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

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

cs.AIAug 28, 2026

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

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

Organizations: Medical AI Co., Ltd. Seoul, Republic of Korea

Abstract

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: 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 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.

Figures & tables

Figure 1. Streaming biosignals at an ICU bedside over 24 hours. Waveforms (gray) arrive continuously while only sparse cuff-BP checks are available.

Figure 2. Blood pressure changes with moving statistics for a certain patient in PulseDB. The solid black and gray lines represent raw SBP and DBP, respectively. The dashed lines denote the moving average. The shaded regions represent the moving standard deviation.

Dataset	Split	#Visits	#Segments	Seg/visit	BP (mmHg)		Within-pt std
Dataset	Split	#Visits	#Segments	Seg/visit	SBP	DBP	SBP	DBP
PulseDB	Train	2,506	902,160	360.0	118.60 $\pm$ 21.03	61.86 $\pm$ 12.65	12.4	6.8
PulseDB	Test	279	111,600	400.0	118.84 $\pm$ 20.59	62.00 $\pm$ 12.27	12.7	6.6
MC-MED	Train	21,342	2,411,288	113.0 $\pm$ 84.4	131.00 $\pm$ 24.88	78.66 $\pm$ 17.13	8.0	6.4
MC-MED	Test	2,756	313,638	113.8 $\pm$ 86.5	131.87 $\pm$ 23.97	79.14 $\pm$ 16.90	7.6	6.1

Table 1. Summary of datasets. Seg/visit is the mean

\pm

std number of segments per visit. SBP and DBP are the population mean

\pm

std over all segments; within-pt std is the median per-patient standard deviation of reference BP. Table summarizing the PulseDB and MC-MED datasets: number of visits and segments, segments per visit, population SBP and DBP mean and standard deviation, and within-patient BP standard deviation for the train and test splits. MC-MED shows larger across-patient spread but smaller within-patient BP variation than PulseDB.

Dataset	Method	MAE $\downarrow$			Corr. $\uparrow$		RMSE $\downarrow$			Std $\downarrow$
Dataset	Method	Total	SBP	DBP	SBP	DBP	Total	SBP	DBP	SBP	DBP
PulseDB	No TTA	24.21	14.85	9.36	0.422	0.423	30.86	18.97	11.89	11.81	7.32
	TTT-MAE	25.60	16.01	9.58	0.232	0.280	32.34	20.14	12.21	12.21	7.56
	TTC	10.05	6.52	3.54	0.891	0.901	14.77	9.41	5.37	6.78	4.04
	RECAST	9.51	6.20	3.30	0.896	0.909	14.35	9.20	5.15	6.80	3.95
MC-MED	No TTA	29.52	17.53	11.99	0.460	0.342	38.78	22.73	16.05	14.47	10.66

Table 2. Performance comparison of streaming TTA methods on PulseDB and MC-MED. MAE and RMSE are the mean absolute and root-mean-square errors for total, systolic (SBP), and diastolic (DBP) blood pressure; Std is the error standard deviation for SBP and DBP (all lower is better); Corr. is the Pearson correlation for SBP and DBP (higher is better). Bold indicates the best value within each dataset block. Table comparing streaming TTA methods on PulseDB and MC-MED. For each method we report total, SBP, and DBP MAE and RMSE, SBP and DBP error standard deviation, and SBP and DBP Pearson correlation. RECAST achieves the best MAE, RMSE, and correlation in both dataset blocks, with minor Std exceptions.

Quintile	Total		# patients
Quintile	% imp.	gain (%)	MC-MED	PulseDB
Excluded	—	—	852	0
Q1	51.6	$+0.9$	336	1
Q2	57.5	$+4.3$	310	22
Q3	62.9	$+4.4$	266	68
Q4	62.9	$+4.3$	244	90

Table 3. Performance by within-patient SBP variability (Total error), pooled across PulseDB and MC-MED. Patients are sorted into quintiles Q1 (most stable) to Q5 (most variable) by within-patient SBP std. % imp. is the fraction of patients RECAST improves over TTC and gain the median relative improvement. Per-quintile table relating within-patient SBP variability to RECAST's Total improvement. The most stable quintile, almost entirely MC-MED, improves about half of patients with about +0.9% median gain; higher-variability quintiles reach 58 to 63% improved with about +4% median gain, where most PulseDB patients fall.

Panel of Figure 3 — Figure 3. Per-patient Total $\Delta$ MAE (MAE TTC $-$ MAE RECAST ) distribution. Positive values on the $x$ -axis indicate RECAST improvement. Histograms of per-patient Total $\Delta$ MAE for PulseDB (left) and MC-MED (right). The horizontal axis is per-patient Total $\Delta$ MAE (TTC minus RECAST), and bars to the right of the dashed zero line count patients that RECAST improves. On PulseDB most sessions lie to the right of zero; on MC-MED the distribution is more balanced, with a right-leaning tail of large gains.

Panel of Figure 4 — Figure 4. Per-patient Total MAE scatter: TTC on the $x$ -axis vs. RECAST on the $y$ -axis. Each point is one patient (or session). Points below the diagonal indicate RECAST improvement; points farther below indicate larger absolute gains. Scatter of per-patient Total MAE, TTC on the horizontal axis and RECAST on the vertical axis, for PulseDB (left) and MC-MED (right). Most points lie below the diagonal, more so on PulseDB, and the gap widens as the TTC error grows.

Panel of Figure 5 — Figure 5. Queue index-wise usage frequency of adaptation samples, aggregated over the entire test stream. Each panel plots how often each buffer position (FIFO queue index, $x$ -axis) is selected into the adaptation batch ( $y$ -axis), shown separately for the labeled and unlabeled queues. The top row (a)–(b) is PulseDB and the bottom row (c)–(d) is MC-MED, each comparing TTC and RECAST. A flat profile indicates index-agnostic selection, whereas a dispersed (non-uniform) profile indicates preference for specific buffer regions. For each method, a plot of how frequently each queue position is chosen for adaptation, shown separately for labeled and unlabeled queues; RECAST's distribution is more spread out than TTC's on both PulseDB and MC-MED.

Setting	MAE $\downarrow$			Correlation $\uparrow$
	Total	SBP	DBP	SBP	DBP
(A) Temporal decay rate $\boldsymbol{\lambda_{t}}$
0.10	9.51	6.20	3.30	0.896	0.909
0.05	10.00	6.51	3.49	0.889	0.902
0.01	10.50	6.82	3.68	0.881	0.894
(B) Similarity threshold $\boldsymbol{\tau_{\mathrm{sim}}}$

Table 4. Ablation study on RECAST components and key hyperparameters on PulseDB. Best results in each block are boldfaced. Ablation table on PulseDB with four blocks: (A) temporal decay rate, (B) similarity threshold, (C) uncertainty threshold, and (D) component ablation. Each block reports total, SBP, and DBP MAE and SBP and DBP correlation. The chosen configuration attains the best values, and removing similarity, uncertainty, or using only the most recent sample degrades performance.

Method	PulseDB (10 s)		MC-MED (60 s)
Method	ms	Ratio (%)	ms	Ratio (%)
No TTA	40.35	0.40	79.78	0.13
TTC	266.27	2.66	314.41	0.52
RECAST	428.87	4.29	435.79	0.73

Table 5. Runtime cost per segment on a single GPU and one CPU core. Ratio (%) is the per-segment latency divided by the segment duration. Per-segment runtime in milliseconds and as a percentage of the segment duration, for No TTA, TTC, and RECAST on PulseDB (10,s segments) and MC-MED (60,s segments). RECAST adds modest latency, around 4% of the window on PulseDB and under 1% on MC-MED.

Panel of Figure 6 — Figure 6. Per-patient $\Delta$ MAE (MAE TTC $-$ MAE RECAST ) distribution, with separate SBP and DBP panels. Positive values indicate RECAST improvement. Per-patient delta-MAE histograms for PulseDB (left) and MC-MED (right), each with an SBP and a DBP panel. Most patients lie on the improvement side, more so on PulseDB.

Panel of Figure 7 — Figure 7. Per-patient MAE scatter: TTC on the $x$ -axis vs. RECAST on the $y$ -axis, with separate SBP and DBP panels. Points below the diagonal indicate RECAST improvement. Per-patient MAE scatter for PulseDB (left) and MC-MED (right), each with an SBP and a DBP panel. Most points lie below the diagonal and the gap widens with TTC error.

Quintile	within-pt std		SBP		DBP		# patients
Quintile	SBP	DBP	% imp.	gain (%)	% imp.	gain (%)	MC-MED	PulseDB
Excluded	—	—	—	—	—	—	852	0
Q1 (most stable)	2.7	4.2	51.6	$+1.3$	51.9	$+1.2$	336	1
Q2	6.2	5.0	54.5	$+3.1$	59.6	$+4.2$	310	22
Q3	9.2	5.9	61.1	$+4.3$	62.9	$+5.6$	266	68
Q4	12.5	7.7	59.9	$+3.9$	61.7	$+5.1$	244	90

Table 6. Per-target within-patient-variability breakdown (SBP and DBP), pooled across PulseDB and MC-MED. Quintiles are by within-patient SBP std (Q1 most stable to Q5 most variable). The within-pt std columns are the median per-patient standard deviation of reference SBP and DBP, in mmHg. % imp. is the fraction of patients RECAST improves over TTC, and gain is the median relative improvement, for SBP and DBP. The bottom rows reconcile the quintiled counts with the full evaluation set. Per-quintile table of RECAST's SBP and DBP improvement against within-patient SBP variability. SBP and DBP improve fewest patients in the most stable quintile and rise to 55 to 63 percent improved with about +4 percent median gain in the higher-variability quintiles. The bottom rows show 852 single-reading MC-MED patients are excluded, so 1,392 quintiled plus 852 equal the 2,244 evaluated patients.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Setting	MAE $\downarrow$			Correlation $\uparrow$
	Total	SBP	DBP	SBP	DBP
(A) Temporal decay rate $\boldsymbol{\lambda_{t}}$
0.10	19.44	11.22	8.22	0.791	0.699
0.05	19.92	11.51	8.41	0.777	0.689
0.01	20.58	11.98	8.60	0.762	0.685
(B) Similarity threshold $\boldsymbol{\tau_{\mathrm{sim}}}$

Appendix

Table 7. Ablation study on RECAST components and key hyperparameters on MC-MED. Best results in each block are boldfaced. Ablation table on MC-MED with four blocks: (A) temporal decay rate, (B) similarity threshold, (C) uncertainty threshold, and (D) component ablation. Each block reports total, SBP, and DBP MAE and SBP and DBP correlation. The chosen configuration is best in the decay and similarity sweeps; the uncertainty threshold has negligible effect, while removing similarity or using only the most recent sample degrades performance.

Adaptive Sampling

Adaptive sampling allocates limited data, measurements, or computation where they are expected to be most useful.

See all 205 papers

RG-TTA: Regime-Guided Meta-Control for Test-Time Adaptation in Streaming Time Series
Mar 29, 2026Indar Kumar, Akanksha Tiwari, Sai Krishna Jasti +1Test-Time Adaptation Continual Test-Time Adaptation
independent researchers
Temporal Structure Matters for Efficient Test-Time Adaptation in Wearable Human Activity Recognition
May 6, 2026Zishu Zhou, Zaipeng Xie, Xuanyao JieTest-Time Adaptation Human Activity Recognition
College of Computer Science and Software Engineering, Hohai University
MedTS-TTT: Test-Time Training for Medical Time Series Classification
Jun 19, 2026Mingzhi Chen, Yiyu Gui, Guibo LuoTime Series Classification Electrocardiogram Classification
Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University, Shenzhen, China

Mar 29, 2026J/K move · Enter open · S save

RG-TTA: Regime-Guided Meta-Control for Test-Time Adaptation in Streaming Time Series

Indar Kumar, Akanksha Tiwari, Sai Krishna Jasti, Ankit Hemant Lade

independent researchers

Test-Time Adaptation Continual Test-Time Adaptation Time Series Forecasting

Open paper AskPDF arXiv

Test-time adaptation (TTA) enables neural forecasters to adapt to distribution shifts in streaming time series, but existing methods apply the same adaptation intensity regardless of the nature of the shift. We propose Regime-Guided Test-Time Adaptation (RG-TTA), a meta-controller that continuously modulates adaptation intensity based on distributional similarity to previously-seen regimes. Using an ensemble of Kolmogorov-Smirnov, Wasserstein-1, feature-distance, and variance-ratio metrics, RG-TTA computes a similarity score for each incoming batch and uses it to (i) smoothly scale the learning rate -- more aggressive for novel distributions, conservative for familiar ones -- and (ii) control gradient effort via loss-driven early stopping rather than fixed budgets, allowing the system to allocate exactly the effort each batch requires. As a supplementary mechanism, RG-TTA gates checkpoint reuse from a regime memory, loading stored specialist models only when they demonstrably outperform the current model (loss improvement >= 30%). RG-TTA is model-agnostic and strategy-composable: it wraps any forecaster exposing train/predict/save/load interfaces and enhances any gradient-based TTA method. We demonstrate three compositions -- RG-TTA, RG-EWC, and RG-DynaTTA -- and evaluate 6 update policies (3 baselines + 3 regime-guided variants) across 4 compact architectures (GRU, iTransformer, PatchTST, DLinear), 14 datasets (6 real-world multivariate benchmarks + 8 synthetic regime scenarios), and 4 forecast horizons (96, 192, 336, 720) under a streaming evaluation protocol with 3 random seeds (672 experiments total). Regime-guided policies achieve the lowest MSE in 156 of 224 seed-averaged experiments (69.6%), with RG-EWC winning 30.4% and RG-TTA winning 29.0%. Overall, RG-TTA reduces MSE by 5.7% vs TTA while running 5.5% faster; RG-EWC reduces MSE by 14.1% vs standalone EWC.

Abstract

Adaptive Sampling

Explore similar work

RG-TTA: Regime-Guided Meta-Control for Test-Time Adaptation in Streaming Time Series

Temporal Structure Matters for Efficient Test-Time Adaptation in Wearable Human Activity Recognition

MedTS-TTT: Test-Time Training for Medical Time Series Classification