SepsisLens: Structure-Preserving Sequence Modelling for Decomposable Early Sepsis Warning

Yikun Ou; Wei Li

cs.LGOct 6, 2026

SepsisLens: Structure-Preserving Sequence Modelling for Decomposable Early Sepsis Warning

Authors: Yikun Ou, Wei Li

Organizations: School of Computer Science, University of Sydney Sydney, Australia

Abstract

Early sepsis warning from ICU records can be cast as a structure-preserving prediction problem. A model needs to detect deterioration from irregular measurements while keeping each alert connected to the physiological signals that support it. Many temporal models fuse clinical variables into a patient-level representation, supporting scalar risk prediction but weakening the structure needed for clinical decomposition. We present SepsisLens, which preserves variable-indexed temporal states until risk composition. Observation-aware representations encode each variable's dynamics and measurement history, while a shared temporal encoder models each trajectory without collapsing the variable axis. The StructuredRiskHead composes multi-horizon risk from explicit variable-level and organ-level components. We evaluate SepsisLens on three public ICU cohorts and one private-hospital cohort under a common pre-onset protocol. SepsisLens achieves strong discrimination on all four cohorts and lower alert burden at matched event recall on MIMIC-IV. Structural ablations support the design, while input-side masking shows that the ranked components reflect variables with greater influence on prediction.

Figures & tables

Figure 1: Motivation for SepsisLens. (a) Irregular ICU observations contain information about variable values, changes, and measurement history. (b) Early fusion collapses variable trajectories into a global representation without explicit variable- or organ-level components. Post-hoc attribution is applied outside the risk computation. (c) SepsisLens preserves variable-indexed temporal states until risk composition, producing multi-horizon risk with explicit variable- and organ-level components.

Figure 2: SepsisLens architecture. (a) Observation-aware representations encode value, trend, elapsed time, and never-observed status for each clinical variable. (b) A shared temporal encoder models variable trajectories and preserves variable-indexed states $\mathbf{H}\in\mathbb{R}^{B\times T\times V\times d}$ . (c) StructuredRiskHead computes gated variable-level components, aggregates them into organ-level components, and adds a global residual correction. (d) A fixed extended assignment maps all variables to six domain groups organised around the SOFA organ systems.

Method	#P (K)	AUC	APC	F1
SepsisLens	91.1	$\mathbf{.939{\pm}.004}$	$\mathbf{.474{\pm}.019}$	$\mathbf{.504{\pm}.017}$
SepsisCalc	225.0	$.910{\pm}.001$	$.271{\pm}.005$	$.396{\pm}.002$
Transformer	400.8	$.909{\pm}.001$	$.293{\pm}.006$	$.407{\pm}.003$
GRU-D	222.1	$.905{\pm}.002$	$.299{\pm}.009$	$.418{\pm}.006$
LSTM	214.7	$.904{\pm}.000$	$.288{\pm}.004$	$.409{\pm}.002$
Raindrop	235.4	$.904{\pm}.002$	$.268{\pm}.009$	$.396{\pm}.006$

Table 1: Primary 6 h early-warning results on MIMIC-IV (mean

\,\pm\,

std over three seeds).

Cohort	Test $n$	AUC	APC	F1
MIMIC-IV	6,578	$.939{\pm}.004$	$.474{\pm}.019$	$.504{\pm}.017$
MIMIC-III	4,660	$.996{\pm}.000$	$.763{\pm}.005$	$.747{\pm}.002$
eICU	2,251	$.948{\pm}.002$	$.219{\pm}.007$	$.287{\pm}.004$
Private	5,823	$.973{\pm}.001$	$.404{\pm}.006$	$.428{\pm}.002$

Table 2: SepsisLens 6 h early-warning results on ICU cohorts.

Figure 3: Alert burden and structured risk outputs. (a) FAR versus event recall on MIMIC-IV over a threshold sweep. The marker denotes SepsisLens at its validation-selected threshold. The horizontal dashed line marks the $\mathrm{eR}=0.95$ target used in Table 3 . (b,c) Organ-level and variable-level components for a representative true-positive pre-onset alert. The StructuredRiskHead produces these components from variable-indexed temporal states without post-hoc attribution.

Method	${\geq}$ .90	${\geq}$ .93	${\geq}$ .95	Method	${\geq}$ .90	${\geq}$ .93	${\geq}$ .95
SepsisLens	.068	.083	.100	Raindrop	.087	.098	.112
GRU-D	.072	.084	.102	GRU	.086	.102	.113
SepsisCalc	.085	.096	.109	Transf.	.092	.108	.122
LSTM	.082	.096	.110	RETAIN	.196	.223	.250

Table 3: Mean FAR over three seeds at fixed event-recall targets on MIMIC-IV. Lower is better.

Variant	AUC	APC	oAUC
SepsisLens	$.939{\pm}.004$	$\mathbf{.474{\pm}.019}$	$\mathbf{.940{\pm}.013}$
$-$ OrgSupv	$.940{\pm}.002$	$.448{\pm}.005$	$.567{\pm}.015$
ICUOnly	$.909{\pm}.002$	$.452{\pm}.001$	$.899{\pm}.003$
PlainHead	$.935{\pm}.001$	$.450{\pm}.008$	—
EarlyFusion	$.944{\pm}.001$	$.464{\pm}.017$	—

Table 4: MIMIC-IV structural ablations at a fixed threshold of 0.5. oAUC denotes organ macro-AUROC.

Masking	Metric	$k=1$	$k=3$	$k=5$
Top- $k$	$\Delta p$	$.066{\pm}.006$	$.176{\pm}.030$	$.252{\pm}.034$
	Flip	$.593{\pm}.028$	$.884{\pm}.037$	$.956{\pm}.021$
Random- $k$	$\Delta p$	$.029{\pm}.006$	$.095{\pm}.014$	$.195{\pm}.018$
	Flip	$.364{\pm}.030$	$.630{\pm}.045$	$.843{\pm}.046$
Bottom- $k$	$\Delta p$	$.007{\pm}.014$	$.021{\pm}.023$	$.037{\pm}.028$
	Flip	$.180{\pm}.115$	$.304{\pm}.154$	$.407{\pm}.180$

Table 5: Input-side variable masking on MIMIC-IV. Flip uses the validation-selected threshold.

Explore similar work

Clinically Interpretable Sepsis Early Warning via LLM-Guided Simulation of Temporal Physiological Dynamics
Apr 22, 2026Weizhi Nie, Zhen Qu, Weijie Wang +4Sepsis Early Warning
School of Electrical and Information Engineering, Tianjin University, Tianjin, 300072, Tianjin, China · Department of Information Engineering and Computer Science, University of Trento, Trento, Trentino, Italy · Department of Orthopedics, Affiliated Kunshan Hospital of Jiangsu University, Kunshan, Jiangsu, China +3
A Framework for Early Sepsis Prediction via Self-Supervised (JEPA) and Federated Representation Learning
Jul 18, 2026Umair bin Mansoor, Munaf Rashid, Roomi NaqviSepsis Biomarker
Electrical Engg. Department, DHA Suffa University, Karachi, Pakistan · Salim Habib University, Karachi, Sindh, Pakistan · Intel Foundry Services, Intel, Tempe, Arizona, United States
CSRA: Controlled Spectral Residual Augmentation for Robust Sepsis Prediction
Apr 16, 2026Honglin Guo, Rihao Chang, He Jiao +3Sepsis Intensive Care Unit
Tianjin University, Tianjin 300072, China · Department of Emergency Medicine, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou 310016, China · Tianjin Medical University General Hospital, Tianjin 300052, China

Apr 22, 2026, cs.LGJ/K move · Enter open · S save

Clinically Interpretable Sepsis Early Warning via LLM-Guided Simulation of Temporal Physiological Dynamics

Weizhi Nie, Zhen Qu, Weijie Wang, Chunpei Li+3

School of Electrical and Information Engineering, Tianjin University, Tianjin, 300072, Tianjin, China · Department of Information Engineering and Computer Science, University of Trento, Trento, Trentino, Italy · Department of Orthopedics, Affiliated Kunshan Hospital of Jiangsu University, Kunshan, Jiangsu, China+3

Sepsis Early Warning Physiological Data Clinical

Open paper AskPDF arXiv

Timely and interpretable early warning of sepsis remains a major clinical challenge due to the complex temporal dynamics of physiological deterioration. Traditional data-driven models often provide accurate yet opaque predictions, limiting physicians' confidence and clinical applicability. To address this limitation, we propose a Large Language Model (LLM)-guided temporal simulation framework that explicitly models physiological trajectories prior to disease onset for clinically interpretable prediction. The framework consists of a spatiotemporal feature extraction module that captures dynamic dependencies among multivariate vital signs, a Medical Prompt-as-Prefix module that embeds clinical reasoning cues into LLMs, and an agent-based post-processing component that constrains predictions within physiologically plausible ranges. By first simulating the evolution of key physiological indicators and then classifying sepsis onset, our model offers transparent prediction mechanisms that align with clinical judgment. Evaluated on the MIMIC-IV and eICU databases, the proposed method achieves superior AUC scores (0.861-0.903) across 24-4-hour pre-onset prediction tasks, outperforming conventional deep learning and rule-based approaches. More importantly, it provides interpretable trajectories and risk trends that can assist clinicians in early intervention and personalized decision-making in intensive care environments.

Abstract

Explore similar work

Clinically Interpretable Sepsis Early Warning via LLM-Guided Simulation of Temporal Physiological Dynamics

A Framework for Early Sepsis Prediction via Self-Supervised (JEPA) and Federated Representation Learning

CSRA: Controlled Spectral Residual Augmentation for Robust Sepsis Prediction