Loading…
Physiological time-series classification model bringing in-context learning to ECG, EEG, and ICU signals with temporal and channel-aware encodings.
No providers recorded yet. Browse all providers
Shuffle the columns of a clinical spreadsheet and a tabular model's prediction should not change. That permutation invariance is precisely what lets TabPFN generalize across arbitrary tables. Shuffle the 140 samples of an ECG heartbeat and the heartbeat is gone. TSPFN — Time-Series Prior-Data Fitted Network — is an attempt to keep what tabular foundation models buy, supervised prediction from a hundred labeled examples with no gradient updates, while putting back the temporal and channel structure that physiological recordings depend on.
Prior-data fitted networks are transformers pretrained to approximate Bayesian inference in context: a support set of labeled examples and an unlabeled query enter the same forward pass, and a prediction comes out without any parameter being updated. TabPFN reaches that behavior by pretraining on synthetic tables drawn from a prior over structural causal models. Sampling realistic synthetic EEG or ECG with plausible clinical labels is far harder, so TSPFN substitutes real data — roughly 140,000 recordings pooled from four public clinical corpora — and adds positional machinery that lets the model tell time steps apart from channels inside a flattened row.
The model comes from the ISIR lab at Sorbonne University, with collaborators at Inria and the CREATIS unit in Lyon. It is a fifth biology-facing member of the PFN family in this catalog, alongside PerturbPFN, KinPFN, GPFN and TabPFN-Wide — each of which swaps TabPFN's generic prior for a domain-specific one.
TSPFN is a 7.3M-parameter transformer, matching the scale of TabPFN-v2, and is initialized from pretrained TabPFN weights before being pretrained on the physiological corpus. Each input row concatenates one or more channels of a patient recording under the constraint T × C ≤ 500, using the configurations (250, 2), (166, 3), (125, 4) and (100, 5). Pretraining draws on TUEV and TUAB for EEG, PTB-XL for ECG and HiRID for multivariate ICU signals, totalling about 140,000 samples, with each sample represented four times at different channel depths. Datasets are partitioned into chunks of 5,000 samples split into balanced support and query halves, and cross-entropy is computed only on the query labels; optimization uses AdamW on a single H200 GPU, converging after roughly 25 epochs. On a five-dataset benchmark — eICU-CRD, ESR, EOS, ECG5000 and CPSC 2018 — evaluated by stratified 5-fold cross-validation with 100 support samples, TSPFN averages 82.5 AUC and 57.3 AUPRC against 77.5 and 49.2 for TabPFN, 78.1 and 53.6 for MiniRocket, and 64.8 and 37.3 for LaBraM. A DeLong test against MiniRocket on AUROC gives p = 3.7 × 10⁻⁵. The ablation is informative about the design: rotary encoding alone degrades AUPRC sharply on the ECG benchmarks, dropping CPSC from 65.1 to 33.2, and only recovers once channel identity embeddings are added.
The target setting is the ordinary clinical study rather than the large annotated benchmark: a single-center cohort, a pilot recording protocol, or a rare diagnostic class with a few dozen positive examples. A clinical researcher supplies a handful of labeled recordings as context and gets predictions without assembling a training pipeline or tuning a deep network. Because the same checkpoint spans ECG, EEG and ICU vital signs, it also serves as a uniform baseline across modalities that would otherwise each need a specialized model.
TSPFN is evidence that the prior-data fitted recipe transfers to signals once the prior is built from real recordings instead of synthetic tables, and that positional structure has to be reintroduced jointly — channel identity and temporal ordering are not separable improvements. The evaluation is narrow: five datasets subsampled to 100 examples per task, with specialized baselines still winning on their home benchmarks, which the authors report plainly. The experiment code, ablations and preprocessing scheme are released under Apache 2.0, but no pretrained checkpoint accompanies them, and for an in-context model the weights are the artifact — without them the reported low-data behavior cannot be reproduced.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.