statistically significant per-patient gains and sub-second deployment latency.

Problem
Streaming biosignal drift makes population-trained models vulnerable to miscalibration, while buffered TTA leaves the sample-selection problem under-specified.
Method
RECAST is a lightweight plug-in that combines temporal recency, contextual similarity, and predictive reliability to select adaptation samples without architectural changes.
Results
RECAST is the most accurate compared method on PulseDB and MC-MED, attaining the lowest MAE and RMSE on all three targets, with statistically significant improvements on both datasets.
Takeaways & Limitations
Sample selection is a decisive buffered-TTA component, and making it explicit yields consistent gains without changing the model or training objective.
Takeaways & Limitations
RECAST’s recency measure may not match elapsed time under irregular supervision, its benefit depends on within-patient variation, and it is instantiated only within TTC.

Abstract

from arXiv · show

Streaming biosignals vary across subjects and drift over time, so population-trained models lose accuracy during long-term monitoring. Test-time adaptation (TTA) enables online personalization by updating the model on incoming samples. But in a stream, a basic question is left open: \emph{which samples should drive each update?} Using all buffered samples blurs the update with irrelevant segments. Using only the latest segment makes the update noisy and unstable. The most useful samples are recent, aligned with the current physiological state, and reliable enough to learn from. We propose \textbf{RECAST} (REcent \& Context-Aware Sampling for TTA), a lightweight sampling module for buffered TTA frameworks. RECAST builds each adaptation batch from three signals: temporal recency, contextual similarity, and predictive reliability. It changes only which samples are used, leaving the model and the training objective unchanged. On two blood-pressure datasets, RECAST improves estimation accuracy and trend tracking over baselines and ablations. The per-patient gains are statistically significant on both datasets, with broad improvement on the regular benchmark and gains concentrated on the hardest patients in the emergency-department setting. RECAST stays practical, adding only sub-second latency per segment on a single GPU and CPU core.

1 Introduction

Streaming biosignals drift within patients and can change abruptly, making sample selection a central but under-specified part of buffered test-time adaptation. RECAST addresses this by selecting recent, contextually similar, and reliable samples without changing the host model or objective.

Population-trained models can become miscalibrated as patient characteristics and operating conditions shift during long-term monitoring.
Gradual drift makes older samples less representative, while abrupt transitions can mix recent segments from different or unreliable physiological regimes.
RECAST combines age-decayed recency, latent-space contextual similarity, and uncertainty-based predictive reliability to construct adaptation batches.
RECAST merges recency and similarity into a ranking, filters uncertain candidates, and preserves a fixed labeled–unlabeled batch mix while leaving the model and TTC objective unchanged.
The evaluation covers accuracy, tracking, sampling behavior, sensitivity, per-patient improvement, and computational cost across two real-world datasets.
On PulseDB and MC-MED, RECAST produces statistically significant per-patient improvements, with broader gains on PulseDB and concentrated gains for harder, more variable MC-MED patients.

2 Related Work

Streaming biosignal monitoring combines dense waveform data with sparse reference measurements, requiring models to adapt online as conditions change. Existing test-time adaptation methods vary in update strategy, while TTC extends adaptation to personalized biosignal regression with periodic supervised calibration.

Continuous waveforms arrive over long monitoring periods, while cuff-based blood-pressure references are available only intermittently.
Figure 1 depicts an ICU patient monitored for 24 hours, with continuously recorded waveforms and sparse cuff-BP checks.
Test-time adaptation updates models online during inference without revisiting training data, making it suitable for non-stationary real-time deployment.
Existing methods adapt normalization or affine parameters, filter unreliable samples, or use self-supervised auxiliary tasks such as masked autoencoding.
Most surveyed methods target classification and operate fully unsupervised, without using labels that may arrive during deployment.
TTC extends TTA to personalized streaming-biosignal regression by combining self-supervised adaptation with periodic supervised calibration in a dual-buffer structure.

3 Challenges in Adapting to Streaming Biosignals

Within-patient biosignals exhibit both gradual distribution drift and abrupt transitions, so neither uniform history nor latest-only updates reliably represent the current physiological state. These opposing dynamics motivate selective adaptation using recency together with contextual consistency.

A PulseDB patient’s blood pressure shows slow moving-average drift and sharp transient spikes within the same recording.
Non-stationarity changes basic statistics gradually, making older samples increasingly outdated and motivating emphasis on recent observations.
Moving-average trends and moving-standard-deviation regions visualize blood-pressure changes over time in PulseDB.
Temporal decay reduces a sample’s contribution as elapsed time grows, helping updates remain aligned with the current distribution.
Abrupt activity changes or acute clinical events can rapidly alter signal characteristics, making the latest samples risky when they are inconsistent with the current state.
Gradual drift favors recency, whereas abrupt transitions can make recent samples misleading; all-history and latest-only buffering therefore represent opposing simplifications.

4 Proposed Method: RECAST

RECAST selects adaptation samples using temporal recency, contextual similarity, and predictive reliability, addressing both gradual drift and abrupt physiological transitions. It integrates this selective sampling into TTC while leaving the model and training objectives unchanged.

Joint temporal-similarity scoring: RECAST ranks buffered samples by combining temporal recency and contextual similarity to identify candidates aligned with the current physiological state.Recency addresses gradual drift, while similarity can retain older samples that match the state after an abrupt transition.
Age-penalized uncertainty filtering: Predictive reliability filters highly uncertain candidates because similar latent representations can still correspond to different physiological states.Epistemic uncertainty is estimated from predictive variance across Monte Carlo dropout passes.
Age-penalized uncertainty filtering: Age-penalized uncertainty imposes stricter criteria on older samples, reducing adaptation to uncertain and outdated patterns.The penalty is intended to keep streaming updates stable as the data distribution drifts over time.
Adaptation batching: Candidates enter the adaptation batch only when they pass both similarity and uncertainty thresholds.This joint filter keeps samples that are both relevant and reliable rather than selecting on either criterion alone.
Adaptation batching: RECAST maintains a fixed labeled-to-total ratio, selects ranked valid samples from labeled and unlabeled queues, and allows batches smaller than B when insufficient candidates remain.The default labeled ratio is r = 0.25, and the final batch is formed from the selected labeled and unlabeled subsets.
Integration with TTC: RECAST changes only sample selection within TTC; the backbone and prediction objectives used for updates remain otherwise unchanged.The method is implemented as a lightweight sampling module on top of TTC's shared-backbone, task-head architecture.

5.1 Experimental Setup

The experiments evaluate streaming blood-pressure adaptation on PulseDB and MC-MED, which differ in segment structure, label density, and within-patient variability. The setup uses sequentially processed inputs, a shared TTC model, buffered test-time updates, and comparisons against established baselines.

Datasets: PulseDB provides 10-second fully labeled segments, but adaptation streams use one BP label per ten segments to simulate sparse supervision.The benchmark combines MIMIC-III and VitalDB recordings under a patient-disjoint split.
Datasets: MC-MED provides 60-second segments arriving roughly once per minute, with BP references aligned to only about 6% of segments.It represents emergency-department stays with heterogeneous visit lengths and intermittent clinical measurements.
Dataset differences: MC-MED has lower within-patient BP variability than PulseDB, with median within-patient SBP standard deviations of 7.6 and 12.7 mmHg, respectively.The lower within-stay variation leaves less change for adaptation to exploit on MC-MED.
Input and preprocessing: Streaming inputs are processed sequentially after resampling to 125 Hz, bandpass filtering from 1–60 Hz, and sample-wise Z-score normalization.PulseDB uses 10-second inputs and MC-MED uses 60-second inputs.
Pre-training: The shared backbone combines a 1D-CNN frontend with a Transformer encoder pretrained by masked autoencoding at a 0.75 masking ratio.Pretraining runs for 50 epochs with Adam at learning rate 10^-4 and batch size 512.
Test-time phase: During test-time adaptation, dual queues have capacity 64, batches contain B = 32 samples with labeled ratio r = 0.25, and the backbone is updated for five iterations while the prediction head remains frozen.The frozen head preserves the population-trained latent-to-BP mapping under sparse supervision.

5.2 Experimental Results

RECAST improves accuracy and trend tracking across PulseDB and MC-MED, with benefits varying by patient variability and baseline difficulty. Its selective sampling also remains computationally practical, while ablations show that combining recency, similarity, and reliability matters.

Performance Comparison: 5.4% lower Total MAE than TTC on PulseDB (9.51 vs. 10.05), while MC-MED improves by 3.0% (19.44 vs. 20.04).RECAST achieves the lowest MAE and RMSE across all three targets on both datasets.
Performance Comparison: SBP correlation rises from 0.42 to 0.90 on PulseDB and from 0.46 to 0.79 on MC-MED with adaptation, with RECAST highest on both.RECAST also has the lowest RMSE on every target; only two standard-deviation metrics differ within rounding.
Performance by Within-Patient Variability: RECAST’s gain increases once within-patient variability exceeds the most stable quintile, reaching about +4% median improvement across Q2–Q4.The gain plateaus and eases slightly in Q5, where abrupt transitions leave fewer context-aligned samples.
Per-Patient Improvement Distribution: 74.6% of PulseDB sessions and 55.0% of MC-MED patients improve in Total MAE, with paired tests significant at p≈1.3 × 10^-16 and p≈1.4 × 10^-9.Improvements exceeding 1 mmHg occur in 29.4% of PulseDB sessions and 42.3% of MC-MED patients.
Per-Patient Improvement Distribution: RECAST helps most patients with larger TTC baseline errors, especially on PulseDB; this dependence is weaker on MC-MED.In the scatter, points farther below the equality diagonal represent larger absolute gains.
Hyperparameter Sensitivity: Removing similarity or uncertainty raises Total MAE to 10.39 or 10.36, while latest-only Recent reaches 10.40 versus 9.51 for full RECAST.The ablation supports the insufficiency of recency alone and the value of the combined selection strategy.

6 Conclusion

RECAST targets sample selection for stable streaming personalization by combining recent, contextually aligned, and reliable samples without changing the host model or training objective. It improves blood-pressure estimation across datasets, with patient-level benefits varying by dataset and patient difficulty.

The study identifies sample selection as a decisive component of buffered streaming TTA for non-stationary biosignals with abrupt regime changes.
RECAST combines temporal recency, contextual similarity, and predictive reliability to select adaptation samples for streaming biosignal TTA.
RECAST is the most accurate compared method on PulseDB and MC-MED, attaining the lowest MAE and RMSE on all three targets.
Per-patient gains are broad on PulseDB but concentrate on the hardest and most variable MC-MED patients, where most patients vary little within a stay.
RECAST adds sub-second per-segment latency on a single GPU with one CPU core while leaving the model and training objective unchanged.

A Appendix: Per-Target Per-Patient Breakdown

RECAST improves per-patient SBP and DBP errors for most patients on both datasets. Its advantage over TTC is strongest for patients with larger baseline errors and is not confined to one blood-pressure component.

The per-patient improvement pattern holds separately for SBP and DBP rather than arising only from aggregate Total performance.
RECAST improves SBP error for 68.5% of PulseDB patients and 54.5% of MC-MED patients.
RECAST improves DBP error for 76.7% of PulseDB patients and 54.9% of MC-MED patients.
For both SBP and DBP, most patient-level comparisons favor RECAST over TTC, with larger gaps at higher baseline TTC error.

B Appendix: Within-Patient Variability, by Target

The within-patient variability analysis shows similar RECAST gains for SBP and DBP. Gains emerge after modest variability, then saturate because abrupt transitions remain difficult for buffered methods to track.

In the most stable quintile, RECAST improves 51.6% of patients for SBP with a +1.3% median gain and 51.9% for DBP with a +1.2% gain.
From the second variability quintile onward, SBP and DBP improvement rises to about 55–63% of patients with +3 to +6% median gain.
The variability effect is not specific to SBP or DBP, because both targets follow the Total pattern.
Gains follow a threshold pattern rather than increasing monotonically: they saturate at very high variability, where abrupt transitions remain difficult to track.

C Appendix: MC-MED Hyperparameter Sensitivity

On MC-MED, tighter contextual matching and stronger recency weighting improve performance, while the uncertainty threshold has little effect once similarity filtering is tight.

Total MAE increases from 19.44 to 20.58 as temporal decay weakens from 𝜆𝑡=0.1 to 𝜆𝑡=0.01.
Total MAE increases from 19.44 to 20.32 as the similarity threshold loosens from 𝜏sim=0.9 to 𝜏sim=0.5.
The chosen configuration, 𝜏sim=0.9, 𝜆𝑡=0.1, and 𝜏unc=0.9, attains the lowest reported Total MAE of 19.44.
Across 𝜏unc ∈ [0.5, 0.9], all rows are identical because tight similarity filtering already removes most off-context candidates.
Removing similarity filtering raises Total MAE from 19.44 to 20.28, whereas removing uncertainty filtering leaves it statistically unchanged at 19.43 versus 19.44.