NPCNet: Navigator-Driven Pseudo Text for Deep Clustering of Early Sepsis Phenotyping

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, the clustering process rarely incorporates clinical relevance, potentially limiting to reflect clinically distinct phenotypes. We propose NPCNet, a novel deep clustering network with a target navigator that integrates temporal Electronic Health Records (EHRs) to better align sepsis phenotypes with clinical significance. We identify four sepsis phenotypes ($α$, $β$, $γ$, and $δ$) with divergence in SOFA trajectories. Notably, while $α$ and $δ$ phenotypes both show severe conditions in the early stage, NPCNet effectively differentiates patients who are likely to improve ($α$) from those at risk of deterioration ($δ$). Furthermore, through the treatment effect analysis, we discover that $α$, $β$, and $δ$ phenotypes may benefit from early vasopressor administration. The results show that NPCNet enhances precision treatment strategies by uncovering clinically distinct phenotypes.

💡 Research Summary

Sepsis remains a leading cause of mortality worldwide, with heterogeneous clinical courses that challenge one‑size‑fits‑all treatment strategies. Existing phenotyping efforts have largely relied on static variables or aggregated time‑series features, often ignoring the rich temporal dynamics of electronic health records (EHR) and failing to align clusters with clinically meaningful outcomes. To address these gaps, the authors present NPCNet, a deep clustering framework that leverages a “target navigator” to embed clinical relevance directly into the representation learning process.

The method first discretizes each time‑varying laboratory or vital sign measurement into bins based on its distribution in the training set. These binned values are then ordered by timestamp and concatenated with static patient information (demographics, comorbidities) to form a pseudo‑text sequence. A BERT‑style transformer encoder converts this pseudo‑text into a dense embedding vector that preserves temporal ordering while handling irregular sampling and missingness without imputation.

Two auxiliary tasks constitute the target navigator: (a) prediction of in‑hospital mortality and (b) prediction of discharge status (alive vs. dead). Both tasks are trained jointly with a composite loss that combines cross‑entropy (probability) and L2 distance components, encouraging the embeddings to be informative for outcomes. The clustering operator then applies a K‑means‑like algorithm on these outcome‑aware embeddings, learning four cluster centroids that define the final phenotypes.

In the development cohort (MIMIC‑IV, N = 19,834), NPCNet outperformed four benchmark methods (standard K‑means, K‑means‑DTW, DCN, DKM, naviDCN) on three internal metrics: Silhouette Index (SI = 0.447), Calinski‑Harabasz (CHI = 2.051), and Davies‑Bouldin (DBI = 0.670), each averaged over ten random seeds. The model identified four phenotypes, labeled α, β, γ, and δ. Phenotype α comprised younger patients with few comorbidities, exhibited early multi‑organ abnormalities but showed rapid SOFA score decline and the lowest 30‑day mortality (1.4%). Phenotype δ, in contrast, consisted of older patients with multiple comorbidities, severe early organ dysfunction, rising SOFA trajectories, and the highest 30‑day mortality (46.2%). Phenotypes β and γ displayed intermediate patterns, with β dominated by inflammatory marker elevation and γ characterized by combined inflammation and renal dysfunction.

Treatment‑effect analysis using multivariable logistic regression revealed phenotype‑specific associations between time to vasopressor initiation and in‑hospital mortality. A one‑hour delay increased odds of death by 31.8 % for α, 15.6 % for β, and 16.0 % for δ, while no significant effect was observed for γ. Sensitivity analysis using E‑values confirmed that unmeasured confounders would need to exert implausibly large effects to nullify these associations. Intravenous fluid volume showed no consistent mortality impact across phenotypes.

Ablation studies demonstrated that (1) the pseudo‑text representation outperformed raw numeric time‑series with timeline encoding, (2) the transformer‑based embedding surpassed MLP and CNN autoencoders, (3) the combined probability‑distance loss yielded the best clustering metrics, and (4) the discharge‑status navigator contributed the strongest clinical differentiation.

External validation on the eICU database reproduced the four phenotypes, their organ‑system abnormality patterns, and divergent SOFA trajectories, supporting the generalizability of NPCNet. However, the vasopressor timing effect was not statistically significant in eICU, suggesting cohort‑specific practice variations.

The authors conclude that NPCNet successfully integrates temporal EHR information and outcome‑driven supervision to produce clinically actionable sepsis phenotypes. Limitations include sensitivity to binning choices, the need for prospective validation to establish causality, and computational considerations for real‑time deployment. Nonetheless, NPCNet represents a promising step toward precision medicine in critical care, offering a scalable pipeline that can be extended to other heterogeneous syndromes.