Machine learning consensus clustering for inflammatory subtype analysis in stroke and its impact on mortality risk: a study based on NHANES (1999–2018)

 Don't tell us what the mortality risk is; REDUCE IT! You blithering idiots can't think at all, can you?

Machine learning consensus clustering for inflammatory subtype analysis in stroke and its impact on mortality risk: a study based on NHANES (1999–2018)

Zhang Chunjuan^*Wang YulongZhou XichengMa Xiaodong^*

• Haiyan People’s Hospital, Jiaxing, Zhejiang, China

Background: Our study aims to utilize unsupervised machine learning methods to perform inflammation clustering on stroke patients via novel CBC-derived inflammatory indicators (NLR, PLR, NPAR, SII, SIRI, and AISI), evaluate the mortality risk among these different clusters and construct prognostic models to provide reference for clinical management.

Methods: A cross-sectional analysis was conducted using data from stroke participants in the U.S. NHANES 1999–2018. Weighted multivariate logistic regression was used to construct different models; consensus clustering methods were employed to subtype stroke patients based on inflammatory marker levels; LASSO regression analysis was used to construct an inflammatory risk score model to analyze the survival risks of different inflammatory subtypes; WQS regression, Cox regression, as well as XGBoost, random forest, and SVMRFE machine learning methods were used to screen hub markers which affected stroke prognosis; finally, a prognostic nomogram model based on hub inflammatory markers was constructed and evaluated using calibration and DCA curves.

Results: A total of 918 stroke patients with a median follow-up of 79 months and 369 deaths. Weighted multivariate logistic regression analysis revealed that high SIRI and NPAR levels were significantly positively correlated with increased all-cause mortality risk in stroke patients (p < 0.001), independent of potential confounders; Consensus clustering divided patients into two inflammatory subgroups via SIRI and NPAR, with subgroup 2 having significantly higher markers and mortality risks than subgroup 1 (p < 0.001); LASSO regression analysis showed subgroup 2 had higher risk scores and shorter overall survival than subgroup 1 [HR, 1.99 (1.61–2.45), p < 0.001]; WQS regression, Cox regression, and machine learning methods identified NPAR and SIRI as hub prognostic inflammatory markers; The nomogram prognostic model with NPAR and SIRI demonstrated the best net benefit for predicting 1, 3, 5 and 10-year overall survival in stroke patients.

Conclusion: This study shows NPAR and SIRI were key prognostic inflammatory markers and positively correlated with mortality risk (p < 0.001) for stroke patients. Patients would been divided into 2 inflammatory subtypes via them, with subtype 2 having higher values and mortality risks (p < 0.001). It suggests that enhanced monitoring and management for patients with high SIRI and NPAR levels to improve survival outcomes.

1 Introduction

Stroke, also termed cerebrovascular accident (CVA), is characterized by rapidly developing neurological deficits due to sudden cerebral blood flow disruption, leading to long-term disability or mortality. As a leading global cause of disability and death (1), stroke affects approximately 15 million individuals annually, with a substantial proportion experiencing recurrent events or persistent functional impairments. The inflammatory cascade plays a dual role in post-stroke pathophysiology, mediating both secondary neuronal injury and repair processes (2). While inflammation has emerged as a promising therapeutic target, the clinical benefits of systemic anti-inflammatory interventions remain controversial. Consequently, stratifying acute ischemic stroke (AIS) patients based on inflammatory heterogeneity may enhance pathophysiological understanding and enable tailored therapeutic modulation of neuroinflammation, thereby optimizing cerebral protection and functional recovery.

Complete blood count (CBC)-derived inflammatory indices—including neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), neutrophil percentage-to-albumin ratio (NPAR), systemic inflammation response index (SIRI), systemic immune-inflammation index (SII), and aggregate index of systemic inflammation (AISI)—provide integrative prognostic insights by quantifying interactions among platelets, neutrophils, and lymphocytes. These indices have demonstrated diagnostic and prognostic utility across multiple diseases. In AIS, compelling evidence highlights their clinical relevance: NLR independently predicts 90-day functional outcomes (3), while NLR, PLR, lymphocyte-to-monocyte ratio (LMR), and SIRI correlate with in-hospital mortality, length of stay (4), and 3-month disability rates (5). Notably, these CBC-based biomarkers reflect real-time inflammatory dynamics through routine hematological testing, offering practical advantages over conventional cytokine assays. Their integration into AIS subtyping frameworks may thus facilitate precision medicine by identifying high-risk phenotypes amenable to targeted immunomodulation.

Unsupervised machine learning (ML), a statistical approach that identifies latent patterns by analyzing the underlying structures of unlabeled data, has demonstrated unique value in medical research for classification and risk stratification (6). As a pivotal branch of unsupervised ML, consensus clustering enables precise phenotypic classification by iteratively validating multidimensional feature heterogeneity across patient populations, independent of outcome variables. This methodology has been successfully applied to elucidate disease mechanisms and optimize therapeutic strategies across various conditions (7–10). Notably, its potential is emerging in stroke research. For instance, Yang et al. (11) classified AIS patients into three molecular subtypes based on peripheral blood monocyte transcriptomic profiles, revealing that the high-inflammatory-response subtype exhibited a significantly elevated risk of hemorrhagic transformation, providing critical insights for timing immunomodulatory therapies. Similarly, Cui et al. (12) integrated clinical and neuroimaging data using unsupervised ML to identify a subgroup showing superior therapeutic responses to a “statin combined with repetitive transcranial magnetic stimulation” regimen. However, existing studies have yet to systematically explore the role of CBC (complete blood count)-derived inflammatory indices (e.g., SIRI, NPAR) in AIS subtyping. These indices not only dynamically reflect key inflammatory pathways, such as neutrophil–platelet interactions, but also offer practical advantages for point-of-care testing.

This study applies consensus clustering to resolve inflammatory heterogeneity in AIS, aiming to achieve dual objectives: (1) identifying pathophysiologically meaningful stroke inflammatory subtypes based on CBC-derived inflammatory indices, and (2) constructing interpretable ML prognostic models to quantify mortality risk disparities across subtypes. Ultimately, this work seeks to provide evidence-based guidance for anti-inflammatory therapy optimization and monitoring frequency stratification (e.g., dynamic inflammation monitoring in high-risk subtypes), thereby facilitating the clinical translation of precision medicine in AIS management.

More at link.