Machine-Learning Prediction of Contrast-Induced Encephalopathy After Neurointerventional Procedures

Summary

An exploratory machine-learning (ML) framework using perioperative data was evaluated to predict contrast-induced encephalopathy (CIE) after neurointerventional procedures. The Naive Bayes model showed the most balanced descriptive performance, and contrast-to-eGFR ratio (CGR) was identified as an important candidate predictor for individualized risk assessment.

Abstract

CIE is a rare yet severe complication following neurointerventional procedures, for which reliable noninvasive predictive tools remain lacking. This study aimed to develop and validate a ML model for predicting CIE using perioperative clinical and procedural data, and to evaluate the predictive value of the CGR for CIE. This was a single-center retrospective study that consecutively enrolled 161 patients who underwent neurointerventional procedures in the Department of Neurosurgery at the Northern Theater General Hospital between January 2024 and December 2025. Candidate perioperative predictors of CIE were identified using univariate analysis combined with LASSO regression. Based on the selected variables, five ML models—Naive Bayes, support vector machine (SVM), k-nearest neighbors (KNN), LightGBM, and multilayer perceptron (MLP)—were trained and optimized. Model performance was comprehensively evaluated using the area under the receiver operating characteristic curve (AUC), and decision curve analysis (DCA). Among the evaluated models, the Naive Bayes model showed the most favorable descriptive performance in the internal test set, with an AUC of 0.952 (95% CI: 0.843–1.000), a sensitivity of 100%, and a specificity of 90.3%; however, these metrics should be interpreted cautiously because the dataset was markedly imbalanced and the internal test cohort was small and contained only a very limited number of CIE events. The DCA results suggested potential clinical net benefit across selected threshold probabilities. Feature-importance analysis indicated that CGR was among the highest-ranked candidate predictors associated with CIE risk in this dataset. The Naive Bayes model evaluated in this study may provide a preliminary risk-stratification framework for perioperative assessment of CIE after neurointerventional procedures. CGR emerged as an important predictive feature and may aid individualized risk stratification. Further external validation is warranted before clinical application.

Introduction

Over the past two decades, minimally invasive interventional techniques have fundamentally transformed the treatment paradigm for cerebrovascular diseases¹. Neurointerventional procedures—with their advantages of high efficacy and minimal invasiveness—have become important treatment options for selected cerebrovascular disorders². However, as the complexity of interventional procedures increases, so do the doses of contrast agents used during procedures, the injection pressures, and the duration of contrast-agent exposure. Consequently, contrast-agent–related complications have gradually come under increasing scrutiny^3,4,5,6,7. CIE, an acute, transient neurological dysfunction induced by iodinated contrast media (ICM), has been increasingly recognized as a clinically relevant complication, particularly in complex neurointerventional procedures. The typical clinical manifestations of CIE usually appear within minutes to 24 h after contrast-agent exposure and include alterations in consciousness (such as somnolence, delirium, or coma), cortical blindness, focal neurological deficits (such as hemiparesis or aphasia), and seizures^4,5,6. Although most cases of CIE are self-limiting and symptoms typically resolve within 48 to 72 h, some patients may experience persistent neurological impairment. Severe cases may even lead to death due to massive cerebral edema^4,7. This unpredictability of outcomes underscores the urgent need to develop a reliable tool for early risk prediction during the perioperative period.

As the volume of clinical data continues to increase, traditional statistical methods are increasingly inadequate when dealing with high-dimensional medical data that exhibit multicollinearity. Moreover, univariate analyses or simple logistic regression models often fail to capture the complex, nonlinear interactions among variables. ML offers a new dimension for addressing these challenges⁸. ML algorithms—such as SVM, gradient boosting machines (LightGBM), and MLP—can automatically identify latent patterns in high-dimensional feature spaces and construct nonlinear models with powerful predictive capabilities^8,9,10,11. However, while pursuing prediction accuracy, the “black-box” nature of these algorithms, characterized by their lack of interpretability, often raises concerns regarding trust in medical decision-making^12,13. In contrast, the Naive Bayes algorithm, with its simple probabilistic logic and inherent interpretability, demonstrates unique advantages in small-sample clinical studies^14,15. By evaluating the posterior probability contribution of each predictor variable to the outcome event, the Naive Bayes model not only ensures predictive performance but also provides clinicians with intuitive decision support insights. Furthermore, to quantitatively assess the imbalance between contrast agent load and individual clearance capacity, this study innovatively introduces the contrast agent-to-eGFR ratio (CGR) as a core feature. For patients with lower renal functional reserve, even moderate doses of contrast agents can result in extremely high CGR values, which may increase susceptibility to CIE. CGR may provide a clinically intuitive composite measure integrating contrast-agent burden and renal-clearance capacity, although its incremental predictive value requires further validation. By incorporating this integrated biomarker, this study aimed to develop a predictive framework that more closely reflects the pathophysiological reality.

In summary, this study aimed to develop and validate an ML-based CIE prediction model using perioperative clinical and procedural variables¹⁶. The specific objectives include: (1) introducing and evaluating CGR as an exploratory composite index; (2) constructing and comparing the performance of five ML models—Naive Bayes, SVM, KNN, LightGBM, and MLP; and (3) identifying the optimal model. Given the limited number of CIE events, the model comparison was considered exploratory and hypothesis-generating. Ultimately, this study may provide a preliminary risk-stratification framework for CIE in neurointerventional practice, but it should not be used as a standalone clinical decision-making tool without external validation. The conceptual basis of CGR and its relationship with contrast exposure and renal clearance are illustrated in Figure 1A. The overall study workflow, including data preprocessing, train/test splitting, feature selection, and model development, is summarized in Figure 1B. The comparative evaluation framework of the five ML models is presented in Figure 1C.

Protocol

The study protocol was reviewed and approved by the Ethics Committee of the General Hospital of Northern Theater Command (Approval No. Ethics Y (2026) 75). Although the clinical data were derived from patients treated between January 2024 and December 2025, the present study was conducted as a retrospective analysis of existing clinical records. The ethics approval obtained in 2026 covered retrospective data extraction, de-identification, analysis, and publication of these previously collected clinical data. No prospective intervention or patient enrollment was performed before ethics approval. The requirement for written informed consent was waived because of the retrospective observational study design. All procedures were conducted in accordance with the Declaration of Helsinki, and all clinical data were de-identified before analysis. The present analysis was conducted as part of an approved retrospective neurointerventional clinical-data research project. The materials and equipment required for the procedures described below are summarized in the Table of Materials.

Patient enrollment and CIE determination

Potentially eligible patients were identified by querying the electronic medical record system for patients treated in the Department of Neurosurgery at the General Hospital of Northern Theater Command between January 2024 and December 2025. The initial search was restricted by admission date, department, cerebrovascular disease diagnosis, and neurointerventional procedure records. The detailed patient screening and cohort allocation workflow is illustrated in Figure 2. Patients were eligible for inclusion if they met all the following criteria: (1) age between 30 and 80 years; (2) diagnosis of cerebrovascular disease requiring neurointerventional evaluation or treatment; (3) availability of complete clinical records and laboratory-test results; and (4) postoperative head CT examination performed within 3 days after surgery.

Patients were excluded if they had incomplete clinical or imaging data, did not undergo endovascular intervention, had severe preoperative neurological deficits (modified Rankin Scale score ≥4), demonstrated acute cerebral infarction on preoperative diffusion-weighted imaging, had chronic kidney disease stage G4 or higher according to the Kidney Disease: Improving Global Outcomes (KDIGO) criteria, or had severe cardiovascular disease. The inclusion and exclusion criteria were applied sequentially. After the initial electronic search, each candidate record was manually reviewed to confirm eligibility and to document the reason for exclusion when applicable. After screening, eligible patients were randomly assigned to training and test cohorts at an 8:2 ratio, including a training cohort (n = 128) and a test cohort (n = 33). Stratified random sampling was used to maintain a consistent distribution of CIE and non-CIE cases between the two cohorts.

The diagnosis of CIE was primarily based on the temporal association between symptom onset and cerebrovascular intervention procedures, combined with exclusionary imaging assessment. Two independent cerebrovascular intervention specialists reviewed the postoperative clinical records, symptom-onset time, neurological manifestations, and postoperative imaging findings for all suspected cases. Disagreements were resolved through discussion until consensus was reached. Patients were considered compatible with CIE if newly developed cortical blindness, altered consciousness (including somnolence, delirium, or coma), seizure, or focal neurological deficits such as hemiparesis or aphasia occurred within 24 h after completion of the neurointerventional procedure.

After symptom onset, alternative causes capable of producing similar neurological manifestations were excluded by reviewing postoperative CT or MRI images together with the clinical course. Postoperative head CT was performed using the institutional standard non-contrast cranial CT protocol. The essential acquisition parameters included tube voltage of 80 kVp, tube current of 260 mA, slice thickness of 0.5 mm, and standard axial reconstruction. Images were reviewed on brain and bone windows by experienced clinicians. Acute intracranial hemorrhage, subarachnoid hemorrhage, intracerebral hemorrhage, newly developed large-area cerebral infarction, seizure-related changes, infection, and metabolic disturbances were considered during differential diagnosis. Post-symptom CT examinations typically demonstrated focal or diffuse cerebral edema and high-density shadows involving the cortex, subcortical regions, or subarachnoid space, mimicking hemorrhagic lesions. MRI findings usually demonstrate cortical swelling, particularly involving the temporoparietal occipital cortex. In some patients, however, CT or MRI examinations showed no obvious abnormalities despite compatible clinical manifestations.

Perioperative variable acquisition and CGR construction

General demographic data, including age, sex, and body weight, were extracted from structured fields in the electronic medical record system. Clinical history variables, including smoking history, alcohol consumption history, hypertension, diabetes mellitus, and coronary heart disease, were extracted from admission notes, past medical history records, and discharge diagnoses, and were cross-checked for consistency. Laboratory data were extracted from the laboratory information system. For each patient, the first venous blood sample collected after admission and before the neurointerventional procedure was used. The extracted laboratory variables included serum creatinine, estimated glomerular filtration rate (eGFR), total cholesterol, and triglyceride levels. eGFR was calculated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) creatinine-based equation and was obtained from the hospital laboratory information system.

Because procedural complexity, contrast-agent exposure, and lesion location may influence blood–brain barrier disruption and contrast-agent retention during neurointerventional procedures, perioperative procedural variables were also incorporated into the predictive framework. Procedure-related variables included lesion location, procedure type, procedural duration, contrast-agent type, and total contrast volume. Lesion location and procedure type were determined from operative reports and angiographic records. Procedural duration was defined as the time from arterial puncture to completion of the neurointerventional procedure. Total contrast volume was defined as the cumulative volume of iodinated contrast agent administered from the beginning to the end of the procedure. The contrast agents used in this study were nonionic ICA, including iodixanol (320 mg of iodine/mL), administered according to institutional clinical practice. All neurointerventional procedures included in this study were performed under digital subtraction angiography guidance by an experienced neurointerventional team at the study center. All procedures were performed using the transfemoral arterial approach. During the intervention, patients received intraoperative anticoagulation with heparin, and activated clotting time was maintained within the target range of 250–300 s. Vital signs were continuously monitored throughout the procedure.

Therapeutic interventions, including aneurysm embolization, angioplasty, and stent placement, were performed according to lesion characteristics and operator judgment. Appropriate microcatheters, coils, stents, or balloons were selected according to procedural requirements. During angiographic imaging and intervention, contrast agents were administered either through high-pressure injection or manual constant-rate infusion. Contrast agent type and total contrast volume were extracted from procedural records and cross-checked with anesthesia or nursing records when available. Records with inconsistent or incomplete contrast-volume information were reviewed manually before inclusion in the final analytical dataset.

All procedure-related variables, including procedural duration, therapeutic intervention type, device usage, and contrast-agent administration records, were systematically reviewed and verified for completeness and consistency before analysis. To quantitatively evaluate the imbalance between contrast-agent burden and renal-clearance capacity, the CGR was established as a core predictive variable and calculated according to the following equation:

(1)

The total contrast volume was recorded in mL, and eGFR was recorded in mL/min/1.73 m². The eGFR value used for CGR calculation was the preoperative eGFR value obtained from the first venous blood sample collected after admission and before the neurointerventional procedure. For each patient, CGR was calculated after verifying both total contrast volume and eGFR. The numerator was total contrast volume in mL, and the denominator was eGFR in mL/min/1.73 m². The same calculation rule was applied to all patients before model development.

Feature selection and ML workflow

To ensure analytical robustness, all clinical and laboratory variables underwent systematic quality control procedures before analysis. The missing data proportion was calculated for each candidate variable before model development. In the final analytical dataset, no missing values were observed among the included predictors or outcome labels; therefore, no variable was excluded because of missingness, and multiple imputation was not required. The detailed missing data summary is provided in Supplementary Table 1.

All candidate predictor variables were initially evaluated using univariate analyses to assess their associations with the occurrence of CIE. Statistical methods were selected according to the distribution characteristics of the data. Variables with normal distribution were expressed as mean ± standard deviation and compared using the independent samples t-test. Variables with non-normal distribution were expressed as median (interquartile range, IQR) and analyzed using the Mann–Whitney U test. Categorical variables were presented as frequencies and percentages and analyzed using the chi-square test or Fisher’s exact test. Statistical significance was defined as p < 0.05.

The entire dataset was randomly divided into training and test cohorts at an 8:2 ratio before model development. To minimize overfitting and reduce multicollinearity, least absolute shrinkage and selection operator (LASSO) regression combined with 10-fold cross-validation was applied for feature selection. Feature selection, hyperparameter tuning, and model development were performed using the training cohort only to avoid data leakage. The internal test cohort was not used during imputation-parameter estimation, feature selection, hyperparameter tuning, or model training, and was used only once for final performance evaluation. Hyperparameter optimization was conducted using a grid search strategy with 10-fold cross-validation within the training cohort. Specifically, the training cohort was randomly divided into 10 mutually exclusive subsets. During each iteration, nine subsets were used for model training, and the remaining subset was used for validation. This process was repeated 10 times to ensure that each subset served as the validation cohort exactly once.

All computational analyses were performed in a Python environment. ML model development and evaluation were performed using scikit-learn. The analysis scripts used for data preprocessing, feature selection, model training, performance evaluation, and DCA were provided as supplementary code files. For model implementation, the selected features from LASSO regression were used as input variables for all candidate classifiers. Continuous variables were standardized using z-score normalization within the training cohort, and the same scaling parameters were applied to the internal test cohort. Categorical variables were encoded according to the predefined coding scheme, and the variable definitions and coding scheme are provided in Supplementary Table 2. Five classifiers were implemented: Gaussian Naive Bayes, SVM, KNN, LightGBM, and MLP. For SVM, the kernel type, regularization parameter C, and kernel coefficient gamma were tuned. For KNN, the number of neighbors and the distance-weighting strategy were tuned. For LightGBM, the number of estimators, learning rate, maximum tree depth, and number of leaves were tuned. For MLP, the hidden-layer structure, activation function, regularization parameter, and learning rate were tuned. Gaussian Naive Bayes was implemented using the specified variance smoothing setting. The optimal hyperparameter combination for each model was selected according to cross-validation performance within the training cohort. The final optimized hyperparameter combinations selected for each model are summarized in Supplementary Table 3. After hyperparameter selection, each final model was refitted on the full training cohort and evaluated once in the internal test cohort.

Comparisons of model performance

Based on the selected predictors, five ML models were developed and compared. Model performance was evaluated in both the training and test cohorts using the area under the receiver operating characteristic curve (AUC), 95% confidence interval (95% CI), sensitivity, specificity, and accuracy. Receiver operating characteristic curves were generated to compare the discriminative performance of the different models. The optimal classification threshold for each model was determined exclusively within the training cohort using the Youden index. The derived threshold was then fixed and applied unchanged to the internal test cohort for calculating threshold-dependent performance metrics.

Decision curve analysis

DCA was performed by calculating the net benefit of each model across a range of threshold probabilities and comparing the model-based strategy with two default strategies: treating all patients and treating no patients. DCA was selected because it allows simultaneous assessment of discrimination performance and potential clinical applicability in perioperative decision-making settings. To facilitate reproducibility, the complete workflow was performed in the following sequence: patient identification from the electronic medical record system, application of inclusion and exclusion criteria, CIE adjudication by two independent specialists, extraction and cross-checking of demographic, clinical history, laboratory, and procedure-related variables, CGR calculation, missing-data assessment and training-test cohort split, LASSO-based feature selection within the training cohort, hyperparameter tuning by 10-fold cross-validation, final model refitting in the full training cohort, internal test-cohort evaluation, ROC-based performance assessment, Youden-index threshold determination, and DCA.

Results

Patient enrollment and CIE determination

A total of 161 patients who underwent neurointerventional procedures were included in this study, among whom 12 (7.5%) developed CIE, while 149 (92.5%) did not. The detailed patient screening and grouping process is shown in Figure 2. The diagnosis of CIE was established based on the temporal association between symptom onset and neurointerventional procedures, combined with exclusionary imaging findings. Representative neuroimaging findings of CIE are presented in Figure 3A–D.

The baseline demographic, clinical, laboratory, and procedural characteristics of the two groups are summarized in Table 1. No significant differences were observed between the CIE and non-CIE groups in terms of age, sex distribution, or body weight (all p > 0.05). Similarly, the prevalence of common comorbidities, including hypertension, diabetes mellitus, and coronary heart disease, did not differ significantly between the two groups (all p > 0.05). Lifestyle factors, such as alcohol consumption and smoking status, were also comparable.

In contrast, several laboratory and procedure-related variables showed significant differences between the two groups. Patients in the CIE group had significantly higher serum creatinine levels compared with those in the non-CIE group (median: 85.2 vs 65.3 µmol/L, p = 0.004), while their eGFR was significantly lower (median: 81 vs 102.7 mL/min/1.73 m², p < 0.001). Regarding procedural characteristics, embolization procedures were significantly more frequent in the CIE group than in the non-CIE group (41.7% vs 2.0%, p < 0.001). In addition, posterior circulation lesions were significantly more common among patients who developed CIE (58.3% vs 6.0%, p < 0.001). No significant difference was observed in the type of contrast media used between the two groups (p = 0.982). Furthermore, patients in the CIE group received a significantly higher volume of contrast agent during the procedure (median: 200.0 vs 165.0 mL, p = 0.005) and had longer procedure durations (median: 78.5 vs 50.0 min, p = 0.008). Notably, the CGR, a composite index proposed in this study, was significantly higher in the CIE group compared with the non-CIE group (median: 2.35 vs 1.63, p < 0.001), indicating a strong association between CGR and the occurrence of CIE.

Perioperative variable acquisition and CGR construction

Perioperative demographic, laboratory, and procedural variables were systematically collected and analyzed as described in the protocol. Among these variables, CGR, defined as the ratio of total contrast volume to eGFR, was constructed as a composite indicator reflecting the balance between contrast-agent exposure and renal clearance capacity. CGR was associated with CIE occurrence in this cohort and was included as a candidate predictor for feature-selection and model-development analyses. To compare CGR with its individual components, we performed a comparative ROC analysis using CGR, total contrast volume, and eGFR. CGR showed the highest discriminative ability for CIE, with an AUC of 0.959 (95% CI: 0.924–0.985), compared with total contrast volume (AUC = 0.745, 95% CI: 0.559–0.886) and eGFR analyzed in the inverse risk direction (AUC = 0.865, 95% CI: 0.764–0.951). The AUC difference between CGR and total contrast volume was 0.214, and the AUC difference between CGR and eGFR was 0.093. Detailed results are provided in Supplementary Table 4.

Feature selection and ML workflow

To identify the most relevant predictors while minimizing multicollinearity, LASSO regression with 10-fold cross-validation was applied to the candidate variables. The coefficient profile and cross-validation curves of the LASSO model are presented in Figure 4A and Figure 4B, respectively. The LASSO coefficient ranking of the final model is presented in Figure 5. At the optimal λ value, a subset of variables was retained, including CGR, procedure type, lesion location, triglyceride levels, serum creatinine, eGFR, and several clinical covariates. The quantitative ranking of retained predictors based on LASSO coefficients is provided in Supplementary Table 5.

Comparisons of model performance

Five ML approaches—Naive Bayes, SVM, KNN, LightGBM, and MLP—were developed to predict the risk of CIE. Their performance in the training and test sets is summarized in Table 2, the accuracy comparison of the five models is shown in Figure 6, and the corresponding confusion matrices are provided in Supplementary Figure 1, Supplementary Figure 2, Supplementary Figure 3, Supplementary Figure 4, and Supplementary Figure 5. Because the dataset was markedly imbalanced, with 12 CIE cases and 149 non-CIE cases, accuracy and threshold-based metrics should be interpreted cautiously and in conjunction with sensitivity, specificity, AUC, and confidence intervals. In the training set, all models exhibited good discriminative ability, with AUC values ranging from 0.961–0.999. In the internal test set, the Naive Bayes model showed the highest numerical (AUC = 0.952, 95% CI: 0.843–1.000), followed by LightGBM (AUC = 0.944, 95% CI: 0.822–1.000) and SVM (AUC = 0.935, 95% CI: 0.796–1.000).

At the optimal classification threshold determined by the Youden index, the Naive Bayes model yielded a sensitivity of 100% and a specificity of 90.3%. Because formal pairwise statistical comparisons among model AUCs were not performed, the observed differences among models should be interpreted as descriptive and exploratory. In addition, because the dataset was markedly imbalanced and the internal test cohort was small with only a very limited number of CIE events, threshold-based metrics such as sensitivity and specificity may be unstable and should be interpreted cautiously. The receiver operating characteristic (ROC) curves of the Naive Bayes model in the training and test sets are presented in Figure 7.

Although the MLP model performed well in the training set (AUC = 0.961), its performance declined substantially in the test set (AUC = 0.742), suggesting potential overfitting. Similarly, the KNN model demonstrated limited discriminative ability in the test set (AUC = 0.718) with a wide confidence interval, indicating instability. The Youden index–based strategy was used to determine the optimal classification threshold for threshold-dependent metrics, including sensitivity, specificity, and accuracy. The ROC-AUC values themselves are threshold-independent. Therefore, the reported threshold-dependent performance metrics reflect model behavior at the optimal operating point. The optimal Youden thresholds and corresponding threshold-based metrics are provided in Supplementary Table 6. Overall, the Naive Bayes model showed relatively stable descriptive performance between the training and internal test sets, with minimal numerical decline in AUC.

Decision curve analysis

DCA was performed to evaluate the clinical utility of the models across a range of threshold probabilities. The DCA curves are shown in Figure 8. All models demonstrated positive net benefit over specific ranges of threshold probabilities compared with the treat-all and treat-none strategies. Among the evaluated models, SVM demonstrated the widest range of positive net benefit, whereas the other models also showed potential clinical usefulness within specific threshold ranges.

DATA AVAILABILITY:

The de-identified raw data used to support the main analyses have been uploaded to GitHub and are available at: https://doi.org/10.5281/zenodo.20043331. All direct patient identifiers were removed before data sharing. Because this was a retrospective clinical study based on hospital medical records, access to any additional patient-level information remains restricted by institutional ethical and privacy requirements. Additional de-identified data may be made available from the corresponding author upon reasonable request and with approval from the institutional ethics committee. The statistical analysis scripts, ML model training code, and figure-generation code have been provided as supplementary files to support reproducibility.

Figure 1: Workflow of CIE risk modeling based on perioperative data. (A) Mechanisms of CIE and derivation of the contrast volume-to-eGFR ratio (CGR). (B) Data integration and LASSO-based feature selection. (C) ML model development and comparison, with Naive Bayes showing the highest numerical AUC in the internal test set (AUC ≈ 0.95). Abbreviations: CIE = contrast-induced encephalopathy; CGR = contrast volume-to-eGFR ratio; eGFR = estimated glomerular filtration rate; AUC = area under the curve. The figure was derived from the authors' own clinical data, and the schematic illustration was originally created by the authors using modified Microsoft PowerPoint icons. Please click here to view a larger version of this figure.

Figure 2: Patient selection and cohort allocation. Flowchart illustrating patient screening, eligibility assessment, exclusion criteria, and cohort assignment. A total of 225 patients were initially assessed. After applying inclusion criteria, including age between 30 and 80 years, availability of complete clinical data, and postoperative head CT within 3 days, 190 patients were eligible. Patients were excluded based on acute cerebral infarction, chronic kidney disease stage ≥ G4 (eGFR < 30 mL/min/1.73 m²), severe cardiovascular disease, absence of endovascular treatment, severe neurological deficit (modified Rankin Scale score ≥ 4), or incomplete data. Finally, 161 patients were included and randomly divided into the training cohort (n = 128) and internal test cohort (n = 33) at an 8:2 ratio. Abbreviations: CT = computed tomography; eGFR = estimated glomerular filtration rate; CKD = chronic kidney disease; mRS = modified Rankin Scale. Please click here to view a larger version of this figure.

Figure 3: Serial non-contrast CT images in a patient with CIE. (A) Baseline CT before procedure. (B) CT at symptom onset showing cortical hyper density in the right parietal lobe (arrow). (C) CT on postoperative day one showing partial resolution. (D) CT on postoperative day two showing further resolution. Please click here to view a larger version of this figure.

Figure 4: LASSO-based feature selection. (A) Ten-fold cross-validation curve used to determine the optimal λ based on the minimum MSE. (B) Coefficient profiles of candidate variables plotted against log(λ), showing the changes in variable coefficients with increasing regularization strength. Abbreviations: LASSO = least absolute shrinkage and selection operator; λ = regularization parameter; MSE = mean squared error. Please click here to view a larger version of this figure.

Figure 5: Coefficients of selected variables. Bar plot showing coefficients of variables retained after LASSO selection. Positive coefficients indicate positive association; negative coefficients indicate inverse association. Please click here to view a larger version of this figure.

Figure 6: Training and test accuracy of the five ML models. Accuracy of five ML models (Naive Bayes, SVM, KNN, LightGBM, MLP) in training and test sets. Abbreviations: SVM = support vector machine; KNN = k-nearest neighbor; LightGBM = light gradient boosting machine; MLP = multilayer perceptron. Please click here to view a larger version of this figure.

Figure 7: ROC curves of the Naive Bayes model. Receiver operating characteristic curves for the training and internal test cohorts with corresponding area under the curve values. Abbreviations: ROC = receiver operating characteristic; AUC = area under the curve. Please click here to view a larger version of this figure.

Figure 8: Decision curve analysis. DCA showing net benefit of different models across threshold probabilities compared with treat-all and treat-none strategies. Shaded regions indicate threshold probability ranges where the model achieved positive net benefit compared with both treat-all and treat-none strategies. Abbreviations: DCA = decision curve analysis. Please click here to view a larger version of this figure.

Supplementary Figure 1: Confusion matrix of the Naive Bayes model in the internal test cohort. The matrix displays the numbers of correctly and incorrectly classified samples at the selected classification threshold.Please click here to download this file.

Supplementary Figure 2: Confusion matrix of the SVM model in the internal test cohort. The matrix displays the numbers of correctly and incorrectly classified samples at the selected classification threshold.Please click here to download this file.

Supplementary Figure 3: Confusion matrix of the KNN model in the internal test cohort. The matrix displays the numbers of correctly and incorrectly classified samples at the selected classification threshold.Please click here to download this file.

Supplementary Figure 4: Confusion matrix of the LightGBM model in the internal test cohort. The matrix displays the numbers of correctly and incorrectly classified samples at the selected classification threshold.Please click here to download this file.

Supplementary Figure 5: Confusion matrix of the MLP model in the internal test cohort. The matrix displays the numbers of correctly and incorrectly classified samples at the selected classification threshold.Please click here to download this file.

Characteristic	CIE ( n = 12 )	Non-CIE ( n = 149 )	p value
Demographics
Age (years), median (IQR)	67 (62.5–69.0 )	63 ( 55.0–70.0 )	0.194
Male, n (%)	9 (75.0)	105 (70.5)	1.000
Weight (kg), median (IQR)	65 (55.0–70.0)	70 (60.0–77.5)	0.160
Comorbidities, n (%)
Hypertension	7 (58.3)	102 (68.5)	0.526
Diabetes	4 (33.0)	57 (38.3)	1.000
Coronary heart disease	2 (16.7)	23 (15.4)	1.000
Habits, n (%)
Alcohol Use	7 (58.3)	78 (52.3)	0.770
Smoking Status	7 (58.3)	59 (39.6)	0.233
Lab Values, median (IQR)
Serum Creatinine (μmol/L)	85.2 (65.0–89.0)	65.3 (54.9–75.7)	0.004
eGFR (mL/min/1.73m²)	81.0 (60.6–90.3)	102.7 (89.0–116.3)	<0.001
Total Cholesterol (mmol/L)	3.3 (3.0–4.4)	3.4 (2.9–4.1)	0.775
Triglycerides (mmol/L)	1.8 (1.2–2.1)	1.2 (0.8–1.6)	0.017
Procedural Characteristics
Procedure Type, n ( % )			<0.001
- Stent Angioplasty	7 (58.3)	146(98.0)
- Embolization	5 (41.7)	3 (2.0)
Lesion Location, n (%)			<0.001
- Posterior Circulation	7 (58.3)	9 (6.0)
- Anterior Circulation	5 (41.7)	140 (94.0)
Contrast Media Type, n (%)			0.982
- 2nd Generation	6 (50.0)	75 (50.3)
- 3rd Generation	6(50.0)	74 (49.7)
Total Contrast Volume (mL), median (IQR)	200 (188.0–200.0)	165 (155.0–180.0)	0.005
Procedure Duration (min), median (IQR)	78.5 (52.5–124.5)	50 (47.0–57.0)	0.008
Novel Index
CGR, median (IQR)	2.35 (2.17–2.70)	1.63 (1.40–1.85)	<0.001

Table 1: Baseline clinical characteristics of patients. Demographic, clinical, laboratory, and procedure-related characteristics of patients were compared between the CIE and non-CIE groups. Continuous variables are presented as mean ± standard deviation or median (interquartile range), and categorical variables are presented as number and percentage. Statistical comparisons between groups were performed using appropriate parametric or nonparametric tests. Abbreviations: CIE = contrast-induced encephalopathy; eGFR = estimated glomerular filtration rate; CGR = contrast volume-to-eGFR ratio; IQR = interquartile range; SD = standard deviation.

Model  name	Accuracy	AUC	95% CI	Sensitivity	Specificity	Dataset
Naive Bayes	0.875	0.961	0.924–0.998	1.0	0.864	train
Naive Bayes	0.909	0.952	0.843–1.000	1.0	0.903	test
SVM	0.992	0.999	0.997–1.000	1.0	0.992	train
SVM	0.879	0.935	0.796–1.000	1.0	0.871	test
KNN	0.891	0.969	0.941–0.997	1.0	0.881	train
KNN	0.939	0.718	0.194–1.000	0.5	0.968	test
LightGBM	0.969	0.988	0.973–1.000	1.0	0.966	train
LightGBM	0.848	0.944	0.822–1.000	1.0	0.839	test
MLP	0.914	0.961	0.908–1.000	0.9	0.915	train
MLP	0.970	0.742	0.228–1.000	0.5	1.000	test

Table 2: Comparison of the performance of the models. The predictive performance of different ML models was evaluated using accuracy, area under the curve (AUC), 95% confidence interval (CI), sensitivity, and specificity in the training and internal test datasets. Abbreviations: CI = confidence interval.

Supplementary Table 1: Summary of missing data assessment before model development. Summary of missing values for all candidate predictor variables before model development. No missing values were observed among the included variables; therefore, no imputation procedure was performed.Please click here to download this file.

Supplementary Table 2: Variable dictionary of the analytical dataset. Definitions, coding methods, variable categories, and descriptions of clinical, laboratory, and procedure-related variables included in the ML analysis.Please click here to download this file.

Supplementary Table 3: Hyperparameter search space and selected parameters for ML models. Candidate hyperparameter ranges and optimized parameters obtained during model tuning for each ML algorithm are summarized.Please click here to download this file.

Supplementary Table 4: Comparative ROC analysis of CGR and its individual components. Receiver operating characteristic analysis was performed to compare the predictive performance of CGR and its individual components for CIE prediction.Please click here to download this file.

Supplementary Table 5: Quantitative feature ranking based on LASSO coefficients. Selected variables were ranked according to their coefficient values after LASSO regression, showing the relative contribution and direction of association of each variable in the predictive model.Please click here to download this file.

Supplementary Table 6: Youden thresholds determined in the training cohort and corresponding classification performance metrics. The optimal threshold for each model was determined exclusively in the training cohort using the Youden index and was subsequently fixed and applied to the internal test cohort. Accuracy, sensitivity, and specificity in the internal test cohort were calculated using these fixed thresholds.Please click here to download this file.

Discussion

CIE is a rare but potentially serious complication following neurointerventional procedures, characterized by acute neurological dysfunction that occurs shortly after exposure to contrast agents. Although the condition is typically reversible, in some cases it can lead to persistent neurological deficits or even life-threatening outcomes, underscoring the importance of early risk identification^17,18,19,20. However, the pathophysiological mechanisms underlying CIE are complex and multifactorial, involving interactions among patient-specific factors, contrast agent properties, and procedural characteristics^{9,10,21,22,23,24}. This complexity makes it extremely challenging to achieve accurate perioperative risk assessment based solely on single clinical variables or empirical judgment.

In this study, five ML models were constructed and evaluated to predict the risk of CIE in patients undergoing cerebrovascular intervention. After descriptive performance comparison, the Naive Bayes model was identified as the model with the most balanced descriptive performance in this study. In terms of model discrimination ability, the Naive Bayes model showed the highest numerical AUC in the internal test cohort (AUC = 0.952; 95% CI: 0.843–1.000), compared with LightGBM (AUC = 0.944; 95% CI: 0.822–1.000), SVM (SVM; AUC = 0.935; 95% CI: 0.796–1.000), MLP (MLP; AUC = 0.742; 95% CI: 0.228–1.000), and KNN algorithm (KNN; AUC = 0.718, 95% CI: 0.195–1.000). Because no formal pairwise statistical comparison among model AUCs was performed, these differences should be interpreted as descriptive rather than definitive evidence of model superiority. It is worth noting that KNN and MLP performed well in the training set (AUC = 0.969 and 0.961, respectively), but their performance declined substantially in the test set, suggesting possible overfitting²⁵. In contrast, the Naive Bayes model maintained similar descriptive AUC values between the training and internal test cohorts (AUC difference = 0.009), suggesting relatively stable descriptive performance in this dataset. However, this apparent stability should be interpreted cautiously because the internal test cohort included only 33 patients and a very limited number of CIE events.

From a practical perspective, the relatively simple probabilistic structure of Naive Bayes may be suitable for small-sample clinical datasets, particularly when the number of outcome events is limited. In this study, the Naive Bayes model showed relatively balanced descriptive performance between the training and internal test cohorts. However, because the internal test cohort was small and contained very few CIE events, this apparent stability should be interpreted cautiously and requires external validation. The relatively favorable descriptive performance of the Naive Bayes model in this dataset may be related to its simple probabilistic structure and its suitability for small-sample settings under the feature-independence assumption^14,15. By contrast, more complex models such as MLP and LightGBM are more prone to overfitting under limited-sample conditions, thereby reducing their generalization capabilities. These findings also suggest that, in high-risk, clinically oriented research fields like neurointerventional, model selection should place greater emphasis on stability, reproducibility, and clinical feasibility rather than blindly pursuing the absolute superiority of complex algorithms or single performance metrics. DCA suggested that the Naive Bayes model may provide net benefit across selected threshold probabilities, but this finding remains exploratory and requires external validation. Moreover, its simple structure, efficient computation, and interpretability may support its future evaluation as an exploratory perioperative risk-stratification model, but external validation is required before clinical implementation. From a clinical perspective, the incidence of CIE in this study was 7.5%, higher than the 2%–5% generally documented in previous reports⁷. This discrepancy is largely attributable to the highly selective nature of the study population: the cohort focused on patients with cerebrovascular disease who underwent complex neurointerventional procedures such as aneurysm embolization. The unique contrast between agent administration patterns associated with these procedures, such as super selective angiography and local high-dose retention—as well as longer procedure durations, are all recognized risk factors for CIE^9,10,26. Moreover, the patients’ higher baseline risk further elevates the overall incidence rate. The study findings underscore that lesions of the posterior circulation and embolization procedures are strong predictors of CIE. Anatomically, the posterior circulation (vertebrobasilar system) supplies the brainstem, cerebellum, and occipital lobe. The occipital lobe exhibits extremely high sensitivity to contrast agent toxicity, which explains why CIE patients often present with cortical blindness. Additionally, the hemodynamic characteristics of the vertebrobasilar system differ from those of the anterior circulation, which may partly explain the observed association between posterior circulation lesions and CIE in this cohort.

Embolization procedures are typically more challenging than simple angioplasty, requiring frequent super selective angiography to confirm the position of the coils and the extent of thrombus occlusion, thereby resulting in a significantly higher peak contrast agent exposure to the local vascular bed per unit time compared to routine procedures. This intense and localized contrast exposure may be a biologically plausible contributor to BBB disruption, although this mechanism could not be directly verified in the present retrospective study. Consistent with the baseline characteristic analysis, patients in the CIE group had significantly higher rates of embolization treatment and a greater proportion of posterior circulation lesions compared to those in the non-CIE group, suggesting that surgical complexity and lesion anatomical location may be associated with the occurrence of CIE. Moreover, the longer procedure duration and higher total contrast agent usage in the CIE group further underscore the cumulative impact of surgery-related stressors on the risk of CIE development.

To enhance the clinical interpretability of the model, this study evaluated CGR as an exploratory composite index. The risk of CIE occurrence depends not only on the absolute dose of contrast agent entering the cerebral vasculature but is also closely related to the rate at which contrast agent is cleared from the systemic circulation. CGR integrates total contrast volume and eGFR into a single composite index reflecting contrast agent burden relative to renal-clearance capacity. This design approach draws on the well-established ratio-based paradigm used in predicting contrast-induced nephropathy (CIN), such as the contrast agent volume to eGFR ratio²⁷, and aligns with current strategies for constructing comprehensive biomarkers like the neutrophil-to-lymphocyte ratio (NLR)²⁸ and the triglyceride-glucose index (TyG)²⁹, all of which aim to reveal the underlying pathophysiological states of diseases. Iodinated contrast agents are primarily excreted by the kidneys in their original form. When renal clearance is reduced, systemic exposure to iodinated contrast agents may be prolonged, which could increase endothelial exposure to contrast agents. During neurointerventional procedures, contrast agents are often injected directly into the cerebral circulation via the internal carotid or vertebral arteries at high concentrations; the locally extreme osmotic pressure gradient may disrupt BBB integrity³⁰. At this point, if systemic clearance is impaired, the accumulated contrast agent in the bloodstream may continue to maintain a concentration gradient across the BBB, driving the diffusion of the contrast agent into the brain interstitial space. Moreover, CGR may provide a clinically intuitive way to integrate contrast burden and renal clearance, but its incremental predictive value over contrast volume or eGFR alone requires further validation. Therefore, CGR may serve as an exploratory index for perioperative CIE risk stratification.

This study found that the type of contrast agent (second-generation low-osmolar vs. third-generation iso-osmolar) did not have a significant independent predictive value for CIE (p = 0.982). This finding is not entirely consistent with some previous reports. One possible explanation is that factors other than osmolarity, such as chemical toxicity or viscosity, could contribute to contrast-agent neurotoxicity; however, this interpretation remains speculative and requires further mechanistic validation. This interpretation is supported by previous clinical and mechanistic studies^31,32. Moreover, in clinical practice, operators may preferentially select third-generation iso-osmolar contrast agents for high-risk patients, introducing potential indication-related confounding bias that could mask any underlying differences among various types of contrast agents.

Several critical protocol steps should be emphasized for the successful application of this workflow. First, rigorous outcome adjudication is essential because inaccurate CIE classification would directly affect model training and evaluation. Suspected CIE cases should be assessed using the temporal relationship between contrast exposure and symptom onset, neurological manifestations, postoperative imaging findings, and exclusion of alternative diagnoses. Second, accurate extraction and verification of total contrast volume and eGFR are important because these two variables determine CGR calculation. Third, CGR should be calculated using the same rule for all patients after confirming both total contrast volume and eGFR. Fourth, feature selection, imputation-parameter estimation, hyperparameter tuning, and model training should be performed within the training cohort only to avoid data leakage. Finally, because CIE is uncommon and the number of events is limited, model performance should be interpreted cautiously and validated externally before clinical implementation³³.

Several method modifications and troubleshooting strategies should also be considered when applying this workflow in other settings. If the proportion of missing data is high, investigators should first determine whether missingness is random and should avoid imputing variables with excessive missingness. If the number of CIE events is very small, model complexity should be reduced, and simpler or penalized models may be preferred to reduce overfitting. If the workflow is applied to another center, variable definitions, coding rules, eGFR calculation methods, contrast-volume recording practices, and imaging-review criteria should be harmonized before model training. If model performance declines during internal or external validation, potential causes such as class imbalance, data leakage, inconsistent preprocessing, unstable feature selection, and differences in patient case mix should be examined. For cases with uncertain CIE diagnosis, adjudication by multiple experienced clinicians and careful exclusion of alternative diagnoses are recommended before model development.

This study has the following limitations: First, this study adopts a single-center retrospective design with a limited number of CIE events. Only 12 CIE events occurred among the 161 included patients, and the internal test cohort included only 33 patients with a very limited number of CIE cases. Therefore, performance metrics such as AUC, sensitivity, and specificity may be unstable and highly sensitive to the classification of only one or two cases. In the future, external validation through multicenter, large-sample prospective studies will be needed to further confirm the model’s reliability. Second, the model developed in this study is primarily based on routine clinical variables collected before and during surgery. In the future, we could further integrate intraoperative contrast agent kinetic parameters, early postoperative imaging features, and serum-specific biomarkers to build a more dynamic and comprehensive CIE risk prediction system. Third, the low incidence of CIE (7.5%) resulted in marked class imbalance, with only 12 CIE cases and 149 non-CIE cases. This imbalance may reduce the reliability of model-performance metrics. Accuracy may overestimate model performance when the non-CIE class dominates the dataset, whereas sensitivity, specificity, positive predictive value, and negative predictive value may be unstable because they are influenced by a very small number of CIE events. Therefore, the proposed model should currently be regarded as an exploratory risk-stratification framework rather than a definitive clinical decision-making tool. Additional repeated cross-validation or resampling-based validation was not performed in the present study because the number of CIE events was very limited, and such analyses may still yield unstable estimates under marked class imbalance. Multicenter external validation and prospective assessment are required before clinical implementation.

In summary, this study systematically compared several ML models and found that the Naive Bayes model showed the most balanced descriptive performance for CIE prediction in the current dataset. This model may provide preliminary support for perioperative risk stratification, but its clinical utility requires further validation. Moreover, this study evaluated CGR as an exploratory composite metric, which quantifies the imbalance between “contrast agent load and renal clearance.” CGR emerged as one of the most influential predictors in the model and may provide useful information for individualized perioperative risk assessment. These findings may help generate hypotheses for identifying high-risk patients undergoing neurointerventional procedures, but further studies with larger samples and external validation are needed. Future studies with larger samples and external validation are needed to determine whether multimodal predictive models can improve CIE risk stratification and support more precise prevention and management of this complication.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Electronic medical-record system	Northern Theater General Hospital	Institutional clinical system	Used for retrospective data extraction Collection of demographic and clinical-history variables
Laboratory information system	Northern Theater General Hospital	Institutional laboratory system	Used for retrospective laboratory-data extraction Collection of serum creatinine, eGFR, lipid profile, and other laboratory data
Digital subtraction angiography system	Philips	Allura Xper FD 20 (with Interventional Workstation R1.3.2)	Used during routine clinical procedures Neurointerventional procedure and recording of contrast use
CT Scanner	Philips	Ingenuity CT	Used for postoperative imaging assessment Postoperative cranial CT evaluation
OnekeyAI structured-data analysis platform	OnekeyAI	20240916	Used as a research tool; no role in clinical decision-making Structured-data preprocessing, model construction, model evaluation, and figure generation
Python	Python Software Foundation	3.7.12	Open-source programming language Data preprocessing, statistical analysis, model development, and visualization
scikit-learn	scikit-learn developers	1.0.2	Open-source Python machine-learning library LASSO, Naive Bayes, SVM, KNN, MLP, ROC analysis, and performance metrics