Early-warning models for the intensive care unit (ICU) are trained to predict a single fixed event, such as "sepsis onset within six hours," encoded as a bare 0/1 label. Hospitals define that event differently. Sepsis-3 onset, the PhysioNet 2019 onset rule, and local alert criteria all disagree on timestamps. Hospitals also keep adding new deterioration endpoints, such as vasopressor start, kidney-injury stages, or lactate escalation. Every new definition currently requires a new labelled cohort and a retrained model.
We propose a model that takes a patient's irregular vital-sign and lab record together with a written definition of an event, including its time horizon, and returns the probability that the event will happen. It is trained on many definitions so that it can score definitions it has never seen. The patient side uses a multitask Gaussian process (MGP), a probabilistic interpolator that fills gaps between irregular measurements using correlations across channels and reports how uncertain each filled-in value is. The model's attention to each channel is driven by the definition text. When a definition hinges on a channel the patient never had measured, the model is forced to say so instead of inventing the value.
In short, the model is a classifier whose label is a sentence. It reads that sentence to decide which physiological channels matter, and it refuses to be confident when those channels are unmeasured.
Core Claim
Semantic labels earn their place in clinical early warning for one reason: they let a model transfer to new or redefined endpoints without new labels. They do not obviously help on a single, fixed binary target, where a sigmoid head is hard to beat.
Our claim has two parts. First, a model trained against a large, programmatically generated family of textual endpoint definitions can score held-out definitions zero-shot. Its accuracy should approach that of a model trained directly on those labels, and it should beat the obvious rule-based alternative of forecasting the MGP forward and applying the written criterion. Second, routing the MGP posterior variance through definition-specific channel attention produces calibrated abstention. When the channels a definition depends on are unmeasured, predicted risk should move toward the prior and expected information gain should rise. This directly targets the failure in which an LLM sepsis model stated lab values that were never recorded.
Where This Sits
The patient encoder comes from the MGP line of work. Learning to Detect Sepsis with a Multitask Gaussian Process RNN Classifier coupled an MGP to an LSTM. It passed posterior mean and variance downstream and used the Lanczos method to make backpropagation through the GP tractable. Early Recognition of Sepsis with Gaussian Process Temporal Convolutional Networks and Dynamic Time Warping replaced the LSTM with a dilated temporal convolutional network (MGP-TCN). We reuse both the Lanczos machinery and the TCN. Both papers end in a single semantically empty binary head trained on one definition, and that head is what we replace.
Learning to Detect Sepsis with a Multitask Gaussian Process RNN Classifier also supplies a design constraint. Its stationary Ornstein-Uhlenbeck kernel reverts to the mean, so MGP extrapolation over several hours is weak. This tells us not to make prediction a pure forecast-then-threshold step. It also predicts that the forecast-then-rule baseline below will degrade with horizon, which is a testable consequence.
The label side comes from text classification. Description-Enhanced Label Embedding Contrastive Learning for Text Classification aligns inputs with dictionary definitions using two-way attention and a contrastive loss. The Benefits of Label-Description Training for Zero-Shot Text Classification shows that training on label descriptions gives large zero-shot gains: 17.5 to 18.9 points of accuracy for RoBERTa. Both methods work only on text inputs. We borrow the dual-encoder and contrastive alignment idea. We differ by matching definitions against probabilistic physiological trajectories, and by letting the definition choose which channels' uncertainty counts.
The motivating evidence is PathwayLLM: Explainable Clinical Trajectory Modeling with Structured Pathways for Sepsis Prediction. It fuses EHR encoders with an LLM and reports MIMIC-IV AUROC 0.891 and AUPRC 0.724. It also hallucinates unmeasured lab values in about 6% of cases and conflates heart-failure deterioration with sepsis. Integrating Physiological Time Series and Clinical Notes with Transformer for Early Prediction of Sepsis brings text in on the input side through ClinicalBERT but keeps a binary output. Our model gives text an output-side role instead. Time pressure motivates the information-gain output: Transformer-Based Unsupervised Learning for Early Detection of Sepsis (Student Abstract) cites roughly 7% added mortality per hour of treatment delay. Li-Fang Cheng et al. used multi-output GPs to guide lab ordering, which we extend to definition-conditioned ordering.
Research Program
Definition corpus. On MIMIC-IV we build a corpus of a few hundred machine-checkable endpoint definitions. Each is written in clinical prose and paired with executable logic that labels every patient-hour. The definitions vary along five axes:
- organ system (cardiovascular, renal, respiratory, metabolic, infection-linked);
- criterion type (threshold, rise from baseline, treatment initiation);
- thresholds;
- horizon (one to twelve hours);
- composites (for example, a Sepsis-3-style definition combining suspected infection with a SOFA rise).
Each definition gets several paraphrases, so the model cannot memorize surface strings. This addresses the hand-curation bottleneck that Description-Enhanced Label Embedding Contrastive Learning for Text Classification and The Benefits of Label-Description Training for Zero-Shot Text Classification both depend on: here every gloss has an executable ground truth.
Model. The MGP turns each patient's record into posterior means $\mu_c(t)$ and variances $\sigma^2_c(t)$ on an hourly grid for each channel $c$. Two features are appended per channel and hour: the measurement indicator and the time since the last measurement. This keeps informative missingness visible rather than smoothing it away, following The Signature-Based Model for Early Detection of Sepsis From Electronic Health Records in the Intensive Care Unit. A TCN produces per-channel states $z_c(t)$. A text encoder maps a definition $d$ to a channel-attention vector $a(d)$ and to per-channel query vectors $q_c(d)$. The risk score is
$$p(d \mid x, t) = \mathrm{sigmoid}\Big(b(d) + \sum_c a_c(d)\, \langle z_c(t), q_c(d)\rangle\Big),$$
where $b(d)$ is a text-predicted base rate that encodes prevalence. The TCN receives $\sigma^2_c$ as input, so a high-variance channel can shrink its own contribution.
Training uses binary cross-entropy summed over definitions and patient-hours. It adds a contrastive term that pulls trajectories toward the definitions they satisfy and away from those they do not. Paraphrases of the same definition share labels. The information-gain output is the expected reduction in the entropy of $p(d \mid x, t)$ from measuring channel $c$ now. It is estimated by sampling hypothetical observations from the MGP posterior and recomputing the score.
Splits. Held-out definitions are split three ways, in increasing difficulty:
- new thresholds or horizons on seen criterion types (interpolation);
- new composites of seen criteria;
- a held-out organ system, kidney injury (extrapolation).
The PhysioNet 2019 onset rule and Moor's Sepsis-3 operationalization are written out as text and always held out. This makes definition drift a zero-shot test.
Baselines
The naive reference point combines forward-fill imputation, a GRU, and one sigmoid head per training definition. Held-out definitions are scored by the head whose definition text is nearest. This tells us whether text-driven scoring beats simply picking the closest known label.
The strongest existing architecture is MGP-TCN from Early Recognition of Sepsis with Gaussian Process Temporal Convolutional Networks and Dynamic Time Warping. It reports AUPRC 0.35 at a seven-hour horizon on its MIMIC-III Sepsis-3 cohort, against 0.40 for DTW-KNN and 0.25 for MGP-RNN. We run MGP-TCN in two ways. As an oracle, it is trained directly on each held-out definition's labels, which bounds what zero-shot transfer could achieve. As a multi-head model, it gets one head per training definition. Both versions use the same MGP, TCN size, data splits, and hyperparameter search budget as our model, measured in GPU-hours. On the original Sepsis-3 cohort we also report the replicated 0.35 for comparability. Beating this baseline means one of two things. Either we come within roughly 80% of the oracle's AUPRC on held-out definitions while the multi-head model falls well short, or we match the oracle on seen sepsis definitions without loss.
The forecast-then-rule baseline extrapolates the MGP posterior to the horizon, samples trajectories, and applies each definition's executable logic. This is the most dangerous competitor because it is zero-shot by construction.
For the LLM comparison we fine-tune a LoRA model on serialized hourly values plus the definition text. This is a reimplementation in the style of PathwayLLM, under a matched budget. PathwayLLM itself reports MIMIC-IV AUROC 0.891 and AUPRC 0.724, but on its own cohort, so those numbers are context rather than a target.
The key ablation feeds MGP means without variances and removes the definition-driven attention by making it uniform. This isolates whether channel attention combined with uncertainty is what produces both transfer and abstention.
For zero-shot scoring of held-out ICU endpoint definitions, no reported result found on Lacuna. For PhysioNet 2019, the reference is the winning test utility of 0.360 from The Signature-Based Model for Early Detection of Sepsis From Electronic Health Records in the Intensive Care Unit. That number comes from a model trained on that challenge's labels, so a zero-shot model reaching it would be a strong result.
Why not just…?
Why not just train a multi-label head with one output per known endpoint? On seen endpoints this will likely do as well as our model, and we expect that and report it. It has no way to score a definition introduced after training, short of nearest-head matching. The held-out-definition splits measure exactly this gap. If the multi-head nearest-neighbor baseline comes within a few points of our model on the interpolation split, then the language component is adding nothing there. In that case we would narrow the claim to composites and new organ systems.
Why not just forecast with the MGP and apply the written rule? This is a real threat for short horizons and threshold-type definitions. A good forecast plus exact logic is hard to beat. Our expectation that it fails at longer horizons rests on the mean-reverting kernel used in Learning to Detect Sepsis with a Multitask Gaussian Process RNN Classifier. It also cannot use precursor patterns that are not part of the criterion. We test this by plotting AUPRC against horizon for both methods. If forecast-then-rule wins beyond three hours, the learned matcher is not justified and the contribution reduces to the uncertainty-gated abstention analysis.
Why not just prompt or fine-tune an LLM on serialized values and the definition? This option also reads definitions, but PathwayLLM documents roughly 6% hallucinated lab values. Serialized tables also carry no calibrated notion of how stale a value is. Our masking test, described below, measures directly whether either model becomes appropriately unsure when a definition's key channel is removed.
Decisive Checks
The first check is zero-shot AUPRC and AUROC on each held-out split of MIMIC-IV. These are reported as a fraction of the per-definition oracle, with patient-level bootstrap intervals.
The second is cross-site definition drift. We train on MIMIC-IV and score the PhysioNet 2019 onset rule, given as text, on the PhysioNet data using the challenge utility metric. We also score the Sepsis-3 text on eICU-CRD and compare against PathwayLLM's reported zero-shot eICU AUROC of 0.842, while noting that the cohorts differ.
The third is the masking test. For definitions that cite lactate or creatinine, we delete those measurements in test patients. A grounded model should move its predictions toward the definition's base rate, and its information gain for the deleted channel should rank first. A hallucinating model keeps its confidence. We score this with expected calibration error on masked versus unmasked patients, and with the rank of the masked channel in the information-gain ordering.
The last check is the horizon curve against forecast-then-rule, described in the previous section. If the zero-shot split shows no gain over nearest-head matching, and the masking test shows no benefit from variance gating, the proposal is falsified.
References
- Learning to Detect Sepsis with a Multitask Gaussian Process RNN Classifier
- Early Recognition of Sepsis with Gaussian Process Temporal Convolutional Networks and Dynamic Time Warping
- Description-Enhanced Label Embedding Contrastive Learning for Text Classification
- The Benefits of Label-Description Training for Zero-Shot Text Classification
- PathwayLLM: Explainable Clinical Trajectory Modeling with Structured Pathways for Sepsis Prediction
- Integrating Physiological Time Series and Clinical Notes with Transformer for Early Prediction of Sepsis
- Transformer-Based Unsupervised Learning for Early Detection of Sepsis (Student Abstract)
- An Optimal Policy for Patient Laboratory Tests in Intensive Care Units
- The Signature-Based Model for Early Detection of Sepsis From Electronic Health Records in the Intensive Care Unit