Sepsis is a life-threatening syndrome characterized by dysregulated host responses to infection and remains a major global health challenge (1). Earlier global estimates suggested that sepsis affects approximately 31 million people annually and contributes to nearly 5 million deaths worldwide, underscoring its profound impact on population health (2). With improvements in critical care, a growing number of patients survive the acute phase of sepsis; however, many survivors experience long-term complications that substantially affect their quality of life (3). Among these sequelae, cognitive impairment has emerged as one of the most prevalent and debilitating outcomes. Studies have reported that a considerable proportion of sepsis survivors develop persistent deficits in memory, attention, executive function, or information processing, which may persist for months or even years after discharge (4, 5). These impairments not only hinder functional recovery and return to daily activities but also increase healthcare utilization, dependency, and long-term mortality. Despite its clinical significance, early recognition of cognitive impairment remains challenging because symptoms are often subtle and obscured by delirium, sedation, and other features of critical illness.

The development of cognitive impairment following sepsis is multifactorial. Proposed mechanisms include sustained systemic inflammation, blood-brain barrier disruption, microglial activation, neuronal injury induced by oxidative stress, and altered cerebral perfusion (6–8). At the clinical level, diverse factors such as disease severity, organ dysfunction patterns, metabolic disturbances, sedative exposure, and hemodynamic instability have all been implicated (5, 9, 10). However, the interaction among these variables is highly complex and non-linear, and traditional regression-based methods often fail to capture such intricate relationships. As a consequence, early and reliable prediction of cognitive impairment in sepsis is still limited, narrowing the potential to implement timely interventions and focused neuroprotective strategies.

Machine learning (ML) offers a powerful approach to address these limitations and has been applied to cognitive impairment following sepsis, both in clinical studies (11, 12) and animal models (13, 14). By leveraging large-scale, multidimensional clinical data, ML algorithms are capable of modeling non-linear interactions and improving predictive accuracy beyond conventional statistical techniques (15). Nevertheless, many ML models function as “black boxes,” producing predictions without revealing how input variables contribute to the outcome (16). In high-stakes clinical decision-making, this opacity raises concerns regarding accountability and interpretability, reducing clinicians’ willingness to adopt ML tools in practice. The emergence of interpretable ML provides a solution to this challenge (17, 18). Techniques such as SHapley additive exPlanations (SHAP) and model-agnostic interpretation frameworks allow visualization of feature contributions at both population and individual levels, enhancing transparency, reliability, and user acceptance (19). Applying interpretable ML to the prediction of cognitive impairment in sepsis may therefore improve early risk stratification, uncover clinically meaningful risk profiles, and support targeted strategies to prevent long-term neurological deterioration.

In this study, we aimed to develop and validate an interpretable ML model to predict the occurrence of cognitive impairment among patients with sepsis. By combining routinely available clinical variables with advanced interpretable ML techniques, we sought not only to optimize predictive performance but also to clarify the relative contribution of critical risk factors. This interpretable framework may assist clinicians in recognizing high-risk patients at an earlier stage and facilitate the implementation of strategies to mitigate long-term cognitive deterioration.

In this retrospective cohort study, we developed an interpretable ML framework to predict cognitive impairment among sepsis survivors and evaluated its performance using multiple complementary validation methods. The study identified seven key predictors using a combined LASSO-Boruta feature selection strategy, including age, years of education, septic shock, benzodiazepine use, APACHE II score, SOFA score, and IL-10. Among the five ML models constructed, the RF model achieved the most favorable and stable performance, with AUCs of 0.947 in the training set and 0.895 in the validation set. Calibration curves, DCA, and confusion matrices consistently confirmed the model’s stability, good clinical applicability, and balanced classification performance. SHAP-based interpretability analysis further clarified the global and individual contributions of the selected predictors. Among them, SOFA score emerged as the strongest contributor, followed by age, APACHE II score, IL-10, and years of education. These findings highlight the potential of an interpretable RF model to enhance early identification of sepsis survivors at high risk for cognitive impairment.

Although cognitive impairment is increasingly recognized as a major complication of sepsis, predictive studies remain scarce. Among the limited studies that have attempted to predict cognitive outcomes after sepsis, Zhu et al. evaluated three serum biomarkers and reported that their combined panel achieved an AUC of 0.887 in predicting persistent cognitive impairment in a cohort of 114 sepsis patients (23). Although the study offers initial insights into potential markers associated with cognitive outcomes after sepsis, its reliance on specialized biomarker assays, small sample size, and the absence of ML guided feature evaluation or interpretability limits its scalability and clinical practicality. In contrast, most other predictive efforts in the sepsis literature have focused on acute brain dysfunction rather than post-discharge cognitive outcomes. Jin et al. used MIMIC-IV data to develop a LASSO-based nomogram for predicting sepsis-associated encephalopathy (SAE), showing moderate discrimination. However, their model targets acute encephalopathy, not cognitive impairment (24). Similarly, Zhao et al. constructed a logistic-regression nomogram to predict acute SAE in elderly ICU patients, achieving an AUC of approximately 0.80, but their work likewise centers on in-hospital encephalopathy (25). Compared with these prior studies, our work provides substantial methodological and clinical advancements by focusing specifically on cognitive impairment, a clinically meaningful outcome that has been largely overlooked in previous predictive models. By leveraging routinely available clinical variables within a rigorous LASSO-Boruta framework and incorporating SHAP-based interpretability to characterize non-linear and individualized risk patterns, our model uses a RF algorithm that achieved excellent discriminative performance, with an AUC of 0.947 in the training cohort and 0.895 in the validation cohort, and offers a robust approach to anticipating cognitive impairment after sepsis.

The importance of the seven retained predictors is supported by existing biological and clinical evidence. Age is one of the strongest predictors of post-sepsis cognitive impairment. Older adults exhibit reduced cognitive reserve, diminished neuroplasticity, and greater vulnerability to inflammatory and hypoperfusion-related brain injury. Sepsis triggers systemic inflammation, microglial activation, oxidative stress, and blood-brain barrier dysfunction, all of which disproportionately impair aging neural circuits (26, 27). Overton et al. reported in a large population-based cohort that the prevalence of mild cognitive impairment rises substantially with advancing age, with older adults showing markedly higher rates of both mild and severe cognitive deficits across all cognitive impairment subtypes (28). Our findings similarly demonstrate that higher age is a key vulnerability factor for post-sepsis cognitive impairment. Individuals with higher educational attainment are believed to have greater cognitive reserve, which enables more efficient neural networks and resilience against brain injury from inflammation or hypoperfusion (29, 30). Concha-Cisternas et al. analyzed more than 2,000 Chilean adults aged ≥ 60 years and found that individuals with lower educational attainment had a markedly higher likelihood of cognitive impairment, with the risk increasing up to 4.53-fold in those with ≤ 8 years of education compared with those with higher education (31). This strong inverse association between education and cognitive performance supports the cognitive-reserve hypothesis and aligns with our finding that fewer years of education increase vulnerability to post-sepsis cognitive impairment. Patients who develop septic shock experience more pronounced circulatory and microvascular disturbances, which increase vulnerability to secondary brain injury through impaired cerebral perfusion and amplified inflammation (32). An MRI study by Götz et al. demonstrated that survivors of severe sepsis or septic shock exhibit structural alterations in key cognitive regions, including the hippocampus and amygdala (33). This evidence supports the strong association between shock physiology and cognitive impairment and aligns with our finding that septic shock significantly heightens the risk of post-sepsis cognitive decline. Benzodiazepine exposure has been reported to be associated with long-term cognitive outcomes in critically ill patients, often reflecting underlying acute brain dysfunction rather than a direct causal effect (34). Pandharipande et al. showed that lorazepam use independently increased the daily risk of transitioning into delirium in ICU patients (35). Stewart reviewed evidence from multiple peer-reviewed studies and reported that long-term benzodiazepine use is associated with impairments across several cognitive domains, with deficits that may persist even after discontinuation (36). The association between benzodiazepine use and post-sepsis cognitive impairment should be interpreted with caution. In critically ill patients, benzodiazepines are often administered in response to agitation, delirium, or severe physiological instability. As such, benzodiazepine exposure is subject to confounding by indication and may primarily reflect underlying acute brain dysfunction or illness severity rather than serving as an independent causal factor of long-term cognitive impairment. From a clinical perspective, benzodiazepine use in this context may act as a surrogate marker identifying patients who experienced more severe neurophysiological stress during the acute phase of sepsis. This interpretation is consistent with prior evidence linking sedative exposure, delirium, and subsequent cognitive decline. Therefore, our findings should not be interpreted as implying a direct neurotoxic effect of benzodiazepines per se, but rather as highlighting a high-risk clinical phenotype characterized by acute neurological vulnerability. Higher APACHE II and SOFA scores reflect more severe physiological derangement and multi-organ dysfunction during sepsis, factors that have been consistently linked to worse long-term neurological outcomes. Severe systemic inflammation, impaired perfusion, and metabolic instability captured by these scoring systems can exacerbate neuronal injury and hinder cognitive recovery (37). Iwashyna et al. reported that survivors of severe sepsis experienced a more than threefold increase in moderate-to-severe cognitive impairment compared with their presepsis baseline (4). A recent systematic review further showed that severe sepsis was associated with a significantly higher risk of dementia or cognitive impairment than milder forms of sepsis (10). These findings reinforce the concept that greater illness severity and organ failure burden markedly increase vulnerability to post-sepsis cognitive impairment, consistent with our observation that higher APACHE II and SOFA scores are key predictors of poor cognitive outcomes. It is important to interpret the contribution of severity-related scores within the context of a survivor-based cohort. In sepsis, patients with extremely high SOFA or APACHE II scores are at substantially increased risk of early mortality and are therefore less likely to survive to undergo post-discharge cognitive evaluation. As a result, the analyzed population represents a subset of sepsis survivors in whom the distribution of disease severity may be partially truncated. Within this framework, the strong influence of SOFA and APACHE II scores observed in our model suggests that, even among survivors, residual organ dysfunction and acute physiological derangement remain key determinants of subsequent cognitive impairment. Rather than reflecting mortality risk alone, these severity scores may capture the cumulative burden of systemic inflammation, hypoperfusion, and multi-organ dysfunction that predispose surviving patients to long-term neurocognitive sequelae. This distinction is clinically relevant, as it highlights that illness severity retains prognostic value beyond survival and may aid in identifying high-risk survivors who could benefit from early cognitive monitoring and intervention. IL-10 also emerged as a significant predictor in our model. Although IL-10 can exert anti-inflammatory or even neuroprotective effects in experimental models of acute neuroinflammation, elevated circulating IL-10 levels in sepsis primarily reflect severe immune dysregulation rather than direct protection (38, 39). Clinical studies have consistently shown that higher IL-10 concentrations correlate with greater organ failure, secondary infections, and increased mortality, indicating a state of sepsis-induced immunosuppression and more severe physiological derangement (40, 41). In this context, IL-10 likely serves as a biomarker of heightened systemic inflammation and immune paralysis—conditions known to amplify neuroinflammatory responses and hinder cognitive recovery. This interpretation aligns with our findings that higher IL-10 levels were associated with an increased risk of post-sepsis cognitive impairment.

Several methodological features of our study strengthen both the robustness of the model and its potential for real-world clinical application. By integrating a two-stage LASSO-Boruta feature selection strategy, we combined the strengths of penalized regression and iterative feature confirmation to minimize redundancy, reduce multicollinearity, and ensure the stability and clinical relevance of the retained predictors. The parallel evaluation of multiple ML algorithms further reduced model-specific bias, allowing an unbiased selection of the most reliable approach. Importantly, embedding SHAP-based interpretability addresses one of the major barriers to clinical adoption of ML—the opacity of prediction mechanisms—by providing transparent, patient-level explanations of how individual variables contribute to risk estimates. The methodological strengths of our model also translate into clinically actionable implications. Although septic shock, APACHE II, and SOFA scores primarily reflect underlying illness severity rather than directly adjustable interventions, the strong association of early physiological derangement with long-term cognitive impairment suggests that meticulous early resuscitation and organ-support strategies may indirectly contribute to neuroprotection. Together, these findings highlight several points in the acute-management trajectory where optimization may help mitigate the risk of post-sepsis cognitive decline. Although k-fold cross-validation is commonly used to enhance robustness in predictive modeling, a fixed training-validation split was intentionally adopted in this study. This design allowed us to retain an independent validation cohort for assessing discrimination, calibration, and clinical utility, as well as for illustrating SHAP-based interpretability on unseen data. Given that the primary objective of this work was to develop a clinically interpretable prediction framework rather than to optimize performance alone, preserving a hold-out validation set was considered appropriate. Nevertheless, we acknowledge that k-fold cross-validation may further improve generalizability, particularly in single-center settings. Future studies incorporating multicenter cohorts and cross-validation–based validation strategies are warranted to further evaluate model robustness and transportability. In addition to interpretability and predictive performance, the clinical applicability of ML models also depends on their generalizability across settings. Although the present model demonstrated strong discriminative performance and stability under internal validation, several issues related to generalizability should be considered. This study was conducted at a single center, and local clinical practices, patient characteristics, and sedation strategies may influence model behavior. Consequently, the current findings should be interpreted as reflecting performance within a similar clinical context rather than as universally generalizable estimates. Nevertheless, the use of conservative feature selection, independent validation splitting, and multiple complementary performance assessments was intended to reduce overfitting and enhance robustness. External validation using multicenter cohorts with standardized cognitive assessments will be essential to further evaluate model transportability and is a priority for future investigation. Several limitations should be acknowledged. First, this study included only sepsis survivors who were able to complete post-discharge cognitive assessments, which may introduce survivorship bias. Patients with extremely severe illness and early mortality were underrepresented, potentially limiting the generalizability of the findings to the entire sepsis population. Future multicenter studies are needed to validate the model across broader severity spectra. Second, this was a single-center retrospective study with internal validation only. Although multiple strategies were applied to reduce overfitting, including conservative feature selection and independent validation splitting, the absence of external validation limits the generalizability of the proposed model. Validation in multicenter cohorts with diverse clinical practices and patient populations is therefore warranted. Third, although MoCA is widely used for cognitive screening, it may not capture subtle deficits in specific domains such as processing speed or executive function. Incorporating comprehensive neuropsychological batteries or digital cognitive tools in future studies would yield more granular insights. Fourth, the follow-up duration in this study was limited to 3 months after hospital discharge, which may not fully capture the trajectory of post-sepsis cognitive impairment. Fifth, benzodiazepine use was treated as a binary variable due to limitations in the availability of detailed dosing and duration data. This simplified representation may not fully capture dose-dependent or time-dependent effects, and residual confounding by indication cannot be completely excluded. Finally, although a broad spectrum of clinical variables was incorporated, residual confounding cannot be completely excluded, particularly with regard to premorbid cognitive reserve, socioeconomic factors, and variations in sedation or hemodynamic management.

In summary, our study demonstrates that age, years of education, septic shock, benzodiazepine exposure, APACHE II and SOFA scores, and IL-10 levels are key features contributing to the model’s prediction of post-sepsis cognitive impairment. Among the ML models evaluated, the RF algorithm demonstrated the most robust and stable predictive performance. By integrating these variables into an interpretable RF framework, we were able to quantify the relative importance of each factor and enable fine-grained, individualized risk prediction to support personalized clinical management and early identification of high-risk patients. While the follow-up period was limited and external validation is needed, this model offers a promising tool for individualized risk stratification and may inform targeted strategies to mitigate neurocognitive sequelae in sepsis survivors. Future studies with extended follow-up and multicenter cohorts are warranted to refine predictive accuracy and elucidate the mechanisms underlying persistent cognitive impairment.