Organizations: Institute of Data Science and Engineering, National Yang Ming Chiao Tung University, Hsinchu, Taiwan · EECS International Graduate Program, National Yang Ming Chiao Tung University, Hsinchu, Taiwan · College of Intelligent Computing and Medical Statistics Research Center, Chang Gung University, Taoyuan, Taiwan · Department of Emergency Medicine, Chang Gung Memorial Hospital, Keelung, Taiwan · Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan · Computational Health Informatics Program, Boston Children’s Hospital, Boston, MA, USA
Sepsis is a heterogeneous syndrome. Identifying clinically distinct phenotypes may enable more precise treatment strategies. In recent years, many researchers have applied clustering algorithms to sepsis patients. However, many existing approaches require aggregating or imputing temporal Electronic Health Records (EHRs), which can limit the model's ability to capture disease progression. Furthermore, existing clustering processes seldom incorporate domain-driven constraints, often resulting in phenotypes that lack clear clinical distinction. We propose a novel clustering network, NPCNet, that comprises a text embedding generator, a clustering operator, and a target navigator. We first transformed EHRs into pseudo texts by discretizing continuous clinical measurements and integrated them with static variables to construct the embeddings. The target navigator then infused clinical knowledge into the embeddings through auxiliary tasks, constraining clustering results to better align sepsis phenotypes with clinical significance. Finally, the clustering operator employed an iterative refinement mechanism to jointly optimize phenotype centroids and patient representations under domain-driven constraints. Extensive experiments on public datasets demonstrated that NPCNet achieved superior performance on both internal clustering and clinical metrics. We identify four distinct sepsis phenotypes with divergent SOFA trajectories. Notably, although α and δ exhibit similarly severe conditions initially, NPCNet distinguishes patients likely to improve (α) from those at risk of deterioration (δ). Furthermore, treatment effect analysis suggests that early vasopressor administration may be associated with improved outcomes in the α, {beta}, and δ phenotypes. These findings highlight NPCNet's potential for phenotype-specific treatment strategies.
Figures & tables
Figure 1: The overview of NPCNet. Through the text embedding generator, we first bin the value of time-varying variables into bin indices according to the distribution of the training set. We then transform the series of time-varying variables into pseudo texts, combining static information. NPCNet trains with an objective function that consists of Lrec , Lclustering , and Lnavigator . Through the clustering network in the clustering operator and clinical relevance through the target navigator, we can derive the patient embeddings and the centroids of computable phenotypes during the training stage.
Figure 2: Illustration of the binning process for converting time-varying EHR into a pseudo-text sequence. Numerical observations are discretized into quantile-based bins, encoded using the [VARIABLE][BIN] format, and arranged chronologically to generate the final pseudo text. For example, a systolic blood pressure (SBP) value of 32 falls into the lowest quantile bin (SBP-1) and is encoded as "SBP SBP-1".
Figure 3: An example of the NPCNet input representation. Temporal clinical variables are converted into pseudo text and mapped to token-level embeddings, which are combined with order encoding to preserve the examination order. Static variables are encoded separately and fused with the temporal representation using a weighted sum to form the final input for NPCNet.
Figure 4: Abnormal clinical variables, grouped into eight organ systems [ 24 ] , among the sepsis computable phenotypes. The ribbon connects from a phenotype to an organ system if the group median is more abnormal than the overall median ( Supplementary Table 2 ). The more clinical variables are abnormal for the phenotype, the broader the ribbon.
Characteristic
Total
\upalpha
\upbeta
\upgamma
\updelta
No. of patients (%)
19834 (100)
4624 (23.3)
6246 (31.5)
6250 (31.5)
2714 (13.7)
Age, mean (SD)
64.7 (15.8)
63.3 (14.2)
63.6 (16.2)
66.2 (16.1)
66.1 (16.3)
Sex, No. (%)
Male
11763 (59.3)
3122 (67.5)
3535 (56.6)
3513 (56.2)
1593 (58.7)
Female
8071 (40.7)
1502 (32.5)
2711 (43.4)
2737 (43.8)
1121 (41.3)
SOFA score, median (IQR)
7 [4, 10]
7 [3, 9]
5 [3, 8]
7 [5, 10]
11 [8, 14]
Table 1: Characteristics of the 4 sepsis computable phenotypes.
Model
SI ↑
CHI ↑
DBI ↓
TDI ↑
consensus K-means [ 34 ]
-0.043 (0.011)***
0.040 (0.005)***
5.426 (0.870)***
0.716 (0.073)
KM-DTW
0.000 (0.000)***
0.335 (0.004)***
1.993 (0.011)***
0.476 (0.093)***
DCN [ 38 ]
0.344 (0.048) ***
1.482 (0.420) **
1.140 (0.083) ***
0.480 (0.080)***
DKM [ 2 ]
0.290 (0.053)***
0.423 (0.029)***
1.690 (0.098)***
0.463 (0.065)***
USNID [ 41 ]
0.711 (0.413)
4.415 (5.146)
0.830 (0.628)
0.213 (0.197)***
naviDCN [ 29 ]
0.335 (0.076)***
0.824 (0.285)***
1.184 (0.129)***
0.533 (0.138) ***
Table 2: The performance of NPCNet and benchmarks in the testing set
Figure 5: Multivariable logistic regression on in-hospital mortality with phenotypes.
Figure 6: SOFA trajectories during the 18 hours following phenotype derivation by NPCNet, stratified by the SOFA score at six hours after ICU admission.
Figure 7: Pairwise comparisons of SOFA trajectories between four phenotypes at each hour within each stratum for NPCNet and benchmarks. The cell in the heatmap represents the number of significantly different phenotype pairs, out of six total possible pairwise combinations, at each time point within each SOFA stratum. The values range from 0 to 6, with darker colors indicating more pairwise differences, thereby reflecting how distinguishable the phenotypes are over time. The Trajectory Divergence Index (TDI) was then derived by normalizing the number of statistically significant pairs by the total number of pairwise comparisons, yielding a metric between 0 and 1 that quantifies the overall performance of models in clinical significance.
Feature modalities
SI ↑
CHI ↑
DBI ↓
TDI ↑
table w/ MLP
0.134 (0.030)***
0.298 (0.020)***
2.180 (0.304)***
0.595 (0.063) ***
matrix w/ CNN
0.428 (0.040) **
1.441 (0.478) **
0.766 (0.102) *
0.587 (0.128)**
NPCNet
0.447 (0.012)
2.051 (0.161)
0.670 (0.022)
0.753 (0.061)
Table 3: Ablation of the feature modalities and corresponding backbones in the testing set. The table uses the most abnormal measurements with imputation and an MLP backbone; temporal matrices aggregate measurements into hourly windows using ADL [ 14 ] with a CNN backbone; pseudo text preserves the original temporal resolution without aggregation or imputation and uses a Transformer backbone (NPCNet).
Text
Positional encoding
SI ↑
CHI ↑
DBI ↓
TDI ↑
value
timeline
0.285 (0.051)***
0.715 (0.364)***
1.167 (0.167)***
0.481 (0.072)***
value
order
0.450 (0.027)
2.135 (0.164)
0.682 (0.024)
0.576 (0.056)***
bin
timeline
0.348 (0.053) ***
0.862 (0.245) **
1.017 (0.083) ***
0.681 (0.063) *
bin
order
0.447 (0.012)
2.051 (0.161)
0.670 (0.022)
0.753 (0.061)
Table 4: Ablation of the text representation by combining two value encodings (raw vs. discretized) and two positional encodings (timeline vs. examination order) in the testing set.
Lprob
Ldist
SI ↑
CHI ↑
DBI ↓
TDI ↑
0.326 (0.012) ***
0.817 (0.104) ***
0.864 (0.017) ***
0.268 (0.062)***
V
0.278 (0.022)***
0.512 (0.053)***
1.262 (0.100)***
0.600 (0.029)***
V
0.211 (0.034)***
0.319 (0.055)***
1.910 (0.317)***
0.633 (0.055) ***
V
V
0.447 (0.012)
2.051 (0.161)
0.670 (0.022)
0.753 (0.061)
Table 5: Ablation of the objective functions in the testing set. Each navigator loss is removed individually while retaining reconstruction and clustering losses to prevent embedding collapse.
Figure 8: SOFA trajectories across phenotypes under different target navigators: No Navigator, ICU LOS Navigator, and Dual Navigator. In the Dual Navigator, ICU LOS guides Lprob , while discharge status guides Ldist . In the primary NPCNet, discharge status is used as the navigator. Pairwise comparisons of SOFA trajectories between phenotypes at each hour within each stratum in the testing set. The darker the color is, the greater the number of pairwise phenotypes that have statistically significant differences.
Early sepsis prediction from electronic health records is challenged by irregular sampling, high missingness, and class imbalance. We systematically compare four modeling paradigms -- self-supervised Joint Embedding Predictive Architecture (JEPA) via masked latent prediction, self-supervised VICReg (variance-invariance-covariance regularization) with two-view augmentation, semi-supervised fine-tuning of a VICReg-pretrained encoder, and supervised Temporal Convolutional Network (TCN) -- alongside raw-feature baselines. All models share a common preprocessing pipeline of hourly binning with forward-fill imputation applied to 7 biomarkers selected via sparsity analysis from the MIMIC-III dataset. Our best model (JEPA + XGBoost + mean pooling) achieves AUPRC 0.636 at the time of onset (H0), approaching the SupMix benchmark (0.667) while using 83% fewer biomarkers. The Tier 1 pipeline -- VICReg pretraining followed by semi-supervised fine-tuning and XGBoost -- achieves AUPRC 0.510 at H0, a 3.1× improvement over the raw-feature baseline (0.165) and a 7.6% improvement over the end-to-end supervised TCN (0.474). Crucially, the fine-tuned VICReg encoder exhibits the most temporally persistent representations, degrading only 16.8% from H0 to H10 compared to 47.5% for supervised TCN and 65.3% for JEPA, demonstrating that self-supervised pretraining with task-aware fine-tuning yields features that are both sharp near onset and robust across prediction horizons.
Umair bin Mansoor, Munaf Rashid, Roomi Naqvi
Electrical Engg. Department, DHA Suffa University, Karachi, Pakistan · Salim Habib University, Karachi, Sindh, Pakistan · Intel Foundry Services, Intel, Tempe, Arizona, United States
The early prediction of sepsis in intensive care unit (ICU) patients is crucial for improving survival rates. However, the development of accurate predictive models is hampered by data fragmentation across healthcare institutions and the complex, temporal nature of medical data, all under stringent privacy constraints. To address these challenges, we propose a novel framework that uniquely integrates federated learning (FL) with a medical knowledge graph and a temporal transformer model, enhanced by meta-learning capabilities. Our approach enables collaborative model training across multiple hospitals without sharing raw patient data, thereby preserving privacy. The model leverages a knowledge graph to incorporate structured medical relationships and employs a temporal transformer to capture long-range dependencies in clinical time-series data. A model-agnostic meta-learning (MAML) strategy is further incorporated to facilitate rapid adaptation of the global model to local data distributions. Evaluated on the MIMIC-IV and eICU datasets, our method achieves an area under the curve (AUC) of 0.956, which represents a 22.4% improvement over conventional centralized models and a 12.7% improvement over standard federated learning, demonstrating strong predictive capability for sepsis. This work presents a reliable and privacy-preserving solution for multi-center collaborative early warning of sepsis.
Yue Chang, Guangsen Lin, Jyun Jie Chuang +3
Chengdu Medical College, Chengdu, Sichuan 610500, China · Kunming Medical University, Kunming, Yunnan 650500, China · National Yang Ming Chiao Tung University, Taipei, Taiwan 900, China +2
Offline sepsis treatment aims to learn treatment policies from retrospective ICU trajectories. Decision Transformer-based methods provide a natural sequence modeling framework for this task, but they mainly rely on raw physiological states and historical clinician actions. Such raw trajectory tokens may not explicitly capture clinical semantics that clinicians use in practice, such as patient severity, recent treatment response, and current risk priorities. To address this limitation, we propose ClinSDT, an LLM-assisted Decision Transformer framework for offline sepsis treatment. ClinSDT uses a frozen LLM as a semantic bridge rather than a treatment decision maker. It constructs three clinical semantic views from ICU trajectories, namely Profile, Context, and Advisory, and inserts their encoded representations as independent tokens into a Decision Transformer policy. To support policy refinement and rollout, ClinSDT further trains a Treatment-Response Forecaster to estimate the next physiological state, SAPS-II change, and predictive uncertainty from state-action inputs. In the second training stage, the forecaster screens local counterfactual treatment candidates, and only low-uncertainty candidates with lower predicted SAPS-II increase are distilled into the policy after semantic regeneration. During inference, ClinSDT performs closed-loop treatment rollout by refreshing semantic inputs from the currently available patient state. Experiments on MIMIC-III and MIMIC-IV show that ClinSDT achieves consistent relative improvements over representative imitation learning, offline reinforcement learning, and Decision Transformer baselines under the same rollout protocol. Code and data are available at https://github.com/KAG778/ClinSDT.
Yanyan Wu, Guanyu Zhu, Changbin Zhao +7
South China Agricultural University, Guangzhou, China · The Ohio State University, Columbus, Ohio, United States