An approach for choosing which stored waveform segments should drive each model update reported lower blood-pressure estimation error than TTC, the paper’s standard adaptation setup, and a no-adaptation baseline on PulseDB and MC-MED. RECAST had the lowest reported mean absolute error (MAE) and root mean squared error (RMSE) for the total measure and for systolic and diastolic blood pressure on both datasets.
On PulseDB, total MAE was 9.51 mmHg with RECAST and 10.05 mmHg with TTC, a reported reduction of 5.4%. On MC-MED, the figures were 19.44 and 20.04 mmHg, respectively, a reported reduction of 3.0%. These values are point estimates; the analysis did not report confidence intervals.
The change is in the data selection
RECAST leaves the host model and the TTC training objective unchanged. It alters only sample selection, building each adaptation batch from three signals: temporal recency, contextual similarity in latent space and predictive reliability, represented by uncertainty filtering. In other words, the method focuses on which examples from the buffer should guide the next update as new waveform segments arrive.
Candidates receive a joint score that rewards similarity to the current segment and recency in the buffer. The method age-penalizes uncertainty, then retains a candidate only when both its similarity and its uncertainty meet their respective thresholds. Recency is measured by buffer insertion order, which may not match real elapsed time when supervision is irregular.
The reported implementation used two buffers with a capacity of 64 samples each. Adaptation batches contained 32 samples, with a labeled-to-total ratio of 0.25; the underlying model received five updates per input, and uncertainty was estimated from 20 stochastic forward passes.
Two benchmarks, with different scales
The evaluation used streaming waveform segments with intermittent blood-pressure references from PulseDB and MC-MED. PulseDB had 2,506 training visits and 279 test visits, containing 902,160 and 111,600 segments respectively. MC-MED had 21,342 training visits and 2,756 test visits, with 2,411,288 training segments and 313,638 test segments.
MC-MED’s cuff references covered about 2% of segments. For the per-patient comparison, the evaluation set included 2,244 of the 2,756 test patients who had at least one cuff reference.
The pattern held beyond one average score
The analysis also looked at trend tracking, using Pearson correlation for systolic blood pressure. RECAST had the highest reported SBP correlation on both datasets. On PulseDB, the rounded correlation rose from 0.42 without test-time adaptation to 0.90 with RECAST; on MC-MED, it rose from 0.46 to 0.79. RMSE was lowest for every target.
RECAST’s lower total error was not confined to a small number of evaluation units. Relative to TTC, total MAE fell for 74.6% of the 279 PulseDB test sessions and 55.0% of the 2,244 MC-MED patients. The paired Wilcoxon signed-rank tests were reported as significant, with approximate p-values of 1.3 times 10 to the minus 16 and 1.4 times 10 to the minus 9, respectively. The supplied analysis did not report confidence intervals.
The improvement was uneven across patients. In a pooled analysis, 51.6% of patients in the most stable within-patient variability group improved, with a median gain of 0.9%. From the second variability quintile onward, 58% to 63% improved and the median gain was about 4%, before the share eased to 59.6% in the most variable quintile. The authors present this as a pooled subgroup analysis, not a prespecified clinical subgroup outcome.
A trade-off in computation
An ablation analysis tested what happened when parts of the selector were removed. On PulseDB, full RECAST had a total MAE of 9.51 mmHg. The version without contextual similarity recorded 10.39 mmHg, the version without uncertainty recorded 10.36 mmHg, and the version using only the most recent samples recorded 10.40 mmHg. The reported comparisons favored the complete strategy over recency alone.
RECAST’s latency was higher than the baselines, at 428.87 milliseconds per segment on PulseDB and 435.79 milliseconds on MC-MED. The paper reports that latency remained below one second, with the two figures equal to 4.29% and 0.73% of segment duration. Runtime was measured using one A100 GPU and one CPU core, so the numbers describe the reported hardware setup.
What the benchmark leaves unanswered
The results are predictive-performance findings from simulated streaming benchmarks, not evidence of improved patient outcomes, clinical safety or prospective bedside effectiveness. The study did not report prospective clinical deployment or patient-outcome evaluation. Its own limitations are that insertion-order recency may fail to reflect elapsed time under irregular supervision, that gains depend on within-patient variation, and that the method was instantiated only within TTC. The findings therefore do not establish superiority outside the two evaluated datasets or the TTC-based implementation.
The document is an arXiv version-1 preprint dated 28 Aug 2026. It lists Medical AI Co., Ltd. as the authors’ affiliation, while no funding source or journal publication status is reported in the supplied text.
Paper data and sources
Original title: RECAST: Recent & Context-Aware Sampling for Test-Time Adaptation in Streaming Biosignals
Authors: Yong-Yeon Jo, Junho Song, Joon-myoung Kwon
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text