Postoperative cerebral infarction after major thoracic and abdominopelvic surgery: a retrospective single-centre cohort study of incidence, risk factors, and outcomes Authors: Xian Zhang, Yan Li, Qiang Li, Quan Zhou Published: 2026 ABSTRACT BackgroundPostoperative cerebral infarction (POCI) is an uncommon but potentially fatal complication of major non-cardiac surgery and is associated with higher in-hospital mortality, intensive care unit (ICU) admission, and subsequent mortality. The contribution of preoperative composite inflammatory and coagulation indices, including the systemic immune-inflammation index (SII), to POCI risk stratification has not been well characterized in Korean thoracic and abdominopelvic surgical cohorts.MethodsWe conducted a retrospective cohort study using the INformative Surgical Patient dataset for Innovative Research Environment (INSPIRE). Of 15,225 eligible patients who underwent elective or emergency thoracic or abdominopelvic surgery, 15,138 met inclusion criteria. The primary outcome was confirmed POCI. Independent associations with POCI were assessed by multivariable logistic regression (adjusted odds ratios (aORs) with 95% confidence intervals). Three machine learning classifiers were trained and evaluated by five-fold stratified cross-validation. Survival was compared using a pre-specified fixed-endpoint analysis (1-year all-cause mortality). Kaplan-Meier estimation was provided as supporting description.ResultsPOCI was confirmed in 277 of 15,138 patients (1.83%; 95% CI: 1.63–2.06%). In multivariable logistic regression, the factors independently associated with POCI were advanced age (aOR: 1.65 per 10-year increment; 95% CI: 1.46–1.88), hypertension (aOR: 5.92; 95% CI: 4.56–7.68), atrial fibrillation (aOR: 4.21; 95% CI: 2.60–6.79), chronic kidney disease (aOR: 2.47; 95% CI: 1.56–3.90), and peripheral arterial disease (aOR: 2.45; 95% CI: 1.27–4.72) (all p < 0.01). In unadjusted descriptive comparisons, POCI patients had higher preoperative neutrophil-to-lymphocyte ratio (NLR; median 3.1 vs. 2.5; p < 0.001), systemic immune-inflammation index (SII; 694.4 vs. 577.5; p = 0.006), and fibrinogen-to-albumin ratio (FAR; 85.5 vs. 76.3; p < 0.001). However, none of these indices was independently associated with POCI after adjustment (e.g. NLR aOR: 1.21; 95% CI: 0.93–1.58; p = 0.155). No consistent pattern of greater intra-operative hypotension was observed. In-hospital mortality (5.4% vs. 1.6%; OR: 3.56; 95% CI: 2.10–6.03) and ICU admission (42.6% vs. 22.8%; OR: 2.52; 95% CI: 1.98–3.20) were higher in the POCI group (both p < 0.001). One-year all-cause mortality was 14.8% in the POCI group versus 5.8% in controls (OR: 2.83; 95% CI: 2.02–3.96; p < 0.001), and the Kaplan-Meier log-rank comparison was non-significant (p = 0.25).ConclusionsPOCI following major thoracic and abdominopelvic surgery occurred in 1.83% of this cohort and was associated with substantially higher in-hospital mortality and higher 1-year all-cause mortality. These findings provide a population-specific incidence benchmark and support structured preoperative cardiovascular risk review, rather than reliance on inflammatory indices or machine-learning scores, whose positive predictive values are limited in this low-prevalence setting.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12871-026-04061-x. Introduction Approximately 300 million adults undergo non-cardiac surgical procedures each year worldwide, and major perioperative complications occur in an estimated 7–15% of these individuals [1]. Among the spectrum of recognized complications, postoperative cerebral infarction (POCI) occupies a particularly concerning position owing to its high case-fatality rate, the frequency of permanent neurological disability among survivors, and its considerable contribution to prolonged hospital stays and downstream health-care costs [2, 3]. The insidious clinical presentation of postoperative stroke, frequently masked by residual anesthetic effects, analgesic sedation, or the patient’s inability to communicate, compounds the problem by delaying diagnosis and the initiation of evidence-based management. Estimates of POCI incidence in non-cardiac, non-neurological surgery derived from large national administrative datasets have ranged from 0.1% to 1.9%.(2, 4, 5) The true burden is likely underestimated because covert cerebral infarctions, detectable on diffusion-weighted magnetic resonance imaging (MRI) but clinically silent, occur in approximately 7% of patients undergoing non-cardiac surgery [4]. The NeuroVISION prospective cohort demonstrated that covert stroke is at least 14 times more frequent than overt stroke after non-cardiac surgery, and that its presence independently predicts a composite of stroke, transient ischemic attack, and delirium within 30 days [5]. Overt POCI carries an in-hospital mortality of 5–26% across published series and leaves 25–45% of survivors with moderate or severe disability. Established risk factors for POCI include older age, hypertension, pre-existing cerebrovascular disease, atrial fibrillation (AF), chronic kidney disease, and elevated ASA physical status [2, 6, 7]. AF is of particular mechanistic interest: new-onset perioperative AF precipitates cardioembolic events, and epidemiological evidence suggests that patients who develop AF in the perioperative setting face a significantly elevated risk of ischemic stroke for at least 12 months following surgery [8, 9]. The 2020 Society for Neuroscience in Anesthesiology and Critical Care (SNACC) guidelines emphasize perioperative hemodynamic stability, bridging anticoagulation decisions, and structured neurological screening as the principal modifiable targets, yet no validated instrument exists for stratifying POCI risk specifically in thoracic or abdominopelvic surgical populations [7]. A growing body of evidence has linked circulating inflammatory and coagulation indices, calculable from routinely collected blood parameters, to incident cerebrovascular events. Elevated neutrophil-to-lymphocyte ratio (NLR), reflecting a shift in immune balance towards a pro-inflammatory state, has been associated with stroke severity, hemorrhagic transformation, and adverse functional outcomes in acute ischemic stroke [10, 11]. The fibrinogen-to-albumin ratio (FAR), a composite marker integrating the pro-thrombotic potential of fibrinogen with the anti-inflammatory and oncotic properties of albumin, has been identified as an independent predictor of early neurological deterioration following intravenous thrombolysis and as a prognostic marker in large-artery atherosclerosis stroke [12, 13]. The systemic immune-inflammation index (SII), derived from neutrophil, platelet, and lymphocyte counts, was recently reported to predict perioperative ischemic stroke in older patients undergoing non-cardiac surgery with superior discriminatory performance to NLR or platelet-to-lymphocyte ratio (PLR) alone [14]. However, the joint preoperative evaluation of these composite indices within a large surgical cohort remains limited. The INformative Surgical Patient dataset for Innovative Research Environment (INSPIRE) is a publicly available perioperative research database containing records from approximately 130,000 surgical admissions to Seoul National University Hospital. It includes time-series vital signs, longitudinal laboratory results, medication administration records, and post-discharge outcomes [1]. Its single-center Korean origin and multispecialty surgical scope provide an opportunity to characterize POCI incidence, risk factor profiles, and clinical consequences in a population that has been less represented in existing perioperative-stroke literature [15, 16]. The present study therefore had two primary aims: (i) to determine POCI incidence across four major surgical categories (esophageal, lung, gastrointestinal, and hepatobiliary surgery); and (ii) to identify independent, adjusted associations between preoperative clinical variables and POCI using multivariable logistic regression. The composite inflammatory indices, machine-learning discrimination, perioperative hemodynamic summaries, and fixed-endpoint clinical outcomes were also evaluated in secondary analyses. Methods Data source and ethics This retrospective cohort study used INSPIRE version 1.3 (PhysioNet DOI https://doi.org/10.13026/4hm3-0x86), a publicly available perioperative dataset released by Seoul National University Hospital (SNUH) that extends the originally published 2011–2020 release through December 2021, under Institutional Review Board approval (IRB No. H-2210-078-1368) and Institutional Data Review Board authorization for public release following de-identification [1]. Because INSPIRE is an openly available, fully anonymized secondary dataset, no additional ethical approval or individual informed consent was required for this analysis. INSPIRE encompasses records from approximately 50% of all surgical admissions to SNUH between January 2011 and December 2021, incorporating patient demographics, ASA classification, procedure codes, intraoperative vital signs recorded at one-minute intervals, all laboratory results from six months before admission to six months after discharge, drug administration data, and post-discharge survival outcomes linked to the national death registry. Study population The eligible population comprised all adult patients (aged ≥ 18 years) undergoing elective or emergency thoracic or abdominopelvic surgery within the INSPIRE dataset, with a primary procedure classified as esophageal, lung, gastrointestinal, or hepatobiliary surgery. Surgeries designated as “other” were retained only to account for the denominator in the overall incidence estimate and were excluded from comparative surgical-subgroup analyses given diagnostic heterogeneity. Patients were included in the primary analysis cohort according to a pre-specified algorithm that excluded: (a) patients with an ICD-10-coded cerebral infarction diagnosis predating the index surgery or recorded within 48 h of admission, classified as pre-existing or peri-admission (“uncertain”) non-postoperative stroke (n = 87; defined as a cerebral-infarction code recorded before the index surgery or within 48 h of admission, where pre- versus postoperative timing could not be unambiguously established from the available timestamps); and (b) patients with incomplete ascertainment of the primary outcome variable. The final primary analysis cohort comprised 15,138 patients. The complete patient selection algorithm, including ICD-10 code-based filtering, temporal criteria for distinguishing pre-existing from new postoperative stroke, the pre-existing versus uncertain breakdown within the 87 excluded patients, and the resulting cohort sizes at each stage, is illustrated in Supplementary Figure S1. Definition of postoperative cerebral infarction The primary outcome, POCI, was defined as a new diagnosis of cerebral infarction (ICD-10 codes I63.0-I63.9) first recorded during or following the index surgical admission, excluding cases where the diagnosis was present on admission or temporally inconsistent with a postoperative event as identified by the INSPIRE curation algorithm. This definition encompasses both clinically overt strokes and those incidentally coded during the admission, but does not capture covert infarctions diagnosed exclusively on imaging without a clinical diagnosis code, consistent with the ICD-based outcome ascertainment used in comparator administrative-dataset studies [2, 6, 16]. Variables and composite indices Pre-specified exposures included: age (modelled as a continuous term, expressed per 10-year increment, in the multivariable analysis; the ≥ 65-year dichotomy was retained only for descriptive tabulation to avoid the nested, collinear ≥ 65/≥75 categories used previously), sex, ASA physical status (≥ 3 as binary), emergency surgery, and a panel of comorbidities derived from ICD-10 diagnosis codes, including hypertension, type 2 diabetes mellitus, chronic kidney disease, atrial fibrillation, coronary artery disease, heart failure, and peripheral arterial disease. A composite vascular comorbidity variable (any of the above vascular conditions) was also derived. A previously included “cardioembolic-prone profile” composite (atrial fibrillation, valvular disease, or heart failure) was removed from the analysis because it overlapped with atrial fibrillation and heart failure entered as individual predictors and incorporated a valvular-disease component not tabulated separately. Laboratory values were extracted from the first preoperative measurement within 30 days preceding surgery and the first available postoperative value. Perioperative hemodynamic variables included baseline mean arterial pressure (MAP), last preoperative MAP, end-of-operation MAP, and intraoperative vasopressor exposure. Four composite inflammatory and coagulation indices were computed from available laboratory data: NLR = absolute neutrophil count ÷ absolute lymphocyte count; SII = neutrophil count × platelet count ÷ lymphocyte count [14]; FAR = fibrinogen (mg/dL) ÷ albumin (g/dL) [12, 13]; and CRP-to-albumin ratio (CAR) = CRP (mg/dL) ÷ albumin (g/dL). Delta (Δ) values for each index were computed as first postoperative minus preoperative value and are reported only as secondary descriptive analyses. Hemoglobin, creatinine, platelet count, fibrinogen, albumin, and CRP were included as individual laboratory variables. Statistical analysis Continuous variables are reported as median (interquartile range, IQR) and compared between groups using two-sided Mann-Whitney U tests. Categorical variables are reported as counts and percentages and compared using two-sided Fisher’s exact test or Pearson χ² test as appropriate. Unadjusted ORs and their 95% confidence intervals for binary exposures were computed from 2 × 2 contingency tables with Haldane-Anscombe 0.5-cell continuity correction applied where any cell contained zero observations; confidence intervals were constructed using the Wilson score method. POCI incidence and 95% confidence intervals for each surgical category were likewise estimated using the Wilson method. All p values are two-sided; p < 0.05 was considered statistically significant. Three machine-learning classifiers were evaluated as supporting prediction analyses: L2-regularised logistic regression (regularization parameter C = 0.05), random forest (300 decision trees, maximum depth 5), and gradient boosting (200 trees, maximum depth 3). All models were trained and validated using five-fold stratified cross-validation to account for the low POCI prevalence (1.83%). Missing preoperative laboratory values were imputed using feature-level medians, and continuous features were standardized before model fitting. Data completeness is summarized in Supplementary Table S1. Fibrinogen and FAR were excluded from the primary binary odds-ratio analysis and from the primary machine-learning feature matrix because FAR was missing in 27.8% of patients. Model performance was summarized using area under the ROC curve (AUC-ROC), area under the precision-recall curve (AUC-PR), Brier score, sensitivity, specificity, positive predictive value, and negative predictive value (Supplementary Figure S2). Survival time was defined from hospital admission to all-cause death as recorded by national death-registry linkage (maximum observed follow-up 12.1 years). Patients alive at the end of their admission were censored at hospital discharge, reflecting the structure of INSPIRE in which long-duration follow-up data are available only for patients in whom a subsequent death event was registered. Because patients alive at discharge are censored at hospital discharge (median 10 days) while long-duration follow-up is available almost exclusively for patients in whom a registry death was subsequently linked, the censoring in INSPIRE is informative and violates the non-informative-censoring assumption underlying both the Kaplan-Meier estimator and the Cox proportional-hazards model. We therefore pre-specified a fixed-endpoint analysis, 1-year all-cause mortality compared by Fisher’s exact test, as the primary survival comparison. Kaplan-Meier curves are presented in the Supplementary Material for descriptive purposes only, with between-group difference assessed by the log-rank test. The Cox model was not fitted because the same informative-censoring violation would bias the hazard ratio. To identify independent preoperative predictors of POCI, a multivariable logistic regression was fitted incorporating age as a continuous term expressed per 10-year increment, sex, ASA physical status ≥ 3, emergency surgery, individual binary comorbidities, and log-transformed preoperative NLR and SII. Missing laboratory values for modelled variables were imputed by feature-level medians, and continuous predictors were standardised before model fitting. Adjusted odds ratios (aORs) with 95% bootstrap confidence intervals (1,000 resamples) were reported. All analyses were conducted in Python 3.12 using scipy (v1.11), scikit-learn (v1.3), numpy (v1.24), and pandas (v2.0). Reporting This study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for cohort studies. Results Study cohort and overview Of the 15,225 patients in the INSPIRE database who underwent elective or emergency thoracic or abdominopelvic surgery at Seoul National University Hospital between January 2011 and December 2021, 15,138 met the eligibility criteria for the primary analysis after the exclusion of 87 individuals with pre-existing or diagnostically uncertain non-postoperative cerebral infarction (Fig. 1A). Postoperative cerebral infarction (POCI) was confirmed in 277 of these patients, corresponding to an overall incidence of 1.83% (95% CI: 1.63–2.06%). Fig. 1Cohort derivation and overview A Flow diagram showing derivation of the primary analysis cohort from the INSPIRE database (Seoul National University Hospital, 2011–2021). Eighty-seven patients with pre-existing or diagnostically uncertain non-postoperative cerebral infarction were excluded B Stacked bar chart showing the distribution of patients across four principal surgical categories. Annotated percentages denote POCI incidence within each category. GI, gastrointestinal C Violin plots overlaid with box-and-whisker plots comparing the age distributions of the control (n = 14,861) and POCI (n = 277) groups. The central horizontal line denotes the median; boxes span the interquartile range (IQR); whiskers extend to 1.5 × IQR, with values beyond this range shown individually. Significance brackets indicate p < 0.001 by Mann-Whitney U test D Grouped bar charts with 95% Wilson confidence interval error bars showing POCI incidence (%) stratified by surgery type and the presence or absence of any vascular comorbidity E Horizontal bar chart comparing the prevalence (%) of seven pre-specified comorbidities between groups. Asterisks denote statistical significance: *p < 0.05; **p < 0.01; ***p < 0.001 by Fisher’s exact test. POCI, postoperative cerebral infarction Cohort derivation and overview A Flow diagram showing derivation of the primary analysis cohort from the INSPIRE database (Seoul National University Hospital, 2011–2021). Eighty-seven patients with pre-existing or diagnostically uncertain non-postoperative cerebral infarction were excluded B Stacked bar chart showing the distribution of patients across four principal surgical categories. Annotated percentages denote POCI incidence within each category. GI, gastrointestinal C Violin plots overlaid with box-and-whisker plots comparing the age distributions of the control (n = 14,861) and POCI (n = 277) groups. The central horizontal line denotes the median; boxes span the interquartile range (IQR); whiskers extend to 1.5 × IQR, with values beyond this range shown individually. Significance brackets indicate p < 0.001 by Mann-Whitney U test D Grouped bar charts with 95% Wilson confidence interval error bars showing POCI incidence (%) stratified by surgery type and the presence or absence of any vascular comorbidity E Horizontal bar chart comparing the prevalence (%) of seven pre-specified comorbidities between groups. Asterisks denote statistical significance: *p < 0.05; **p < 0.01; ***p < 0.001 by Fisher’s exact test. POCI, postoperative cerebral infarction The distribution of patients across the four principal surgical categories is illustrated in Fig. 1B. Esophageal procedures constituted the largest category (n = 5,523, 36.5%), followed by lung (n = 3,189, 21.1%), gastrointestinal (n = 3,136, 20.7%), hepatobiliary (n = 2,253, 14.9%), and other procedures retained only for overall incidence accounting (n = 1,037, 6.9%). The age distribution of the two groups is shown in Fig. 1C, highlighting the pronounced rightward shift in the POCI group. POCI incidence stratified by surgery type and vascular comorbidity status is depicted in Fig. 1D, and the prevalence of pre-specified comorbidities in both groups is compared in Fig. 1E. Baseline demographic and clinical characteristics Demographic and surgical profile Complete baseline characteristics for both groups are presented in Table 1. POCI patients were substantially older than controls (median age 70.0 years, IQR 65.0–75.0 vs. 60.0 years, IQR 55.0–70.0; p < 0.001), with 80.9% aged ≥ 65 years and 41.2% aged ≥ 75 years, compared with 46.4% and 17.1% in the control group, respectively. Male sex was more prevalent among POCI patients (71.1% vs. 60.5%; p < 0.001). The POCI group also had a higher proportion of ASA physical status grade ≥ 3 (28.9% vs. 8.7%; p < 0.001), greater median number of co-existing diagnoses (8.0, IQR 5.0–13.0 vs. 3.0, IQR 2.0–5.0; p < 0.001), and a higher rate of emergency procedures (11.6% vs. 7.2%; p = 0.009). Table 1Baseline demographic and clinical characteristicsVariableControl (n = 14,861)POCI (n = 277)p valueDemographic and perioperative profile Age (years)60.0 (55.0–70.0)70.0 (65.0–75.0)< 0.001 BMI (kg/m²)22.9 (20.8–25.4)23.4 (20.8–25.4)0.731 Operation duration (hr)2.8 (1.8–4.0)2.5 (1.3–4.1)0.017 Anaesthesia duration (hr)3.5 (2.4–4.8)3.3 (2.0–4.8)0.021 Pre-operative stay (days)2.5 (2.3–3.4)2.5 (2.4–4.4)0.014 Number of diagnoses3.0 (2.0–5.0)8.0 (5.0–13.0)< 0.001 Male sex, n (%)8,990 (60.5%)197 (71.1%)< 0.001 Age ≥ 65 years, n (%)6,902 (46.4%)224 (80.9%)< 0.001 Age ≥ 75 years, n (%)2,536 (17.1%)114 (41.2%)< 0.001 ASA grade ≥ 3, n (%)1,287 (8.7%)80 (28.9%)< 0.001 Emergency surgery, n (%)1,066 (7.2%)32 (11.6%)0.009Surgical category Lung, n (%)3,128 (21.0%)61 (22.0%)0.710 Gastrointestinal, n (%)3,074 (20.7%)62 (22.4%)0.501 Esophageal, n (%)5,446 (36.6%)77 (27.8%)0.002 Hepatobiliary, n (%)2,237 (15.1%)16 (5.8%)< 0.001 Other procedure, n (%)†976 (6.6%)61 (22.0%)Descriptive onlyComorbidity profile Hypertension, n (%)1,201 (8.1%)125 (45.1%)< 0.001 Diabetes mellitus, n (%)701 (4.7%)31 (11.2%)< 0.001 Chronic kidney disease, n (%)282 (1.9%)27 (9.7%)< 0.001 Angina pectoris, n (%)598 (4.0%)35 (12.6%)< 0.001 Atrial fibrillation, n (%)142 (1.0%)28 (10.1%)< 0.001 Coronary artery disease, n (%)475 (3.2%)30 (10.8%)< 0.001 Heart failure, n (%)43 (0.3%)8 (2.9%)< 0.001 Peripheral arterial disease, n (%)107 (0.7%)13 (4.7%)< 0.001 Malignant neoplasm, n (%)11,584 (77.9%)206 (74.4%)0.165 Any vascular comorbidity, n (%)2,463 (16.6%)172 (62.1%)< 0.001Note: Data are median (interquartile range) for continuous variables and n (%) for categorical variablesPOCI Post-operative cerebral infarction, LOS Length of stay, ICU Intensive care unit, ASA American Society of Anesthesiologists physical status, BMI Body mass indexp values: Mann-Whitney U test for continuous variables; Fisher exact or chi-square test for categorical variables†Other procedures were retained only to account for the overall cohort denominator and were excluded from surgical-subgroup inference Table 1Baseline demographic and clinical characteristicsVariableControl (n = 14,861)POCI (n = 277)p valueDemographic and perioperative profile Age (years)60.0 (55.0–70.0)70.0 (65.0–75.0)< 0.001 BMI (kg/m²)22.9 (20.8–25.4)23.4 (20.8–25.4)0.731 Operation duration (hr)2.8 (1.8–4.0)2.5 (1.3–4.1)0.017 Anaesthesia duration (hr)3.5 (2.4–4.8)3.3 (2.0–4.8)0.021 Pre-operative stay (days)2.5 (2.3–3.4)2.5 (2.4–4.4)0.014 Number of diagnoses3.0 (2.0–5.0)8.0 (5.0–13.0)< 0.001 Male sex, n (%)8,990 (60.5%)197 (71.1%)< 0.001 Age ≥ 65 years, n (%)6,902 (46.4%)224 (80.9%)< 0.001 Age ≥ 75 years, n (%)2,536 (17.1%)114 (41.2%)< 0.001 ASA grade ≥ 3, n (%)1,287 (8.7%)80 (28.9%)< 0.001 Emergency surgery, n (%)1,066 (7.2%)32 (11.6%)0.009Surgical category Lung, n (%)3,128 (21.0%)61 (22.0%)0.710 Gastrointestinal, n (%)3,074 (20.7%)62 (22.4%)0.501 Esophageal, n (%)5,446 (36.6%)77 (27.8%)0.002 Hepatobiliary, n (%)2,237 (15.1%)16 (5.8%)< 0.001 Other procedure, n (%)†976 (6.6%)61 (22.0%)Descriptive onlyComorbidity profile Hypertension, n (%)1,201 (8.1%)125 (45.1%)< 0.001 Diabetes mellitus, n (%)701 (4.7%)31 (11.2%)< 0.001 Chronic kidney disease, n (%)282 (1.9%)27 (9.7%)< 0.001 Angina pectoris, n (%)598 (4.0%)35 (12.6%)< 0.001 Atrial fibrillation, n (%)142 (1.0%)28 (10.1%)< 0.001 Coronary artery disease, n (%)475 (3.2%)30 (10.8%)< 0.001 Heart failure, n (%)43 (0.3%)8 (2.9%)< 0.001 Peripheral arterial disease, n (%)107 (0.7%)13 (4.7%)< 0.001 Malignant neoplasm, n (%)11,584 (77.9%)206 (74.4%)0.165 Any vascular comorbidity, n (%)2,463 (16.6%)172 (62.1%)< 0.001Note: Data are median (interquartile range) for continuous variables and n (%) for categorical variablesPOCI Post-operative cerebral infarction, LOS Length of stay, ICU Intensive care unit, ASA American Society of Anesthesiologists physical status, BMI Body mass indexp values: Mann-Whitney U test for continuous variables; Fisher exact or chi-square test for categorical variables†Other procedures were retained only to account for the overall cohort denominator and were excluded from surgical-subgroup inference Baseline demographic and clinical characteristics Note: Data are median (interquartile range) for continuous variables and n (%) for categorical variables POCI Post-operative cerebral infarction, LOS Length of stay, ICU Intensive care unit, ASA American Society of Anesthesiologists physical status, BMI Body mass index p values: Mann-Whitney U test for continuous variables; Fisher exact or chi-square test for categorical variables †Other procedures were retained only to account for the overall cohort denominator and were excluded from surgical-subgroup inference Comorbidity profile The POCI group demonstrated a markedly higher burden of cardiovascular and cardiometabolic comorbidities across all domains examined. Hypertension was present in 45.1% of POCI patients versus 8.1% of controls (OR 9.36, 95% CI: 7.33–11.94; p < 0.001). Atrial fibrillation was identified in 10.1% of POCI patients compared with 1.0% of controls (OR 11.80, 95% CI: 7.74–17.98; p < 0.001), and chronic kidney disease was present in 9.7% versus 1.9% (OR 5.67, 95% CI: 3.76–8.55; p < 0.001). Coronary artery disease (10.8% vs. 3.2%; OR 3.73, 95% CI: 2.53–5.49), heart failure (2.9% vs. 0.3%; OR 10.74, 95% CI: 5.10-22.63), and peripheral arterial disease (4.7% vs. 0.7%; OR 7.01, 95% CI: 3.93–12.50) were all significantly more prevalent in the POCI group (all p < 0.001). In aggregate, at least one vascular comorbidity was present in 62.1% of POCI patients versus 16.6% of controls (OR 8.23, 95% CI: 6.43–10.52; p < 0.001). Preoperative and postoperative laboratory values Conventional biochemical indices Preoperative and postoperative laboratory values are presented in Table 2. POCI patients had lower preoperative albumin (median 3.7 g/dL, IQR 3.1-4.0 vs. 3.9 g/dL, IQR 3.5–4.1; p < 0.001) and hemoglobin (12.2 g/dL, IQR 10.7–13.7 vs. 12.9 g/dL, IQR 11.6–13.7; p < 0.001), and higher creatinine (0.9 mg/dL, IQR 0.7–1.2 vs. 0.8 mg/dL, IQR 0.7–0.9; p < 0.001) and fibrinogen (327.0 mg/dL, IQR 276.0-376.0 vs. 301.0 mg/dL, IQR 249.0-357.0; p < 0.001). Preoperative CRP concentrations did not differ significantly between groups (both medians 1.1 mg/dL; p = 0.885), and platelet counts were comparable (211.0 × 10⁹/L vs. 222.0 × 10⁹/L; p = 0.054). Table 2Pre-operative and first post-operative laboratory valuesLaboratory parameterControl (n = 14,861)POCI (n = 277)p valuePre-operative laboratory values Albumin (g/dL)3.9 (3.5–4.1)3.7 (3.1–4.0)< 0.001 CRP (mg/dL)1.1 (0.1–6.8)1.1 (0.1–5.6)0.885 Fibrinogen (mg/dL)†301.0 (249.0–357.0)327.0 (276.0–376.0)< 0.001 Haemoglobin (g/dL)12.9 (11.6–13.7)12.2 (10.7–13.7)< 0.001 White blood cell (×10⁹/L)6.7 (5.2–8.9)7.0 (5.6–10.2)0.007 Creatinine (mg/dL)0.8 (0.7–0.9)0.9 (0.7–1.2)< 0.001 Platelet (×10⁹/L)222.0 (178.0–273.0)211.0 (166.0–258.0)0.054 PT-INR1.0 (0.9–1.1)1.0 (1.0–1.1)0.005 APTT (seconds)30.6 (28.5–33.0)31.2 (27.2–33.8)0.841 HbA1c (%)6.1 (5.6–6.8)6.4 (6.0–7.4)0.006 Glucose (mg/dL)110.0 (96.0–132.0)120.0 (101.0–153.0)< 0.001 Total protein (g/dL)6.6 (6.1–7.0)6.5 (5.8–7.0)0.001 Total bilirubin (mg/dL)0.7 (0.5–0.8)0.5 (0.4–1.0)0.229 NLR2.5 (1.6–5.2)3.1 (1.9–6.6)< 0.001 SII577.5 (338.2–1129.0)694.4 (380.7–1414.0)0.006 CAR0.4 (0.0–1.8)0.3 (0.0–1.6)0.743 FAR†76.3 (64.1–96.4)85.5 (71.5–116.8)< 0.001First post-operative laboratory values Albumin (g/dL)3.3 (3.0–3.7)3.2 (2.9–3.7)< 0.001 CRP (mg/dL)4.6 (0.5–8.3)3.7 (0.3–8.3)0.296 Fibrinogen (mg/dL)†288.0 (234.0–357.0)301.0 (263.0–357.0)0.003 Haemoglobin (g/dL)12.2 (11.0–13.2)11.6 (10.1–13.2)< 0.001 Creatinine (mg/dL)0.8 (0.7–0.9)0.8 (0.7–1.1)< 0.001 NLR6.4 (4.0–10.0)6.6 (3.9–11.4)0.624 SII1267.2 (740.5–2105.1)1316.2 (702.2–2216.7)0.592 CAR1.3 (0.1–2.7)1.2 (0.1–2.8)0.546 FAR†84.6 (68.6–108.2)92.2 (74.8–120.0)< 0.001Note: Data are median (IQR). Delta (post-operative minus pre-operative) laboratory analyses are reserved for the Supplementary Material as exploratory analysesNLR Neutrophil-to-lymphocyte ratio, SII Systemic immune-inflammation index, CAR CRP-to-albumin ratio, FAR Fibrinogen-to-albumin ratio, CRP C-reactive protein, PT-INR Prothrombin time-international normalized ratio, APTT Activated partial thromboplastin time†Fibrinogen-derived variables, including FAR, are reported descriptively because pre-operative FAR was available in 72.2% of patients (27.8% missing) and post-operative FAR in 69.1%; they were excluded from the primary inferential models. Table 2Pre-operative and first post-operative laboratory valuesLaboratory parameterControl (n = 14,861)POCI (n = 277)p valuePre-operative laboratory values Albumin (g/dL)3.9 (3.5–4.1)3.7 (3.1–4.0)< 0.001 CRP (mg/dL)1.1 (0.1–6.8)1.1 (0.1–5.6)0.885 Fibrinogen (mg/dL)†301.0 (249.0–357.0)327.0 (276.0–376.0)< 0.001 Haemoglobin (g/dL)12.9 (11.6–13.7)12.2 (10.7–13.7)< 0.001 White blood cell (×10⁹/L)6.7 (5.2–8.9)7.0 (5.6–10.2)0.007 Creatinine (mg/dL)0.8 (0.7–0.9)0.9 (0.7–1.2)< 0.001 Platelet (×10⁹/L)222.0 (178.0–273.0)211.0 (166.0–258.0)0.054 PT-INR1.0 (0.9–1.1)1.0 (1.0–1.1)0.005 APTT (seconds)30.6 (28.5–33.0)31.2 (27.2–33.8)0.841 HbA1c (%)6.1 (5.6–6.8)6.4 (6.0–7.4)0.006 Glucose (mg/dL)110.0 (96.0–132.0)120.0 (101.0–153.0)< 0.001 Total protein (g/dL)6.6 (6.1–7.0)6.5 (5.8–7.0)0.001 Total bilirubin (mg/dL)0.7 (0.5–0.8)0.5 (0.4–1.0)0.229 NLR2.5 (1.6–5.2)3.1 (1.9–6.6)< 0.001 SII577.5 (338.2–1129.0)694.4 (380.7–1414.0)0.006 CAR0.4 (0.0–1.8)0.3 (0.0–1.6)0.743 FAR†76.3 (64.1–96.4)85.5 (71.5–116.8)< 0.001First post-operative laboratory values Albumin (g/dL)3.3 (3.0–3.7)3.2 (2.9–3.7)< 0.001 CRP (mg/dL)4.6 (0.5–8.3)3.7 (0.3–8.3)0.296 Fibrinogen (mg/dL)†288.0 (234.0–357.0)301.0 (263.0–357.0)0.003 Haemoglobin (g/dL)12.2 (11.0–13.2)11.6 (10.1–13.2)< 0.001 Creatinine (mg/dL)0.8 (0.7–0.9)0.8 (0.7–1.1)< 0.001 NLR6.4 (4.0–10.0)6.6 (3.9–11.4)0.624 SII1267.2 (740.5–2105.1)1316.2 (702.2–2216.7)0.592 CAR1.3 (0.1–2.7)1.2 (0.1–2.8)0.546 FAR†84.6 (68.6–108.2)92.2 (74.8–120.0)< 0.001Note: Data are median (IQR). Delta (post-operative minus pre-operative) laboratory analyses are reserved for the Supplementary Material as exploratory analysesNLR Neutrophil-to-lymphocyte ratio, SII Systemic immune-inflammation index, CAR CRP-to-albumin ratio, FAR Fibrinogen-to-albumin ratio, CRP C-reactive protein, PT-INR Prothrombin time-international normalized ratio, APTT Activated partial thromboplastin time†Fibrinogen-derived variables, including FAR, are reported descriptively because pre-operative FAR was available in 72.2% of patients (27.8% missing) and post-operative FAR in 69.1%; they were excluded from the primary inferential models. Pre-operative and first post-operative laboratory values Note: Data are median (IQR). Delta (post-operative minus pre-operative) laboratory analyses are reserved for the Supplementary Material as exploratory analyses NLR Neutrophil-to-lymphocyte ratio, SII Systemic immune-inflammation index, CAR CRP-to-albumin ratio, FAR Fibrinogen-to-albumin ratio, CRP C-reactive protein, PT-INR Prothrombin time-international normalized ratio, APTT Activated partial thromboplastin time †Fibrinogen-derived variables, including FAR, are reported descriptively because pre-operative FAR was available in 72.2% of patients (27.8% missing) and post-operative FAR in 69.1%; they were excluded from the primary inferential models. Composite inflammatory and coagulation indices Among the novel composite indices, the preoperative NLR was higher in the POCI group (median 3.1, IQR 1.9–6.6 vs. 2.5, IQR 1.6–5.2; p < 0.001), as was the systemic immune-inflammation index (SII: 694.4, IQR 380.7–1414.0 vs. 577.5, IQR 338.2–1129.0; p = 0.006). The fibrinogen-to-albumin ratio (FAR) was also elevated preoperatively in the POCI group (85.5, IQR 71.5-116.8 vs. 76.3, IQR 64.1–96.4; p < 0.001), although this comparison is descriptive. By contrast, the CRP-to-albumin ratio (CAR) did not differ significantly between groups (median 0.3 vs. 0.4; p = 0.743). These distribution comparisons are illustrated in Fig. 2. Fig. 2Preoperative and postoperative composite inflammatory and coagulation indices. Violin plots overlaid with box-and-whisker plots compare the distributions of NLR, SII, CAR, and FAR between the control and POCI groups at the preoperative and first postoperative measurements, as applicable. Values are presented on a natural-logarithm scale where indicated. Significance brackets are from two-sided Mann-Whitney U tests. NLR, neutrophil-to-lymphocyte ratio; SII, systemic immune-inflammation index (neutrophil count × platelet count / lymphocyte count); CAR, CRP-to-albumin ratio; FAR, fibrinogen-to-albumin ratio; POCI, postoperative cerebral infarction Preoperative and postoperative composite inflammatory and coagulation indices. Violin plots overlaid with box-and-whisker plots compare the distributions of NLR, SII, CAR, and FAR between the control and POCI groups at the preoperative and first postoperative measurements, as applicable. Values are presented on a natural-logarithm scale where indicated. Significance brackets are from two-sided Mann-Whitney U tests. NLR, neutrophil-to-lymphocyte ratio; SII, systemic immune-inflammation index (neutrophil count × platelet count / lymphocyte count); CAR, CRP-to-albumin ratio; FAR, fibrinogen-to-albumin ratio; POCI, postoperative cerebral infarction Perioperative hemodynamic parameters Perioperative hemodynamic summaries are presented in Table 3; Fig. 3. Baseline MAP, end-of-operation MAP, and relative MAP decrease were similar between groups (all p > 0.50). Vasopressor use was less frequent in the POCI group (30.0% vs. 36.6%; p = 0.023), but this secondary observation is likely confounded by surgical profile, anesthetic approach, and case complexity. Overall, the available hemodynamic summaries did not show a consistent pattern of greater hypotension among patients who developed POCI. Table 3Perioperative haemodynamic parametersHaemodynamic parameterControl (n = 14,861)POCI (n = 277)p valueBaseline haemodynamics MAP (mmHg)75.0 (66.0–88.0)74.0 (64.0–88.0)0.631 SBP (mmHg)106.0 (94.0–120.0)110.0 (94.0–130.0)0.006 DBP (mmHg)58.0 (52.0–68.0)58.0 (50.0–66.0)0.027 Heart rate (bpm)64.0 (56.0–74.0)66.0 (56.0–78.0)0.204 Shock index0.6 (0.5–0.7)0.6 (0.5–0.7)0.097Last pre-operative measurement MAP (mmHg)71.0 (64.0–81.5)71.0 (64.0–86.0)0.661 SBP (mmHg)100.0 (92.0–112.0)104.0 (94.0–120.0)< 0.001 DBP (mmHg)55.0 (50.0–62.5)54.0 (47.0–64.0)0.167 Heart rate (bpm)64.0 (56.0–74.0)66.0 (56.0–78.0)0.057 Shock index0.6 (0.5–0.7)0.6 (0.5–0.8)0.169End of operation room MAP (mmHg)88.0 (78.0–98.0)88.0 (77.0–106.0)0.565 SBP (mmHg)122.0 (109.5–140.0)130.0 (112.0–150.0)< 0.001 DBP (mmHg)66.0 (58.0–76.0)64.0 (54.0–76.0)0.037 Heart rate (bpm)74.0 (64.0–84.0)76.0 (62.0–88.0)0.693 Shock index0.6 (0.5–0.7)0.6 (0.5–0.7)0.026First post-operative measurement MAP (mmHg)94.0 (83.0–112.0)98.0 (86.0–112.0)0.106 SBP (mmHg)123.0 (113.0–132.0)129.0 (115.0–137.0)< 0.001 DBP (mmHg)76.0 (69.0–85.0)74.0 (67.0–82.0)0.034 Heart rate (bpm)76.0 (66.0–85.0)76.0 (62.0–85.8)0.075Shock index0.6 (0.5–0.7)0.6 (0.5–0.7)< 0.001 Intraoperative haemodynamic variability and vasopressor exposure Relative MAP drop, proportion0.1 (0.0–0.2)0.1 (0.0–0.2)0.835 Relative SBP drop, proportion0.1 (0.0–0.2)0.1 (0.0–0.2)0.232 MAP range (mmHg)26.0 (16.0–38.0)28.0 (19.0–42.0)< 0.001 SBP range (mmHg)38.5 (28.0–52.0)40.0 (28.0–54.0)0.058 Heart rate range (bpm)22.0 (14.0–29.0)20.0 (12.0–28.0)0.199 ΔMAP: last pre-op minus baseline (mmHg)-2.0 (-12.0–4.0)0.0 (-9.0–4.0)0.156 ΔMAP: OR end minus last pre-op (mmHg)16.0 (3.5–28.0)14.0 (0.0–28.0)0.296 Any vasopressor use, n (%)5,434 (36.6%)83 (30.0%)0.023Note: Data are median (IQR) unless otherwise statedRelative MAP/SBP drop = (baseline value − OR-end value) / baseline valuep values from Mann-Whitney U test for continuous variables and Fisher exact test for vasopressor useMAP Mean arterial pressure, SBP Systolic blood pressure, DBP Diastolic blood pressure, OR Operating room Table 3Perioperative haemodynamic parametersHaemodynamic parameterControl (n = 14,861)POCI (n = 277)p valueBaseline haemodynamics MAP (mmHg)75.0 (66.0–88.0)74.0 (64.0–88.0)0.631 SBP (mmHg)106.0 (94.0–120.0)110.0 (94.0–130.0)0.006 DBP (mmHg)58.0 (52.0–68.0)58.0 (50.0–66.0)0.027 Heart rate (bpm)64.0 (56.0–74.0)66.0 (56.0–78.0)0.204 Shock index0.6 (0.5–0.7)0.6 (0.5–0.7)0.097Last pre-operative measurement MAP (mmHg)71.0 (64.0–81.5)71.0 (64.0–86.0)0.661 SBP (mmHg)100.0 (92.0–112.0)104.0 (94.0–120.0)< 0.001 DBP (mmHg)55.0 (50.0–62.5)54.0 (47.0–64.0)0.167 Heart rate (bpm)64.0 (56.0–74.0)66.0 (56.0–78.0)0.057 Shock index0.6 (0.5–0.7)0.6 (0.5–0.8)0.169End of operation room MAP (mmHg)88.0 (78.0–98.0)88.0 (77.0–106.0)0.565 SBP (mmHg)122.0 (109.5–140.0)130.0 (112.0–150.0)< 0.001 DBP (mmHg)66.0 (58.0–76.0)64.0 (54.0–76.0)0.037 Heart rate (bpm)74.0 (64.0–84.0)76.0 (62.0–88.0)0.693 Shock index0.6 (0.5–0.7)0.6 (0.5–0.7)0.026First post-operative measurement MAP (mmHg)94.0 (83.0–112.0)98.0 (86.0–112.0)0.106 SBP (mmHg)123.0 (113.0–132.0)129.0 (115.0–137.0)< 0.001 DBP (mmHg)76.0 (69.0–85.0)74.0 (67.0–82.0)0.034 Heart rate (bpm)76.0 (66.0–85.0)76.0 (62.0–85.8)0.075Shock index0.6 (0.5–0.7)0.6 (0.5–0.7)< 0.001 Intraoperative haemodynamic variability and vasopressor exposure Relative MAP drop, proportion0.1 (0.0–0.2)0.1 (0.0–0.2)0.835 Relative SBP drop, proportion0.1 (0.0–0.2)0.1 (0.0–0.2)0.232 MAP range (mmHg)26.0 (16.0–38.0)28.0 (19.0–42.0)< 0.001 SBP range (mmHg)38.5 (28.0–52.0)40.0 (28.0–54.0)0.058 Heart rate range (bpm)22.0 (14.0–29.0)20.0 (12.0–28.0)0.199 ΔMAP: last pre-op minus baseline (mmHg)-2.0 (-12.0–4.0)0.0 (-9.0–4.0)0.156 ΔMAP: OR end minus last pre-op (mmHg)16.0 (3.5–28.0)14.0 (0.0–28.0)0.296 Any vasopressor use, n (%)5,434 (36.6%)83 (30.0%)0.023Note: Data are median (IQR) unless otherwise statedRelative MAP/SBP drop = (baseline value − OR-end value) / baseline valuep values from Mann-Whitney U test for continuous variables and Fisher exact test for vasopressor useMAP Mean arterial pressure, SBP Systolic blood pressure, DBP Diastolic blood pressure, OR Operating room Perioperative haemodynamic parameters Note: Data are median (IQR) unless otherwise stated Relative MAP/SBP drop = (baseline value − OR-end value) / baseline value p values from Mann-Whitney U test for continuous variables and Fisher exact test for vasopressor use MAP Mean arterial pressure, SBP Systolic blood pressure, DBP Diastolic blood pressure, OR Operating room Fig. 3Perioperative hemodynamic profiles A, B Line graphs (mean ± standard error) showing the trajectory of mean arterial pressure (MAP) and systolic blood pressure (SBP), respectively, across four perioperative time-points, baseline, last preoperative measurement, end of operation room (OR), and first postoperative measurement, in the control (teal) and POCI (gold) groups. Asterisks above individual time-points denote between-group differences by Mann-Whitney U test (*p < 0.05; ***p < 0.001) C Density histogram and kernel density estimate curves showing the distribution of the relative MAP drop (defined as [baseline MAP − OR-end MAP] / baseline MAP) in each group. Dashed vertical lines indicate group medians. The annotated p value is from the Mann-Whitney U test. MAP, mean arterial pressure; SBP, systolic blood pressure Perioperative hemodynamic profiles A, B Line graphs (mean ± standard error) showing the trajectory of mean arterial pressure (MAP) and systolic blood pressure (SBP), respectively, across four perioperative time-points, baseline, last preoperative measurement, end of operation room (OR), and first postoperative measurement, in the control (teal) and POCI (gold) groups. Asterisks above individual time-points denote between-group differences by Mann-Whitney U test (*p < 0.05; ***p < 0.001) C Density histogram and kernel density estimate curves showing the distribution of the relative MAP drop (defined as [baseline MAP − OR-end MAP] / baseline MAP) in each group. Dashed vertical lines indicate group medians. The annotated p value is from the Mann-Whitney U test. MAP, mean arterial pressure; SBP, systolic blood pressure Univariate risk-factor analysis Unadjusted ORs for all pre-specified binary risk factors are displayed as a forest plot in Fig. 4 and tabulated in Table 4. Among the comorbidities examined, atrial fibrillation carried the largest effect estimate (OR 11.80; 95% CI: 7.74–17.98), followed by heart failure (OR 10.74; 95% CI: 5.10-22.63), hypertension (OR 9.36; 95% CI: 7.33–11.94), peripheral arterial disease (OR 7.01; 95% CI: 3.93–12.50), and chronic kidney disease (OR 5.67; 95% CI: 3.76–8.55). Among demographic and surgical variables, age ≥ 65 years yielded an OR of 4.84 (95% CI: 3.58–6.53; p < 0.001), ASA grade ≥ 3 an OR of 4.30 (95% CI: 3.30–5.60; p < 0.001), and emergency surgery an OR of 1.71 (95% CI: 1.18–2.48; p = 0.009). Unadjusted odds ratios estimated in the complete-case cohort (patients with non-missing values across all machine-learning predictor variables; n = 10,694; POCI = 213) were consistent in direction and magnitude with those from the full imputed cohort (Supplementary Table S2). Fig. 4Forest plot of univariate odds ratios for postoperative cerebral infarction. Each row represents a binary exposure. Squares indicate unadjusted odds ratios (OR); horizontal lines represent 95% confidence intervals (CI). The vertical dashed reference line is drawn at OR = 1.0. Variables with p < 0.05 (Fisher’s exact test) are shown in red; those with p ≥ 0.05 are shown in grey. ORs and p values are annotated to the right of each confidence interval. All ORs are computed from 2 × 2 contingency tables with Haldane-Anscombe 0.5-cell continuity correction and sorted by ascending OR Forest plot of univariate odds ratios for postoperative cerebral infarction. Each row represents a binary exposure. Squares indicate unadjusted odds ratios (OR); horizontal lines represent 95% confidence intervals (CI). The vertical dashed reference line is drawn at OR = 1.0. Variables with p < 0.05 (Fisher’s exact test) are shown in red; those with p ≥ 0.05 are shown in grey. ORs and p values are annotated to the right of each confidence interval. All ORs are computed from 2 × 2 contingency tables with Haldane-Anscombe 0.5-cell continuity correction and sorted by ascending OR Table 4Univariate and multivariable risk-factor analysis for post-operative cerebral infarctionRisk factorControl, n/N (%)POCI, n/N (%)Unadjusted OR (95% CI)p valueAdjusted OR (95% CI)p valueDemographic and surgical factors Age ≥ 65 years6,902/14,861 (46.4%)224/277 (80.9%)4.84 (3.58–6.53)< 0.001—— Age ≥ 75 years2,536/14,861 (17.1%)114/277 (41.2%)3.40 (2.67–4.34)< 0.001—— Age, per 10-year increment————1.65 (1.46–1.88)< 0.001 Male sex8,990/14,861 (60.5%)197/277 (71.1%)1.60 (1.23–2.08)< 0.001—— ASA grade ≥ 31,287/14,861 (8.7%)80/277 (28.9%)4.30 (3.30–5.60)< 0.001—— Emergency surgery1,066/14,861 (7.2%)32/277 (11.6%)1.71 (1.18–2.48)0.0091.57 (1.04–2.37)0.030Vascular comorbidities Hypertension1,201/14,861 (8.1%)125/277 (45.1%)9.36 (7.33–11.94)< 0.0015.92 (4.56–7.68)< 0.001 Diabetes mellitus701/14,861 (4.7%)31/277 (11.2%)2.58 (1.77–3.77)< 0.0011.21 (0.80–1.84)0.369 Chronic kidney disease282/14,861 (1.9%)27/277 (9.7%)5.67 (3.76–8.55)< 0.0012.47 (1.56–3.90)< 0.001 Angina pectoris598/14,861 (4.0%)35/277 (12.6%)3.49 (2.43–5.01)< 0.001—— Atrial fibrillation142/14,861 (1.0%)28/277 (10.1%)11.80 (7.74–17.98)< 0.0014.21 (2.60–6.79)< 0.001 Coronary artery disease475/14,861 (3.2%)30/277 (10.8%)3.73 (2.53–5.49)< 0.0011.27 (0.82–1.97)0.284 Heart failure43/14,861 (0.3%)8/277 (2.9%)10.74 (5.10–22.63)< 0.0011.29 (0.50–3.32)0.591 Peripheral arterial disease107/14,861 (0.7%)13/277 (4.7%)7.01 (3.93–12.50)< 0.0012.45 (1.27–4.72)0.008 Any vascular comorbidity (≥ 1)2,463/14,861 (16.6%)172/277 (62.1%)8.23 (6.43–10.52)< 0.001——Pre-operative composite inflammatory indices NLR, log-transformed per SD increase————1.21 (0.93–1.58)0.155 SII, log-transformed per SD increase————0.94 (0.71–1.23)0.641Perioperative interventions and support Any vasopressor exposure5,434/14,861 (36.6%)83/277 (30.0%)0.74 (0.58–0.96)0.023—— Red blood cell transfusion306/14,861 (2.1%)8/277 (2.9%)1.50 (0.75–2.99)0.289—— Fresh frozen plasma transfusion336/14,861 (2.3%)9/277 (3.2%)1.53 (0.79–2.94)0.303—— Any blood product exposure1,134/14,861 (7.6%)26/277 (9.4%)1.27 (0.85–1.91)0.256—— Mechanical ventilation9,264/14,861 (62.3%)154/277 (55.6%)0.76 (0.60–0.96)0.024——Note: Unadjusted ORs were computed from 2 × 2 contingency tables with Haldane-Anscombe 0.5-cell continuity correction where required; p values are from two-sided Fisher exact tests. Adjusted estimates are shown for variables included in the multivariable logistic-regression model; age was modelled as a per-10-year increment, so age ≥ 65 and age ≥ 75 are retained only as unadjusted/descriptive risk strata. Dashes indicate not applicable or not included in the final adjusted modelOR Odds ratio, aOR Adjusted odds ratio, CI Confidence interval, ASA American Society of Anesthesiologists, NLR Neutrophil-to-lymphocyte ratio, SII Systemic immune-inflammation index, POCI Post-operative cerebral infarction Table 4Univariate and multivariable risk-factor analysis for post-operative cerebral infarctionRisk factorControl, n/N (%)POCI, n/N (%)Unadjusted OR (95% CI)p valueAdjusted OR (95% CI)p valueDemographic and surgical factors Age ≥ 65 years6,902/14,861 (46.4%)224/277 (80.9%)4.84 (3.58–6.53)< 0.001—— Age ≥ 75 years2,536/14,861 (17.1%)114/277 (41.2%)3.40 (2.67–4.34)< 0.001—— Age, per 10-year increment————1.65 (1.46–1.88)< 0.001 Male sex8,990/14,861 (60.5%)197/277 (71.1%)1.60 (1.23–2.08)< 0.001—— ASA grade ≥ 31,287/14,861 (8.7%)80/277 (28.9%)4.30 (3.30–5.60)< 0.001—— Emergency surgery1,066/14,861 (7.2%)32/277 (11.6%)1.71 (1.18–2.48)0.0091.57 (1.04–2.37)0.030Vascular comorbidities Hypertension1,201/14,861 (8.1%)125/277 (45.1%)9.36 (7.33–11.94)< 0.0015.92 (4.56–7.68)< 0.001 Diabetes mellitus701/14,861 (4.7%)31/277 (11.2%)2.58 (1.77–3.77)< 0.0011.21 (0.80–1.84)0.369 Chronic kidney disease282/14,861 (1.9%)27/277 (9.7%)5.67 (3.76–8.55)< 0.0012.47 (1.56–3.90)< 0.001 Angina pectoris598/14,861 (4.0%)35/277 (12.6%)3.49 (2.43–5.01)< 0.001—— Atrial fibrillation142/14,861 (1.0%)28/277 (10.1%)11.80 (7.74–17.98)< 0.0014.21 (2.60–6.79)< 0.001 Coronary artery disease475/14,861 (3.2%)30/277 (10.8%)3.73 (2.53–5.49)< 0.0011.27 (0.82–1.97)0.284 Heart failure43/14,861 (0.3%)8/277 (2.9%)10.74 (5.10–22.63)< 0.0011.29 (0.50–3.32)0.591 Peripheral arterial disease107/14,861 (0.7%)13/277 (4.7%)7.01 (3.93–12.50)< 0.0012.45 (1.27–4.72)0.008 Any vascular comorbidity (≥ 1)2,463/14,861 (16.6%)172/277 (62.1%)8.23 (6.43–10.52)< 0.001——Pre-operative composite inflammatory indices NLR, log-transformed per SD increase————1.21 (0.93–1.58)0.155 SII, log-transformed per SD increase————0.94 (0.71–1.23)0.641Perioperative interventions and support Any vasopressor exposure5,434/14,861 (36.6%)83/277 (30.0%)0.74 (0.58–0.96)0.023—— Red blood cell transfusion306/14,861 (2.1%)8/277 (2.9%)1.50 (0.75–2.99)0.289—— Fresh frozen plasma transfusion336/14,861 (2.3%)9/277 (3.2%)1.53 (0.79–2.94)0.303—— Any blood product exposure1,134/14,861 (7.6%)26/277 (9.4%)1.27 (0.85–1.91)0.256—— Mechanical ventilation9,264/14,861 (62.3%)154/277 (55.6%)0.76 (0.60–0.96)0.024——Note: Unadjusted ORs were computed from 2 × 2 contingency tables with Haldane-Anscombe 0.5-cell continuity correction where required; p values are from two-sided Fisher exact tests. Adjusted estimates are shown for variables included in the multivariable logistic-regression model; age was modelled as a per-10-year increment, so age ≥ 65 and age ≥ 75 are retained only as unadjusted/descriptive risk strata. Dashes indicate not applicable or not included in the final adjusted modelOR Odds ratio, aOR Adjusted odds ratio, CI Confidence interval, ASA American Society of Anesthesiologists, NLR Neutrophil-to-lymphocyte ratio, SII Systemic immune-inflammation index, POCI Post-operative cerebral infarction Univariate and multivariable risk-factor analysis for post-operative cerebral infarction Note: Unadjusted ORs were computed from 2 × 2 contingency tables with Haldane-Anscombe 0.5-cell continuity correction where required; p values are from two-sided Fisher exact tests. Adjusted estimates are shown for variables included in the multivariable logistic-regression model; age was modelled as a per-10-year increment, so age ≥ 65 and age ≥ 75 are retained only as unadjusted/descriptive risk strata. Dashes indicate not applicable or not included in the final adjusted model OR Odds ratio, aOR Adjusted odds ratio, CI Confidence interval, ASA American Society of Anesthesiologists, NLR Neutrophil-to-lymphocyte ratio, SII Systemic immune-inflammation index, POCI Post-operative cerebral infarction Multivariable risk-factor analysis In multivariable logistic regression mutually adjusting for all pre-specified comorbidities, age (per 10-year increment), sex, ASA physical status ≥ 3, emergency surgery, and log-transformed preoperative NLR and SII (n = 15,138), six variables were independently associated with POCI (Table 4). For specific estimates, they were hypertension (aOR 5.92; 95% CI: 4.56–7.68; p < 0.001), atrial fibrillation (aOR 4.21; 95% CI: 2.60–6.79; p < 0.001), chronic kidney disease (aOR 2.47; 95% CI: 1.56–3.90; p < 0.001), peripheral arterial disease (aOR 2.45; 95% CI: 1.27–4.72; p = 0.008), age (aOR 1.65 per 10-year increment; 95% CI: 1.46–1.88; p < 0.001), and emergency surgery (aOR 1.57; 95% CI: 1.04–2.37; p = 0.030). Coronary artery disease (aOR 1.27; 95% CI: 0.82–1.97; p = 0.284), heart failure (aOR 1.29; 95% CI: 0.50–3.32; p = 0.591), and diabetes mellitus (aOR 1.21; 95% CI: 0.80–1.84; p = 0.369) were not independently associated with POCI after adjustment. Neither log-transformed NLR (aOR 1.21 per SD; 95% CI: 0.93–1.58; p = 0.155) nor SII (aOR 0.94 per SD; 95% CI: 0.71–1.23; p = 0.641) was independently associated with POCI after adjustment for the comorbidity profile. Variance-inflation factors for all predictors were below 5 (Supplementary Table S3), indicating acceptable collinearity. Machine learning-based prediction of POCI The machine-learning analysis is summarized in Fig. 5. L2-regularised logistic regression achieved the highest AUC-ROC (0.819), followed by random forest (0.806) and gradient boosting (0.801) (Fig. 5A). However, precision-recall performance was limited: the AUC-PR values were 0.118, 0.081, and 0.089 for logistic regression, random forest, and gradient boosting, respectively, against a baseline event prevalence of 1.83%. The low positive predictive values (6.0-16.7%) indicate limited screening utility despite acceptable AUC-ROC values. Fig. 5Machine-learning model performance and feature importance A Receiver-operating-characteristic (ROC) curves for L2-regularised logistic regression, random forest, and gradient boosting, derived from five-fold stratified cross-validation. The diagonal dashed line represents a random classifier (AUC = 0.500). Area under the ROC curve (AUC-ROC) values are given in the legend B Horizontal bar chart of the top 15 features ranked by mean decrease in Gini impurity from the random-forest classifier. Raw dataset variable names have been replaced with publication-ready clinical labels. All features were standardized by z-score before modelling; missing values were imputed using feature medians. NLR, neutrophil-to-lymphocyte ratio; SII, systemic immune-inflammation index; CAR, CRP-to-albumin ratio; ASA, American Society of Anesthesiologists physical status Machine-learning model performance and feature importance A Receiver-operating-characteristic (ROC) curves for L2-regularised logistic regression, random forest, and gradient boosting, derived from five-fold stratified cross-validation. The diagonal dashed line represents a random classifier (AUC = 0.500). Area under the ROC curve (AUC-ROC) values are given in the legend B Horizontal bar chart of the top 15 features ranked by mean decrease in Gini impurity from the random-forest classifier. Raw dataset variable names have been replaced with publication-ready clinical labels. All features were standardized by z-score before modelling; missing values were imputed using feature medians. NLR, neutrophil-to-lymphocyte ratio; SII, systemic immune-inflammation index; CAR, CRP-to-albumin ratio; ASA, American Society of Anesthesiologists physical status Random-forest feature-importance analysis (Fig. 5B) identified vascular comorbidity count, age, operation duration, atrial fibrillation, hypertension, chronic kidney disease, peripheral arterial disease, and log-transformed preoperative NLR and SII as leading contributors to model discrimination. Permutation-based importance estimates and decision-curve analysis are provided in Supplementary Figures S3 and S4. Clinical outcomes Clinical outcomes are detailed in Table 5 and illustrated in Fig. 6. In-hospital mortality was 5.4% in the POCI group versus 1.6% in controls (OR: 3.56, 95% CI: 2.10–6.03; p < 0.001). ICU admission rates were 42.6% versus 22.8% (OR: 2.52, 95% CI: 1.98–3.20; p < 0.001), and a prolonged hospital stay exceeding 14 days was recorded in 39.0% of POCI patients compared with 24.7% of controls (OR: 1.95, 95% CI: 1.53–2.48; p < 0.001). Hospital LOS was longer in the POCI group (median 12.0 days, IQR 9.0–22.0 vs. 10.0 days, IQR 8.0–13.0; p < 0.001). Among patients admitted to the ICU, ICU LOS was also greater in the POCI group (median 1.1 days, IQR 0.9–3.2 vs. 1.0 days, IQR 0.8–1.9; p = 0.005). All-cause mortality recorded during the observation period was 49.1% in the POCI group and 26.4% in controls (OR: 2.68, 95% CI: 2.12–3.41; p < 0.001). One-year all-cause mortality, the pre-specified survival endpoint (Sect. 3.10), was 14.8% in the POCI group versus 5.8% in controls (OR: 2.83; 95% CI: 2.02–3.96; p < 0.001). Table 5Clinical outcomes in patients with and without POCIOutcomePOCI (n = 277)Control (n = 14,861)OR or statistic (95% CI)p valueIn-hospital outcomes In-hospital mortality15/277 (5.4%)242/14,861 (1.6%)OR 3.56 (2.10–6.03)< 0.001 ICU admission118/277 (42.6%)3,388/14,861 (22.8%)OR 2.52 (1.98–3.20)< 0.001 Hospital LOS, days12.0 (9.0–22.0)10.0 (8.0–13.0)MWU< 0.001 Hospital LOS > 14 days108/277 (39.0%)3,671/14,861 (24.7%)OR 1.95 (1.53–2.48)< 0.001 ICU LOS, days (ICU-admitted only)1.1 (0.9–3.2)1.0 (0.8–1.9)MWU0.005Mortality outcomes All-cause mortality during registry observation136/277 (49.1%)3,928/14,861 (26.4%)OR 2.68 (2.12–3.41)< 0.001 One-year all-cause mortality (fixed endpoint)41/277 (14.8%)868/14,861 (5.8%)OR 2.83 (2.02–3.96)< 0.001 Thirty-day mortality0/277 (0.0%)approximately 0.3%Not stableNot tested Median time to death, days (IQR)†841 (287–2,028)907 (415–1,709)——Note: Data are n/N (%) for binary outcomes and median (IQR) for continuous outcomes. Binary outcomes were compared by Fisher exact test; continuous outcomes by Mann-Whitney U test. One-year all-cause mortality is the primary fixed-endpoint survival comparison because INSPIRE has informative censoring for longer-term Kaplan-Meier estimation. Kaplan-Meier curves and log-rank testing are supplementary/descriptive only and should not be interpreted as population-level actuarial survivalICU Intensive care unit, LOS Length of stay, MWU Mann-Whitney U test†Median time from admission to death is reported only among patients with a linked death record Table 5Clinical outcomes in patients with and without POCIOutcomePOCI (n = 277)Control (n = 14,861)OR or statistic (95% CI)p valueIn-hospital outcomes In-hospital mortality15/277 (5.4%)242/14,861 (1.6%)OR 3.56 (2.10–6.03)< 0.001 ICU admission118/277 (42.6%)3,388/14,861 (22.8%)OR 2.52 (1.98–3.20)< 0.001 Hospital LOS, days12.0 (9.0–22.0)10.0 (8.0–13.0)MWU< 0.001 Hospital LOS > 14 days108/277 (39.0%)3,671/14,861 (24.7%)OR 1.95 (1.53–2.48)< 0.001 ICU LOS, days (ICU-admitted only)1.1 (0.9–3.2)1.0 (0.8–1.9)MWU0.005Mortality outcomes All-cause mortality during registry observation136/277 (49.1%)3,928/14,861 (26.4%)OR 2.68 (2.12–3.41)< 0.001 One-year all-cause mortality (fixed endpoint)41/277 (14.8%)868/14,861 (5.8%)OR 2.83 (2.02–3.96)< 0.001 Thirty-day mortality0/277 (0.0%)approximately 0.3%Not stableNot tested Median time to death, days (IQR)†841 (287–2,028)907 (415–1,709)——Note: Data are n/N (%) for binary outcomes and median (IQR) for continuous outcomes. Binary outcomes were compared by Fisher exact test; continuous outcomes by Mann-Whitney U test. One-year all-cause mortality is the primary fixed-endpoint survival comparison because INSPIRE has informative censoring for longer-term Kaplan-Meier estimation. Kaplan-Meier curves and log-rank testing are supplementary/descriptive only and should not be interpreted as population-level actuarial survivalICU Intensive care unit, LOS Length of stay, MWU Mann-Whitney U test†Median time from admission to death is reported only among patients with a linked death record Clinical outcomes in patients with and without POCI Note: Data are n/N (%) for binary outcomes and median (IQR) for continuous outcomes. Binary outcomes were compared by Fisher exact test; continuous outcomes by Mann-Whitney U test. One-year all-cause mortality is the primary fixed-endpoint survival comparison because INSPIRE has informative censoring for longer-term Kaplan-Meier estimation. Kaplan-Meier curves and log-rank testing are supplementary/descriptive only and should not be interpreted as population-level actuarial survival ICU Intensive care unit, LOS Length of stay, MWU Mann-Whitney U test †Median time from admission to death is reported only among patients with a linked death record Fig. 6In-hospital and follow-up clinical outcomes A Grouped bar charts comparing the prevalence (%) of four binary outcomes, in-hospital mortality, ICU admission, all-cause mortality, and prolonged hospital stay (> 14 days), between control (teal) and POCI (gold) patients. Significance annotations are from Fisher’s exact test B Box-and-whisker plots of total hospital length of stay (LOS), restricted to 60 days for display clarity C Box-and-whisker plots of ICU LOS, restricted to 30 days, for patients admitted to the ICU. In (B) and (C), boxes span the IQR, horizontal lines within boxes indicate medians, and significance brackets are from Mann-Whitney U tests. ICU, intensive care unit; LOS, length of stay In-hospital and follow-up clinical outcomes A Grouped bar charts comparing the prevalence (%) of four binary outcomes, in-hospital mortality, ICU admission, all-cause mortality, and prolonged hospital stay (> 14 days), between control (teal) and POCI (gold) patients. Significance annotations are from Fisher’s exact test B Box-and-whisker plots of total hospital length of stay (LOS), restricted to 60 days for display clarity C Box-and-whisker plots of ICU LOS, restricted to 30 days, for patients admitted to the ICU. In (B) and (C), boxes span the IQR, horizontal lines within boxes indicate medians, and significance brackets are from Mann-Whitney U tests. ICU, intensive care unit; LOS, length of stay POCI incidence stratified by surgery type POCI incidence and associated odds ratios stratified by the four pre-defined surgical categories are presented in Table 6. Hepatobiliary surgery carried the lowest incidence at 0.71% (95% CI: 0.44–1.15%; OR 0.36, 95% CI: 0.22–0.59; p < 0.001 vs. the remainder of the cohort). Esophageal surgery had an incidence of 1.39% (1.12–1.74%; OR 0.67, 95% CI: 0.51–0.87; p = 0.002). Lung (1.91%, 95% CI: 1.49–2.45%; OR 1.07, 95% CI: 0.80–1.42; p = 0.710) and gastrointestinal (1.98%, 95% CI: 1.55–2.53%; OR 1.11, 95% CI: 0.84–1.48; p = 0.501) procedures did not show a statistically significant deviation from the overall cohort incidence. The surgery-by-vascular-comorbidity stratification in Fig. 1D demonstrated consistently higher POCI rates in vascular-comorbidity-positive patients across surgical categories. Table 6POCI incidence and odds ratios stratified by the four predefined surgery typesSurgery typeTotal NPOCI nPOCI incidence (95% CI)OR (95% CI)p valueLung3,189611.91% (1.49–2.45%)1.07 (0.80–1.42)0.710Gastrointestinal3,136621.98% (1.55–2.53%)1.11 (0.84–1.48)0.501Esophageal5,523771.39% (1.12–1.74%)0.67 (0.51–0.87)0.002Hepatobiliary2,253160.71% (0.44–1.15%)0.36 (0.22–0.59)< 0.001Note: POCI incidence and 95% confidence intervals were estimated using the Wilson score method. OR compares each surgery type against the remainder of the cohort. The “Other” procedure category was excluded from this surgical-subgroup analysis in accordance with the revised objective focusing on the four predefined surgical categories Table 6POCI incidence and odds ratios stratified by the four predefined surgery typesSurgery typeTotal NPOCI nPOCI incidence (95% CI)OR (95% CI)p valueLung3,189611.91% (1.49–2.45%)1.07 (0.80–1.42)0.710Gastrointestinal3,136621.98% (1.55–2.53%)1.11 (0.84–1.48)0.501Esophageal5,523771.39% (1.12–1.74%)0.67 (0.51–0.87)0.002Hepatobiliary2,253160.71% (0.44–1.15%)0.36 (0.22–0.59)< 0.001Note: POCI incidence and 95% confidence intervals were estimated using the Wilson score method. OR compares each surgery type against the remainder of the cohort. The “Other” procedure category was excluded from this surgical-subgroup analysis in accordance with the revised objective focusing on the four predefined surgical categories POCI incidence and odds ratios stratified by the four predefined surgery types Note: POCI incidence and 95% confidence intervals were estimated using the Wilson score method. OR compares each surgery type against the remainder of the cohort. The “Other” procedure category was excluded from this surgical-subgroup analysis in accordance with the revised objective focusing on the four predefined surgical categories One-year mortality and survival analysis In view of the informative censoring structure of INSPIRE, the primary survival comparison was a pre-specified fixed-endpoint analysis. One-year all-cause mortality was 14.8% in the POCI group (41 of 277) versus 5.8% in controls (868 of 14,861), corresponding to an OR of 2.83 (95% CI: 2.02–3.96; p < 0.001). Thirty-day mortality was too infrequent in this predominantly elective cohort to support a stable comparison (0% vs. 0.3%), so the 1-year endpoint was adopted as the primary survival measure. For descriptive purposes, Kaplan-Meier curves and the corresponding log-rank test are provided in Supplementary Figure S5. Supporting perioperative biomarker analyses Pre-to-postoperative trajectories for six key biomarkers are illustrated in Supplementary Figure S6. Both groups exhibited the expected postoperative changes, including reductions in albumin and hemoglobin and increases in creatinine. None of the evaluated delta values reached statistical significance. Accordingly, perioperative biomarker changes were treated as secondary exploratory observations and were not considered central results. Vascular comorbidity-count gradient The bar chart in Fig. 7 displays POCI incidence according to the cumulative number of vascular comorbidities (0, 1, 2, or ≥ 3). A near-monotonic increase in incidence was observed across increasing comorbidity counts, consistent with the univariate odds ratios for individual vascular comorbidities and supporting the clinical usefulness of a simple additive vascular-risk gradient. Fig. 7Age- and comorbidity-stratified POCI incidence A Heatmap displaying POCI incidence (%) within cells defined by six age strata (rows) and four surgical categories (columns). Cell values are annotated to one decimal place. The color scale ranges from yellow (low incidence) to red (high incidence) proportional to the cell rate. Empty cells indicate strata without eligible patients B Bar chart showing POCI incidence (%) stratified by the cumulative number of vascular comorbidities (0, 1, 2, or ≥ 3). Error bars represent 95% Wilson confidence intervals. Patient counts for each stratum are annotated above the bars Age- and comorbidity-stratified POCI incidence A Heatmap displaying POCI incidence (%) within cells defined by six age strata (rows) and four surgical categories (columns). Cell values are annotated to one decimal place. The color scale ranges from yellow (low incidence) to red (high incidence) proportional to the cell rate. Empty cells indicate strata without eligible patients B Bar chart showing POCI incidence (%) stratified by the cumulative number of vascular comorbidities (0, 1, 2, or ≥ 3). Error bars represent 95% Wilson confidence intervals. Patient counts for each stratum are annotated above the bars Additional CAR analyses Additional CAR analyses are presented in Supplementary Figure S7. Preoperative CAR, postoperative CAR, and ΔCAR-ΔNLR patterns showed substantial overlap between groups and did not support CAR dynamics as an independent discriminator of POCI. Machine learning model performance metrics Full cross-validation performance metrics for all three models are presented in Table 7. The gradient boosting classifier produced a comparable AUC-ROC of 0.801 but an extremely low Brier score of 0.018, attributable to the near-zero predicted probabilities assigned to virtually all patients at the cross-validation operating point, with a resulting sensitivity of only 1.4% despite near-perfect specificity (99.9%). The negative predictive value exceeded 99% for logistic regression and random forest, indicating reliable identification of low-risk patients. Table 7Machine Learning Model Performance for POCI Prediction (5-Fold Cross-Validation)Prediction modelAUC-ROCAUC-PRSensitivitySpecificityPPVNPVBrierLogistic regression (L2)0.8190.11865.3%80.9%6.0%99.2%0.158Random forest0.8060.08157.4%86.9%7.6%99.1%0.129Gradient boosting0.8010.0891.4%99.9%16.7%98.2%0.018Note: All metrics were derived from five-fold stratified cross-validation. AUC-PR is emphasized because POCI was rare (1.83%). Primary input features included age per 10-year increment, sex, ASA ≥ 3, emergency surgery, operation duration, hypertension, diabetes, chronic kidney disease, angina/coronary artery disease, atrial fibrillation, heart failure, peripheral arterial disease, vascular comorbidity count, log(NLR), log(SII), log(CAR), relative MAP drop, vasopressor use, and blood product use. Fibrinogen-derived variables and the cardioembolic-prone composite were excluded from the primary machine-learning feature matrix. Missing values were imputed by feature median; continuous features were standardized by z-scorePPV Positive predictive value, NPV Negative predictive value Table 7Machine Learning Model Performance for POCI Prediction (5-Fold Cross-Validation)Prediction modelAUC-ROCAUC-PRSensitivitySpecificityPPVNPVBrierLogistic regression (L2)0.8190.11865.3%80.9%6.0%99.2%0.158Random forest0.8060.08157.4%86.9%7.6%99.1%0.129Gradient boosting0.8010.0891.4%99.9%16.7%98.2%0.018Note: All metrics were derived from five-fold stratified cross-validation. AUC-PR is emphasized because POCI was rare (1.83%). Primary input features included age per 10-year increment, sex, ASA ≥ 3, emergency surgery, operation duration, hypertension, diabetes, chronic kidney disease, angina/coronary artery disease, atrial fibrillation, heart failure, peripheral arterial disease, vascular comorbidity count, log(NLR), log(SII), log(CAR), relative MAP drop, vasopressor use, and blood product use. Fibrinogen-derived variables and the cardioembolic-prone composite were excluded from the primary machine-learning feature matrix. Missing values were imputed by feature median; continuous features were standardized by z-scorePPV Positive predictive value, NPV Negative predictive value Machine Learning Model Performance for POCI Prediction (5-Fold Cross-Validation) Note: All metrics were derived from five-fold stratified cross-validation. AUC-PR is emphasized because POCI was rare (1.83%). Primary input features included age per 10-year increment, sex, ASA ≥ 3, emergency surgery, operation duration, hypertension, diabetes, chronic kidney disease, angina/coronary artery disease, atrial fibrillation, heart failure, peripheral arterial disease, vascular comorbidity count, log(NLR), log(SII), log(CAR), relative MAP drop, vasopressor use, and blood product use. Fibrinogen-derived variables and the cardioembolic-prone composite were excluded from the primary machine-learning feature matrix. Missing values were imputed by feature median; continuous features were standardized by z-score PPV Positive predictive value, NPV Negative predictive value Discussion In this retrospective cohort of 15,138 patients who underwent major thoracic or abdominopelvic surgery at a large academic center in South Korea, POCI occurred in 277 individuals, yielding an overall incidence of 1.83% (95% CI: 1.63–2.06%). This estimate is at the upper end of the 0.1–1.9% range reported in studies of non-cardiac, non-neurological surgery based on Western administrative databases [2, 6, 17]. But it is consistent with the higher rates observed in series restricted to thoracic and abdominal procedures and in studies reporting outcomes from older cohorts with higher vascular comorbidity burdens [7, 16]. Several factors may account for this distribution in the present data: the mean age of the POCI group was 70 years and 80.9% were aged ≥ 65 years, consistent with the strong age-dependency of perioperative stroke risk. The INSPIRE cohort also has a higher proportion of oncological and upper gastrointestinal resections than many general surgical registries, which contributes a patient profile with elevated pre-existing cardiovascular risk and physiological reserve limitations. The strongest unadjusted comorbidity signal for POCI was AF. New-onset AF in the peri-operative context is of particular concern. Prior registry data indicate that patients who develop AF during a non-cardiac surgical admission face approximately a threefold increase in stroke risk over the subsequent 12 months compared with patients in sinus rhythm [8], and that the risk of thromboembolism attributable to peri-operative AF may persist even after the arrhythmia has resolved [9]. The high prevalence of pre-existing AF (10.1%) in the POCI group, more than ten times that in controls, underscores the need for structured preoperative cardiovascular review and individualized perioperative anticoagulation planning in patients undergoing major abdominal and thoracic procedures, consistent with the 2020 SNACC guidance for patients at high risk of perioperative stroke [7] and the 2024 multisociety guideline for perioperative cardiovascular management before non-cardiac surgery [18]. Multivariable adjustment provided an important qualification to the univariate comorbidity findings [18]. After mutual adjustment for age and other covariates, hypertension (aOR 5.92), atrial fibrillation (aOR 4.21), chronic kidney disease (aOR 2.47), peripheral arterial disease (aOR 2.45), advancing age (aOR 1.65 per decade), and emergency surgery (aOR 1.57) remained independently associated with POCI, whereas coronary artery disease, heart failure, and diabetes attenuated to non-significance. Advancing age, hypertension, and atrial fibrillation are the most readily identifiable preoperative flags, while chronic kidney disease and peripheral arterial disease identify a broader vascular-risk phenotype [19]. Because these estimates derive from an observational design, they should be interpreted as associations rather than causal effects. NLR, SII, and FAR were higher in POCI patients than controls in unadjusted comparisons. However, neither NLR nor SII was independently associated with POCI after adjustment, and FAR was excluded from adjusted models because it was missing in 27.8% of patients. These indices therefore did not add demonstrable predictive value beyond established risk factors in this cohort and should be interpreted cautiously [10–14]. Prospective validation with complete laboratory ascertainment would be required before these indices could be recommended for perioperative POCI risk stratification. The hemodynamic findings were also secondary and should be interpreted cautiously. Baseline MAP, end-of-operation MAP, relative MAP decrease, and vasopressor exposure did not show a consistent pattern of greater hypotension in POCI patients. The lower observed vasopressor use in the POCI group (30.0% vs. 36.6%) is most plausibly explained by confounding from surgical profile, anesthetic approach, operation duration, and case complexity rather than by a protective effect. Because the analysis was limited to coarse MAP summaries and vasopressor exposure, it cannot exclude associations involving cumulative time or area under MAP thresholds, postoperative blood-pressure variability, or cerebral perfusion surrogates. The machine-learning analysis was retained as a compact supporting analysis. Although L2-regularised logistic regression achieved an AUC-ROC of 0.819, the AUC-PR was only 0.118 and the positive predictive value remained low. Tree-based models did not improve practical utility; in particular, gradient boosting produced near-complete specificity at the expense of clinically unacceptable sensitivity. These results indicate that machine-learning scores should not be used as standalone screening tools in this low-prevalence setting. The adjusted clinical model and the vascular comorbidity-count gradient provide a simpler and more interpretable approach for focusing postoperative neurological vigilance on patients with higher baseline vascular risk. The surgical category analysis revealed meaningful heterogeneity across the four predefined surgical groups. Hepatobiliary surgery carried the lowest POCI rate (0.71%; OR 0.36 relative to the rest of the cohort), whereas lung and gastrointestinal procedures had numerically higher rates (1.91% and 1.98%, respectively) without statistically significant deviation from the overall cohort incidence. The low POCI rate in hepatobiliary surgery may reflect patient selection, lower baseline vascular comorbidity burden among elective hepatobiliary candidates at this tertiary center, and procedure-specific differences in perioperative physiology. Esophageal procedures, despite being among the most physiologically demanding operations in the cohort, had an intermediate POCI incidence (1.39%), possibly reflecting the comparatively lower vascular comorbidity burden in patients undergoing oncological esophagectomy at this center, although unmeasured confounders including surgical approach and tumor-related pro-coagulant states cannot be excluded. POCI was also associated with worse short- and longer-term outcomes. In-hospital mortality, ICU admission, prolonged hospital stay, and all-cause mortality recorded during the observation period were all more frequent in the POCI group. One-year all-cause mortality was significantly higher in the POCI group (14.8% vs. 5.8%; OR 2.83; p < 0.001). The Kaplan-Meier log-rank comparison, by contrast, was non-significant (p = 0.25) because of informative censoring. We therefore base the survival conclusion on the fixed-endpoint analysis rather than on the survival curves. The excess 1-year mortality is consistent with POCI being a marker of severe systemic vascular disease and perioperative vulnerability, although the observational design precludes causal inference. It should be acknowledged that the censoring structure of INSPIRE, in which patients alive at discharge are censored at the time of hospital discharge rather than at a formal administrative censoring date, limits interpretation of the Kaplan-Meier curves as population-level actuarial estimates. Limitations This study has several limitations. First, POCI ascertainment was based on ICD-10 diagnostic coding rather than neuroimaging adjudication, which may have led to under-coding of mild events and occasional misclassification of pre-existing silent infarcts. Second, the observational design precludes causal interpretation, and residual confounding by unmeasured variables such as preoperative anticoagulation status, antiplatelet therapy, anesthetic technique, intraoperative blood loss, and surgical approach remains possible. Third, INSPIRE is a single-centre tertiary-hospital dataset from South Korea, which limits generalizability to community hospitals and to broader Korean, Asian, or non-Asian surgical populations. Fourth, informative censoring precludes definitive long-term actuarial survival estimation. Therefore, the fixed 1-year endpoint is more interpretable than the supplementary Kaplan-Meier curves. Fifth, covert cerebral infarctions were not captured because brain MRI data were unavailable. Sixth, the machine-learning models lacked external validation and had limited screening utility because of severe class imbalance, low AUC-PR, and low positive predictive value. Seventh, fibrinogen and FAR were missing in more than 20% of patients, limiting inference for fibrinogen-derived indices. Finally, inflammatory indices, perioperative hemodynamics, and machine-learning were secondary supporting analyses around the primary aims of incidence estimation and adjusted risk-factor identification. Conclusions POCI following major thoracic and abdominopelvic surgery occurred at an overall rate of 1.83% in this single-center cohort and was associated with higher in-hospital mortality, more frequent ICU admission, and higher 1-year all-cause mortality. After multivariable adjustment, advancing age, hypertension, atrial fibrillation, chronic kidney disease, peripheral arterial disease, and emergency surgery were independently associated with POCI, with age, hypertension, and atrial fibrillation the most clinically tractable preoperative flags. The composite inflammatory indices (NLR, SII, FAR) differed between groups in unadjusted comparisons but were not independently associated with POCI after adjustment. Likewise, although a logistic-regression model reached an AUC-ROC of 0.819, its low positive predictive value in this low-prevalence setting limits its standalone screening utility. These results provide a population-specific incidence benchmark and support structured preoperative cardiovascular risk review, with prospective external validation required before any composite-index or machine-learning tool could be recommended for clinical use. Supplementary Information Supplementary Material 1: Supplementary Figure S1. Cohort derivation flow diagram (full exclusion algorithm including ICD-10 code-based filtering and temporal admission-surgery window criteria). Supplementary Material 1: Supplementary Figure S1. Cohort derivation flow diagram (full exclusion algorithm including ICD-10 code-based filtering and temporal admission-surgery window criteria). Supplementary Material 2: Supplementary Figure S2. Precision-recall curves for all three machine-learning classifiers (L2 logistic regression, random forest, gradient boosting) with five-fold cross-validation confidence bands. Supplementary Material 2: Supplementary Figure S2. Precision-recall curves for all three machine-learning classifiers (L2 logistic regression, random forest, gradient boosting) with five-fold cross-validation confidence bands. Supplementary Material 3: Supplementary Figure S3. Permutation-based and Gini feature-importance analysis for the random-forest classifier, showing the top 15 predictive features with publication-ready clinical labels rather than raw variable names. Supplementary Material 3: Supplementary Figure S3. Permutation-based and Gini feature-importance analysis for the random-forest classifier, showing the top 15 predictive features with publication-ready clinical labels rather than raw variable names. Supplementary Material 4: Supplementary Figure S4. Decision-curve analysis (DCA) for the three machine-learning classifiers predicting postoperative cerebral infarction (POCI) in the INSPIRE cohort (n = 15,138). Supplementary Material 4: Supplementary Figure S4. Decision-curve analysis (DCA) for the three machine-learning classifiers predicting postoperative cerebral infarction (POCI) in the INSPIRE cohort (n = 15,138). Supplementary Material 5: Supplementary Figure S5. Kaplan-Meier survival analysis (descriptive). (A) Kaplan-Meier overall survival curves for the POCI (gold, n = 277) and control (teal, n = 14,861) groups. Survival time is defined as the interval from hospital admission to all-cause death (national death-registry linkage), expressed in months. Patients alive at last observation are censored at hospital discharge. Shaded bands represent 95% confidence intervals derived from the Nelson-Aalen cumulative hazard estimator. Tick marks indicate censored observations. The at-risk table beneath the plot shows the number of patients with follow-up time exceeding each threshold. The annotated p value is from the log-rank test. Owing to the censoring structure of INSPIRE, the curves are descriptive and should not be interpreted as population-level actuarial survival estimates. (B) Kaplan-Meier survival curves stratified by surgical category within the POCI subgroup. Supplementary Material 5: Supplementary Figure S5. Kaplan-Meier survival analysis (descriptive). (A) Kaplan-Meier overall survival curves for the POCI (gold, n = 277) and control (teal, n = 14,861) groups. Survival time is defined as the interval from hospital admission to all-cause death (national death-registry linkage), expressed in months. Patients alive at last observation are censored at hospital discharge. Shaded bands represent 95% confidence intervals derived from the Nelson-Aalen cumulative hazard estimator. Tick marks indicate censored observations. The at-risk table beneath the plot shows the number of patients with follow-up time exceeding each threshold. The annotated p value is from the log-rank test. Owing to the censoring structure of INSPIRE, the curves are descriptive and should not be interpreted as population-level actuarial survival estimates. (B) Kaplan-Meier survival curves stratified by surgical category within the POCI subgroup. Supplementary Material 6: Supplementary Figure S6. Perioperative biomarker trajectories. Line graphs (mean ± standard error) illustrating the change from the preoperative to the first postoperative measurement for six biomarkers, albumin, CRP, fibrinogen, hemoglobin, creatinine, and NLR, in a random subsample of controls (teal, up to n = 800) and all POCI patients (gold, n = 277). CRP and NLR are displayed on a log(1 + x) scale. Annotated p values in each panel correspond to the between-group comparison of the delta value (postoperative minus preoperative) by Mann-Whitney U test. CRP, C-reactive protein; NLR, neutrophil-to-lymphocyte ratio. Supplementary Material 6: Supplementary Figure S6. Perioperative biomarker trajectories. Line graphs (mean ± standard error) illustrating the change from the preoperative to the first postoperative measurement for six biomarkers, albumin, CRP, fibrinogen, hemoglobin, creatinine, and NLR, in a random subsample of controls (teal, up to n = 800) and all POCI patients (gold, n = 277). CRP and NLR are displayed on a log(1 + x) scale. Annotated p values in each panel correspond to the between-group comparison of the delta value (postoperative minus preoperative) by Mann-Whitney U test. CRP, C-reactive protein; NLR, neutrophil-to-lymphocyte ratio. Supplementary Material 7: Supplementary Figure S7. C-reactive-protein-to-albumin ratio (CAR) dynamics and ΔCAR-ΔNLR relationship. (A, B) Violin plots overlaid with box-and-whisker plots comparing preoperative (A) and postoperative (B) CAR distributions between control (teal) and POCI (gold) patients, displayed on a log(1 + x) scale. Significance brackets are from Mann-Whitney U tests; annotated p values correspond to these comparisons. (C) Scatter plot of log(1 + ΔNLR) against log(1 + ΔCAR) for a random subsample of controls (teal, up to n = 600) and all POCI patients (gold), where Δ denotes the postoperative minus preoperative change. CAR, CRP-to-albumin ratio; NLR, neutrophil-to-lymphocyte ratio. Supplementary Material 7: Supplementary Figure S7. C-reactive-protein-to-albumin ratio (CAR) dynamics and ΔCAR-ΔNLR relationship. (A, B) Violin plots overlaid with box-and-whisker plots comparing preoperative (A) and postoperative (B) CAR distributions between control (teal) and POCI (gold) patients, displayed on a log(1 + x) scale. Significance brackets are from Mann-Whitney U tests; annotated p values correspond to these comparisons. (C) Scatter plot of log(1 + ΔNLR) against log(1 + ΔCAR) for a random subsample of controls (teal, up to n = 600) and all POCI patients (gold), where Δ denotes the postoperative minus preoperative change. CAR, CRP-to-albumin ratio; NLR, neutrophil-to-lymphocyte ratio. Supplementary Material 8: Supplementary Table S1. Data completeness report: proportion of non-missing values for all key variables used in the primary analysis (n = 15,138). Supplementary Table S2. Sensitivity analysis: univariate odds ratios for POCI restricted to patients with complete data across all machine-learning predictor variables (n = 10,694), compared with the main imputed-data cohort. Supplementary Table S3. Variance-inflation factors for the final multivariable logistic-regression model. Supplementary Material 8: Supplementary Table S1. Data completeness report: proportion of non-missing values for all key variables used in the primary analysis (n = 15,138). Supplementary Table S2. Sensitivity analysis: univariate odds ratios for POCI restricted to patients with complete data across all machine-learning predictor variables (n = 10,694), compared with the main imputed-data cohort. Supplementary Table S3. Variance-inflation factors for the final multivariable logistic-regression model. Supplementary Material 9. Supplementary Material 9.