We conducted a retrospective clinical study using data from the Affiliated Hospital of Xuzhou Medical University. The study period spanned from July 2023 to July 2025. The primary aim was to investigate risk factors for cognitive impairment in patients with spontaneous intracerebral hemorrhage (sICH) and to construct a corresponding predictive nomogram model. The prediction time point was defined as the time of hospital discharge. Consequently, the model variables include both baseline characteristics available at admission and clinical complications (e.g., pulmonary infection) that developed during the hospital stay. The study protocol adhered to the ethical principles of the Declaration of Helsinki, with ethical approval number: XYFY2025-KL670-01.

All patients were diagnosed with intracerebral hemorrhage via CT scan within 24 h of symptom onset. sICH was defined according to the AHA/ASA 2022 guidelines (11). To ensure cohort homogeneity, we strictly included patients aged ≥18 years with a first-ever episode of primary spontaneous (non-traumatic) ICH. To minimize confounding factors, the following exclusion criteria were applied: (1) history of dementia or significant pre-existing cognitive deficits; (2) secondary ICH confirmed clinically or radiologically (e.g., vascular malformation, tumor); (3) inability to complete cognitive assessment due to severe clinical condition; (4) extremely poor initial prognosis severely affecting survival. Ultimately, 264 eligible patients were identified (Figure 1). The cohort was randomly divided into a training set (n = 198, 75%) for model development and a validation set (n = 66, 25%) for internal validation using a random number generator.

Data extraction was performed via the hospital’s electronic medical record system. We compiled a comprehensive dataset covering demographic information (age, sex, years of education), admission vital signs [body temperature, mean arterial pressure (MAP)], and baseline health indicators [e.g., body mass index (BMI) and length of hospital stay]. Clinical severity was assessed using the Glasgow Coma Scale (GCS) score and calculated hematoma volume. Baseline laboratory parameters were derived from the first peripheral venous blood samples collected immediately upon hospital admission (strictly within 24 h of symptom onset) to reflect the physiological state prior to in-hospital medical intervention. These parameters were obtained from the institution’s standardized emergency admission panels. Specifically, mean corpuscular hemoglobin concentration (MCHC) was extracted as a standard calculated component of the routine complete blood count (CBC). Gamma-glutamyl transferase (γ-GT) was extracted as part of the routine hepatic function profile, which is standardly ordered at admission to screen for comorbidities such as alcohol use disorder. The comprehensive panel also included hemoglobin, C-reactive protein (CRP), creatinine (Cr), albumin, potassium (K⁺), D-dimer, and random blood glucose (GLU). Inflammatory indices, namely the neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR), were calculated. Treatment-related variables included optimal intervention timing, surgical method (conservative management, craniotomy, or trepanation and drainage), and specific hemorrhage location (basal ganglia, cerebral cortex, or thalamus). Additionally, comorbidities (hypertension, diabetes, epilepsy, prior cerebral infarction) and complications (pulmonary infection, intracranial infection, need for tracheotomy) were recorded. The primary endpoint was the occurrence of cognitive impairment. This study is a retrospective analysis of data generated through the routine Stroke center quality control pathway of our institution. This clinical standard of care mandates that all survivors of cerebrovascular events (including sICH) undergo a comprehensive face-to-face assessment at 3 months post-discharge. During this visit, trained neurologists routinely administer the Montreal Cognitive Assessment (MoCA). We retrospectively extracted these pre-existing clinical data for the current analysis. Cognitive impairment was defined as a MoCA score <22.

Candidate predictor variables were pre-specified based on their established clinical relevance to post-stroke cognitive outcomes in the literature and their availability at the time of hospital discharge, aligning with the intended use of the model. These included demographic factors (age, sex, years of education), indices of initial injury severity (hematoma volume, Glasgow Coma Scale score), key laboratory markers (C-reactive protein, neutrophil-to-lymphocyte ratio), and clinically significant in-hospital complications (pulmonary infection). For baseline characteristics, normally distributed continuous variables are presented as mean ± standard deviation (SD) and compared using Student’s t-test; non-normally distributed variables are presented as median and interquartile range (IQR) and compared using the Mann–Whitney U test. Categorical data are summarized as frequencies and percentages, compared using the chi-square test or Fisher’s exact test. To mitigate the risk of overfitting associated with traditional univariate screening followed by multivariate regression, we employed the least absolute shrinkage and selection operator (LASSO) regression for variable selection within the training set. The optimal tuning parameter (lambda) was determined via 10-fold cross-validation based on the minimum criterion. Variables with non-zero coefficients retained by LASSO were then entered into a standard multivariate logistic regression model to obtain the final odds ratios and construct the nomogram. This approach balances model parsimony with predictive stability. Model performance was rigorously evaluated across three dimensions: discrimination, calibration, and clinical utility. Discrimination was quantified using the area under the receiver operating characteristic (ROC) curve (AUC). Calibration curves were plotted to visualize the agreement between predicted probabilities and observed outcomes. Finally, decision curve analysis (DCA) was employed to determine the clinical net benefit of using the nomogram across different threshold probabilities. To obtain a more robust and stable estimate of the final model’s performance and to mitigate the variability inherent in a single split, we further performed repeated 10-fold cross-validation exclusively on the training set using the fixed predictors identified by LASSO.

All variables included in the analysis were extracted from mandatory fields in the hospital’s structured electronic medical record system and the standardized 3-month follow-up protocol. Consequently, there were no missing data for any of the demographic, clinical, laboratory, or outcome variables analyzed in this cohort of 264 patients. A complete-case analysis was therefore inherently performed.

Statistical analyses were performed using R software (version 4.2) and SPSS (version 25.0). A two-sided p-value <0.05 was defined as statistically significant.

The final analysis included 264 patients diagnosed with spontaneous intracerebral hemorrhage (Figure 1). Within this cohort, 117 patients (44.3%) were identified as having cognitive impairment. Comparative analysis between the training set (n = 198) and validation set (n = 66) showed balanced distribution of baseline characteristics (Table 1). No statistically significant differences were observed between the two sets regarding demographics (e.g., age, sex, years of education) or clinical indicators (e.g., hematoma volume, GCS score) (p > 0.05), confirming the validity of the split. The prevalence of cognitive impairment was 47.0% in the training group and 36.4% in the validation group, with no significant difference (p = 0.174).

Initial screening via univariate logistic regression in the training set identified several factors significantly associated with cognitive decline (p < 0.05), including hemorrhage location, length of hospital stays, pulmonary infection, tracheotomy, years of education, GCS score, hematoma volume, and hemoglobin level (Table 2). To address potential multicollinearity among these variables and prevent overfitting, we employed LASSO regression for rigorous feature selection. This penalized shrinkage method identified five variables with non-zero coefficients as the most robust predictors: pulmonary infection, years of education, hematoma volume, GCS score, and hemoglobin (Figure 2). These five candidate variables were subsequently entered into a multivariate logistic regression model (Table 3). Ultimately, three variables remained as independent predictors of cognitive impairment: pulmonary infection (OR = 3.028, 95% CI 1.480–6.166, p = 0.002), years of education (OR = 0.868, 95% CI 0.809–0.931, p < 0.001), and hematoma volume (OR = 1.030, 95% CI 1.014–1.047, p < 0.001). The results are consistent with those obtained after conducting a univariate logistic regression analysis and including the significant variables in a multivariate logistic regression. The longer education demonstrated a protective effect, whereas pulmonary infection and larger hematoma volume were associated with an increased risk of cognitive impairment.

Based on the multivariate analysis results, we constructed a prognostic nomogram to assist in personalized risk assessment (Figure 3). By locating a patient’s specific years of education, hematoma volume, and pulmonary infection status on the corresponding variable axes, clinicians can calculate a total score corresponding to a specific probability of cognitive impairment occurrence. The nomogram’s predictive accuracy was validated via ROC analysis (Figure 4). The model achieved an AUC of 0.771 (95% CI: 0.706–0.837) in the training set and a higher AUC of 0.820 (95% CI: 0.726–0.914) in the validation set, indicating stable discriminative ability. In the training set, the AUC of the nomogram (0.771) was significantly higher than that of using pulmonary infection alone (AUC = 0.640, p = 0.005), hematoma volume alone (AUC = 0.675, p = 0.030), or years of education alone (AUC = 0.654, p = 0.012). In the validation set, the AUC of the nomogram (0.820) was also significantly higher than that of pulmonary infection (AUC = 0.607, p = 0.012), hematoma volume (AUC = 0.680, p = 0.063), and years of education (AUC = 0.790, p = 0.361). Calibration curves (Figure 5) demonstrated high consistency between the probabilities predicted by the nomogram and the observed frequencies of cognitive impairment in both datasets. In addition to visual calibration curves, the goodness-of-fit of the nomogram was formally assessed using the Hosmer–Lemeshow test. The test indicated no significant deviation between predicted and observed outcomes in either the training set (χ² = 7.24, p = 0.512) or the validation set (χ² = 5.89, p = 0.660), confirming excellent model calibration. The model’s overall discriminative ability, expressed as the concordance index (C-index), was identical to the AUC values (training set: 0.771; validation set: 0.820). Decision curve analysis (DCA) further quantified the clinical utility, demonstrating that the nomogram provided significant net benefit across a wide range of threshold probabilities (approximately 10 to 80%) when compared to the default strategies of intervening for all patients or for none (Figure 6).

To complement the single split validation and provide a more stable performance estimate, repeated 10-fold cross-validation was conducted on the training cohort (n = 198). Across 100 repeats, the nomogram demonstrated consistent discriminative ability, with a mean AUC of 0.763 (standard deviation, SD = 0.032). This result closely aligns with the AUC obtained from the single training set (0.771) and the independent hold-out validation set (0.820), indicating that the model’s performance is robust and not heavily dependent on a particular data partition.

Bootstrap internal validation (1,000 resamples) was performed to evaluate the robustness of the parameter estimates from the final multivariate model. The distributions of the odds ratios for pulmonary infection, years of education, and hematoma volume across all bootstrap samples were analyzed. As summarized in Table 4, the original OR estimates from the model (pulmonary infection: 3.028; years of education: 0.868; hematoma volume: 1.030) were all located close to the median (50th percentile) of their respective bootstrap distributions. Furthermore, the 95% confidence intervals derived from the bootstrap percentiles were narrow and contained the original point estimates. These results indicate that the identified predictors are stable and not artifacts of overfitting or sample variability, supporting the reliability of the nomogram’s foundation.

This study has several limitations. First, we observed that the discriminative accuracy (AUC) in the validation set (0.820) was slightly higher than in the training set (0.771). This atypical finding must be interpreted with caution. Observation of the calibration curves reveals that the validation cohort exhibits noticeable overestimation in the intermediate predicted probability range (0.3–0.6). This suggests that while the model is effective at separating low-risk and high-risk patients, the apparent superiority in AUC is likely a stochastic artifact of the small sample size (n = 66) and data distribution, rather than a true performance advantage. In smaller datasets, stochastic variation can occasionally result in a data distribution that allows for clearer separation between groups. Therefore, this performance metric should be interpreted with caution and requires confirmation in larger external cohorts. Second, as a single-center retrospective study with a relatively limited sample size, it is susceptible to selection bias, which may restrict the model’s generalizability to populations of different ethnicities or regions. Third, the study primarily focused on admission baseline indicators and did not incorporate genetic markers like APOE genotype or dynamic evolution indicators such as hematoma expansion, potentially omitting some underlying biological predictive information. Furthermore, cognitive assessment relied mainly on the MoCA scale. Although clinically practical, it may be less sensitive to subtle deficits in specific cognitive domains (e.g., executive function, memory) compared to a full neuropsychological test battery. Future improvements should involve conducting multicenter prospective cohort studies and integrating multimodal imaging and liquid biomarkers to further optimize the model’s predictive performance. Finally, it is imperative to distinguish between risk prediction and causal inference. Although our multivariate model adjusted for key confounders, observational data inevitably contain residual confounding. For instance, pulmonary infection is often intrinsically linked to overall stroke severity, aspiration risk, and prolonged immobilization. While we identified infection as a marker of high cognitive risk, this does not confirm a direct causal pathway where treating the infection reverses cognitive decline. The association may partially reflect the “sick brain” being more prone to systemic complications. Thus, our results should be interpreted as identifying a high-risk phenotype rather than validating a specific therapeutic target.