Bedside monitors stop guessing from 10-second clips: BioLT streams ECG, PPG and EEG into one trajectory, topping detection AUROC in 15 of 21 settings
Synopsis
Clinical ECG, PPG and EEG now arrive as hours-long streams, yet most analysis still runs on isolated windows of about 10 seconds and cannot handle long, irregular, frequently interrupted signals. BioLT turns any window-level feature extractor or pretrained biosignal foundation model into a streaming model: it feeds the elapsed time between windows as an explicit input channel and keeps long-range context in a fixed-size compressed memory bank, so training and inference stay linear in recording length. Across CPSC-2021, VitalDB, MC-MED and CHB-MIT, covering arrhythmia, sepsis and seizure detection plus forecasting up to 20 minutes, BioLT had the highest detection AUROC in 15 of 21 dataset-feature settings and the highest 20-minute forecasting AUROC in 13.
Interpretation
BioLT moves the unit of biosignal analysis from a single 10-second window to the whole continuous recording, giving a causal prediction at each moment from windows already received and updating it as new windows arrive. Earlier approaches split into two extremes: coarse summaries such as heart-rate variability and entropy computed over long recordings, which keep global state but discard temporal ordering, and fine-grained snapshot analysis on roughly 10-second windows that treats each window as an independent sample and ignores the physiological trajectory. On four datasets, CPSC-2021, VitalDB, MC-MED and CHB-MIT, three modalities of ECG, PPG and EEG, and domain features plus embeddings from six foundation models, 21 dataset-feature settings in all, BioLT reached the highest detection AUROC in 15 settings and the highest 20-minute forecasting AUROC in 13.
BioLT is a decoder-only temporal backbone that takes each valid window's representation together with its elapsed time since the previous window as input, and carries history across chunks in a fixed-size memory bank. Elapsed time is no longer treated as missingness to be imputed but as an explicit input channel, so the model can learn gaps as clinically meaningful information; the memory bank uses learnable queries to compress processed chunks into a few memory slots updated first-in first-out, avoiding quadratic attention over every past window. Memory slots are supervised by cumulative summary statistics, the mean, standard deviation, linear trend and dynamic range of hidden states up to that point; on MC-MED sepsis forecasting at 10 minutes, removing the gap input lowered AUROC from 60.29 to 58.02 on ECG and from 63.19 to 60.61 on PPG, and removing the memory bank as well lowered them further to 53.82 and 53.74.
Perspective
Evaluation is limited to arrhythmia, sepsis and seizure detection and forecasting up to 20 minutes on four public datasets, CPSC-2021, VitalDB, MC-MED and CHB-MIT, with patient-level 60/20/20 splits, downstream training sequences cropped to 500 windows and testing on full recordings; these are preprint results not yet peer reviewed and should not guide clinical practice.
Forecasting margins are generally smaller than detection margins, which the authors attribute to forecasting being inherently harder; on CHB-MIT seizure forecasting at 20 minutes, BioLT with domain features reached AUROC 51.10, below the snapshot method with CBraMod embeddings, a relative difference of minus 6.61. The authors also note smaller gains on atrial fibrillation tasks where local rhythm morphology is already informative.