CIN-RiskNet: a dynamic feature-enhanced TabTransformer with hybrid SMOTE-noise augmentation for contrast-induced nephropathy prediction

Abstract

Purpose:

To propose a dynamic feature-enhanced TabTransformer framework that providing a more effective and accurate tool for predicting Contrast-Induced Nephropathy (CIN).

Methods:

This study proposes CIN-RiskNet, a dynamic feature-enhanced TabTransformer model integrated with a hybrid SMOTE-Noise augmentation strategy. The approach includes adaptive feature gating to suppress noise, synthetic minority oversampling to address class imbalance, and multi-head self-attention to capture complex feature interactions. The model was trained and evaluated under a leakage-free stratified five-fold cross-validation protocol, where SMOTE and Gaussian noise were applied only to the training split within each fold trained on a clinical dataset from Tianjin University Chest Hospital that including a total of 1,679 patients who underwent percutaneous coronary intervention for coronary heart disease.

Results:

Under leakage-free five-fold evaluation, CIN-RiskNet achieved strong performance with an accuracy of 95.40%, a recall of 95.40%, and an F1-score of 95.42%. It attained the highest F1-score and recall among all evaluated configurations. It outperformed not only traditional machine learning models including XGBoost, Random Forest, and Support Vector Machine, but also the Mehran risk score, a widely used clinical scoring system for CIN prediction. Ablation studies confirmed the contributions of each module, demonstrating improved robustness and generalization.

Conclusion:

The proposed model effectively addresses key challenges in CIN prediction, including class imbalance and feature noise, through an integrated deep learning framework. It shows promising potential as a decision-support tool, while external multicenter validation remains necessary before broad clinical deployment.

1 Introduction

Contrast-induced nephropathy (CIN) has emerged as a critical complication following invasive cardiovascular procedures, such as percutaneous coronary intervention (PCI), driven by the increasing global use of iodinated contrast media (CM) (1). CIN is defined as a significant increase in serum creatinine levels within 48 to 72 h after exposure to contrast media, representing a reversible yet dangerous form of acute renal failure (2). Studies have shown that patients who develop CIN after PCI have a significantly increased risk of adverse cardiovascular events, regardless of whether their renal function ultimately recovers. Especially among high-risk groups such as those with diabetes, the incidence rate of CIN can reach up to 50% (3). Currently, there is no effective treatment available, making early screening and prevention the only clinically effective strategy (4).

Risk prediction for CIN has long been a focal point of clinical research. Early methods for predicting CIN primarily relied on statistical models. For instance, Mehran identified eight independent risk factors using multiple regression and developed a CIN risk score (5). Simultaneously, with the continuous advancement of machine learning, methods such as random forests and ensemble models have been gradually introduced to enhance prediction accuracy (6–9). However, these methods are highly dependent on manual feature engineering, struggle to capture nonlinear relationships, and are highly sensitive to class imbalance issues.

Current research can effectively assist doctors in predicting the occurrence of CIN, but these methods still face two persistent challenges: Firstly, clinical datasets are filled with a large number of feature attributes and noise, requiring manual feature selection, which introduces a significant amount of subjective bias (10); Secondly, severe class imbalance leads to models favoring majority class predictions, thereby increasing the risk of missed diagnoses.

To address the issues, we propose CIN-RiskNet, a dynamic feature-enhanced TabTransformer framework. This model integrates the SMOTE-Noise enhancement strategy, which balances class distribution while simulating clinical measurement errors, and a dynamic feature gating mechanism that can adaptively suppress redundant information. The experimental results on the dataset from the Department of Cardiology at Tianjin University Chest Hospital demonstrate that CIN-RiskNet significantly outperforms traditional models in multiple evaluation metrics, including accuracy, precision, and recall, providing a more effective and accurate tool for predicting CIN.

2 Materials and methods

2.1 Clinical characterization of dataset

2.1.1 Patients

To evaluate the performance of our proposed method, we utilized Tianjin University Chest Hospital Cardiology Department Dataset. This study consecutively enrolled patients who underwent PCI from 2023.04 to 2025.04. We conducted this study in accordance with the Declaration of Helsinki. Ethics approval was obtained from the Ethics Committee of Tianjin Chest Hospital (2021YS-031-01). All patients signed informed consent.

2.1.2 Inclusion and exclusion criteria

Patients were eligible if they were 18 years of age or older, met the diagnostic criteria for coronary heart disease and subsequently undergo elective PCI. Exclusion criteria included: (1) acute renal failure or estimated Glomerular Filtration Rate (eGFR) < 30 mL/min/1.73 m²; (2) Severe heart failure [left ventricular ejection fraction (LVEF)] < 30% or B-type natriuretic peptide (BNP) > 5,000 ng/L) or cardiogenic shock; (3) Perform emergency PCI; (4) Have used contrast medium or are allergic to contrast medium within the past 7 days; (5) Primary kidney disease, long-term dialysis, or undergoing renal replacement therapy within 24 h after admission; (6) Severe liver and lung failure; (7) Abnormal coagulation function; (8) Infectious diseases; (9) Malignant tumor; (10) Death within 48 h of admission or hospital stay less than 48 h.

2.1.3 Primary endpoint definition

The primary endpoint was occurrence of CIN, defined as a rise in serum creatinine (Scr) ≥ 44.2 μmol/L or ≥ 25% above baseline within 72 h after contrast medium administration (2). Baseline Scr was defined as the first serum creatinine value obtained within 24 h of hospital admission, before major surgery or specific nephrotoxic interventions, representing the initial renal function at admission. Postoperative Scr measurements were standardized across all included patients at 48 h and 72 h after the procedure. Sample collection and testing followed a uniform protocol. Cases with non-standard or missing key postoperative Scr time points were excluded to ensure temporal consistency and data reliability.

2.1.4 Statistical analysis

Continuous variables were analyzed using Student's t-test and non-parametric tests, with outcomes presented as mean ± standard deviation (SD) or median and interquartile range, respectively. Categorical variables were assessed via Chi-square test and Fisher's exact test, and results were expressed as counts and percentages. All statistical analyses were performed using SPSS software (version 25.0; SPSS Inc., New York, USA).

2.1.5 Dataset

Finally, a total of 1,679 patients who underwent PCI for coronary heart disease were finally included, among whom 103 were patients with CIN (Figure 1). The clinical baseline characteristics of the patients are presented in Table 1, a summary of its characteristics is provided in Table 2. We performed temporal feature governance and leakage prevention. To prevent outcome leakage, we predefined PCI/contrast administration as the prediction index time. Only variables available before or at this index time were eligible as model inputs. Post-procedural laboratory values, creatinine-change variables, and any endpoint-derived features were strictly excluded from model training, hyperparameter tuning, and SHAP interpretation. As shown in Table 3, only variables available before or at the index time were eligible as model inputs; hydration volume and all post-procedural creatinine-derived variables were excluded to prevent outcome leakage.

Figure 1

2.2 Proposed method

This study proposes the CIN-RiskNet designed for clinical tabular data. The core innovation lies in the integration of a dynamic feature gating mechanism with a TabTransformer encoder, enabling adaptive interaction modeling of multidimensional clinical features. This design addresses common challenges in medical data analysis, including noise sensitivity, class imbalance, and complex feature dependencies. The overall framework is illustrated in Figure 2.

Figure 2

Prior to feeding data into the model, we employed a SMOTE-Noise hybrid augmentation strategy to mitigate the effects of small sample size, high noise levels, and class imbalance inherent in clinical datasets. Specifically, SMOTE oversampling was first applied to the CIN-positive samples to generate synthetic instances and balance the positive/negative class ratio. To further prevent overfitting, Gaussian noise was randomly injected into both the synthetic and original samples.

The augmented dataset was then passed through the gating unit. Given that clinical data often contain a large number of low-informative or noisy features, traditional static feature selection approaches fail to adapt to patient-specific feature distributions. To address this, we adopted a gating mechanism for dynamic feature selection and representation learning. The gating unit assigns adaptive weights to input features, allowing the model to focus on clinically relevant variables while suppressing noise and redundancy.

The output of the gating unit was subsequently fed into an embedding layer to obtain higher-dimensional representations, which were then passed into the TabTransformer encoder. Each of the 23 clinical variables was treated as an individual feature token. The input was represented as (batch,23,d), and self-attention was performed across feature tokens to model inter-feature dependencies. After processing through the encoder, the output was reduced in dimension to match the original feature space, followed by a multilayer perceptron (MLP)classifier that generated the final prediction. The whole architecture and workflow of the CIN-RiskNet framework is shown in Figure 3. The mathematical formulations of the proposed method are detailed in Equations 1–19.

Figure 3

2.3 SMOTE-noise hybrid augmentation strategy

Given the inherent class imbalance in medical datasets, resampling techniques can be employed to balance class distributions. One such method is undersampling, which creates a new dataset by removing samples from the majority class (11). Conversely, oversampling increases the number of minority class instances to approximate the size of the majority class, thereby generating a more balanced dataset. Since medical datasets typically suffer from both class imbalance and limited overall sample size, oversampling is generally more appropriate.

Synthetic Minority Oversampling Technique (SMOTE) (12) is a widely used and effective oversampling method that generates synthetic minority class samples by interpolating between existing instances and their nearest neighbors. Compared to undersampling, SMOTE often yields better model performance and is considered one of the most prevalent oversampling strategies across domains (13), including widespread use in medical data applications (14, 15).

In this study, SMOTE is applied to generate synthetic samples for the minority class—CIN-positive cases. The algorithm flow is illustrated in Figure 4. For each minority sample, SMOTE identifies its K nearest neighbors in the feature space, with K set to 3 in this study. Euclidean distance is used to measure the similarity between data points. After identifying the neighbors, SMOTE randomly selects one of them and computes the difference between the feature vectors of the minority sample and the selected neighbor. This difference is then multiplied by a random number between 0 and 1 and added to the feature vector of the minority sample, thereby generating a new synthetic sample. These synthetic samples lie along the line segments connecting the minority sample to its neighbors. This process is repeated until the desired level of class balance is achieved.

Figure 4

The number of synthetic samples generated via SMOTE is controlled by the imbalance ratio between the majority and minority classes. The oversampling ratio is defined as follows:where and represent the sample sizes of the majority and minority classes, respectively. Based on the calculated oversampling ratio N, SMOTE determines the number of synthetic samples to generate .

These samples are created by linear interpolation between a minority instance and one of its nearest neighbors , as shown below:Here, is a minority-class sample, is one of its nearest neighbors, is a random interpolation coefficient drawn from a uniform distribution , and is the generated synthetic sample.

However, when the class imbalance is extreme, the large volume of synthetic samples may lead to high redundancy and overfitting. Additionally, such interpolated samples can be overly idealized and may not accurately reflect real-world variability. To address these limitations and simulate the measurement errors and data noise commonly observed in clinical scenarios, this study introduces a Gaussian noise-based hybrid augmentation strategy following SMOTE. The formula is as follows:Here, X denotes the original or synthetic feature vector, is multivariate Gaussian noise with zero mean and identity covariance matrix, and is a noise scaling factor that controls the degree of perturbation. The augmented data are subsequently fed into the feature gating unit for downstream processing.

2.4 Feature gating mechanism

The concept of gated units originated with the Long Short-Term Memory (LSTM) network (16), which uses a forget gate, an input gate, and an output gate to control information flow, allowing the network to selectively retain long-term information or discard irrelevant data, thereby addressing the vanishing and exploding gradient problems of traditional recurrent neural network. Building on LSTM, Chung et al. introduced the Gated Recurrent Unit (GRU) (17) in 2014, which simplifies the gating structure by merging LSTM's three gates into two (reset gate and update gate) and eliminating the cell state in favor of relying solely on the hidden state for information propagation. Gating mechanisms have been widely applied in domains such as speech recognition (18) and image processing (19), and have also been reported in medical applications (20). It serves as a dynamic feature selection technique, designed to apply learnable weight matrices to adaptively emphasize informative features and suppress noisy or redundant ones. Clinical tabular data typically contain numerous low-informative features and significant noise. Traditional static feature selection approaches, which rely on manual screening or fixed weighting schemes, often fail to adapt to the heterogeneous feature distributions across different patients. To address this issue, we incorporate a dynamic feature gating module based on a trainable gating weight matrix. After the original input undergoes SMOTE-Noise augmentation, the enhanced data are passed into the gating unit, computed as follows:Here, denotes the sigmoid activation function， X represents the input feature vector, and are trainable parameters, and indicates element-wise multiplication. This mechanism generates a feature importance score for each input dimension via linear transformation followed by a non-linear gating operation. The resulting weighted feature representation is obtained by element-wise multiplication of the original input with the learned gate vector. All gating parameters are optimized through backpropagation, enabling the model to learn a task-specific feature selection strategy in an end-to-end manner.

Through this mechanism, the model automatically enhances features that are critical for CIN prediction while suppressing irrelevant or noisy dimensions. The dimensionality of the output from the gating unit remains the same as the input and is subsequently fed into the embedding layer.

In this study, each sample in the Tianjin University Chest Hospital Cardiology Department Dataset contains 23 clinical features. The input batch size is set to 32, and after processing through a DataLoader, each input batch is represented as a tensor of shape .The embedding layer transforms this input from a 23-dimensional space to a 64-dimensional space, using the following operation:Here . and are the trainable weight matrix and bias vector, respectively. The embedded feature representation is then forwarded to the TabTransformer encoder for further modeling.

2.5 Training and testing strategy

The Transformer network, initially proposed by Vaswani et al. for machine translation, introduced the groundbreaking self-attention mechanism, which dynamically calculates the pairwise dependencies between elements in a sequence, granting the model strong global contextual awareness. While the Transformer has achieved state-of-the-art results in fields such as image recognition and natural language processing, it has shown limitations in handling tabular data when compared to conventional machine learning models (21, 22). TabTransformer addresses these limitations through targeted modifications for structured tabular data. Unlike the standard Transformer, TabTransformer introduces the following key adaptations: (1) Removal of positional encoding: Tabular features are typically unordered and do not follow a natural sequential structure. (2) Feature-wise embedding: Each column (feature) is individually mapped to a dense vector to capture semantic relationships between features. (3) Context-aware attention: Self-attention models nonlinear inter-feature dependencies, replacing the need for manually designed feature crosses.

In this study, we adopt only the encoder module of the TabTransformer for the CIN classification task, as it effectively captures global inter-feature interactions without requiring a decoder. The encoder consists of multi-head self-attention (MSA) and a feed-forward network (FFN). MSA splits the input feature space into multiple subspaces (attention heads), with each head learning a distinct interaction pattern. Self-attention operates by computing similarity scores between query (Q) and key (K) pairs and applying these weights to the value (V) representations. The computation is defined as:Here, are projections of the input, and are the learned projection matrices for the attention head. projects the concatenated output back to the desired dimension. The number of heads is set to 8, and the feature dimension with a scaling factor to stabilize gradients.

Although tabular data lack a natural sequence structure, TabTransformer requires a sequence dimension in its input. In the revised TabTransformer design, each of the 23 clinical variables is treated as an individual feature token rather than collapsing all variables into a single token, and then reshaping the input tensor as .

The encoder is composed of identical layers, each containing an MSA sub-layer and an FFN sub-layer. Every sub-layer is followed by residual connections and layer normalization, as defined by:Each FFN module comprises two linear layers with a ReLU activation between them. The hidden dimension is expanded to four times the original size and then projected back to the original dimension:Where .

The output of the final encoder layer is flattened (removing the dummy sequence dimension) and passed into a single-layer fully connected classification head, transforming the 64-dimensional embedding into a 2-dimensional class probability vector:Where , and . The unsqueeze and squeeze operations are purely for tensor reshaping, and do not alter the content of the data—only its shape, to match the model's interface requirements.

2.6 Model training and evaluation protocol

This study employs five-fold cross-validation, where the entire patient dataset is randomly divided into five mutually exclusive subsets. In each iteration, one subset is selected as an independent test set, while the remaining four are combined to form the training set. Importantly, in each fold, SMOTE and Gaussian noise were applied strictly to the training subset only, while validation and test subsets were kept untouched to prevent synthetic-sample leakage. This process is repeated five times to ensure that every data instance is used for testing. In each experiment, the training and validation sets are strictly non-overlapping, ensuring that the validation results reflect the model's generalization performance on unseen data. Furthermore, validation based on five distinct data splits mitigates potential biases arising from chance-based data partitioning, thereby enhancing the reliability of experimental conclusions. Within the training set, 20% of the samples are further separated as a validation set, and the remaining 80% are used for model training to prevent information leakage.

The model is trained for 75 epochs with a batch size of 32, using the Adam optimizer (initial learning rate of 0.001 and weight decay of 1 × 10⁻⁵), minimizing the cross-entropy loss. Our model is implemented in Python 3.9 and Torch 2.5, and is configured to run on a GeForce RTX 2080 Ti GPU.

2.7 Comparison with clinical scoring

Mehran score is a widely used clinical risk stratification tool for CIN prediction. However, several variables required by the original Mehran score (e.g., hypotension, intra-aortic balloon pump use, congestive heart failure) were not routinely documented in our structured electronic health records, a common limitation in retrospective studies. We therefore constructed a modified Mehran-like comparator using available variables (age, diabetes, contrast volume, baseline Scr/eGFR) to provide a pragmatic local benchmark.

We emphasize that this comparison is intended solely as a supplementary clinical reference, not as a primary evaluation metric. Our principal comparative assessment relies on multiple machine learning baselines (LOG, RF, XGBoost, SVM, MLP) trained under identical data partitions and preprocessing pipelines (Table 4), where CIN-RiskNet demonstrated consistent superiority. The modified Mehran comparator is included to contextualize model performance within our institutional data environment, acknowledging that structured datasets often lack variables required by established risk scores.

3 Results

We compared the proposed model against mainstream classifiers, including logistic regression (LOG) (23), random forest (RF) (24), support vector machine (SVM) (25), XGBoost (26) and MLP.

To ensure fairness in model comparisons, all methods adopted identical training/testing splits, uniform data preprocessing pipelines, and 5-fold cross-validation. Comparative performance across four key metrics (accuracy, precision, recall, and F1-score) is detailed in Table 4 and visually highlighted in Figure 5(a).

Figure 5

Among all compared models, CIN-RiskNet achieved the highest recall (0.9540) and F1-score (0.9542), substantially outperforming traditional machine learning baselines in detecting minority-class samples (Table 4). Although RF and XGBoost attained marginally higher accuracy (0.9621 and 0.9588) and exhibited strong precision (1.0000 and 0.9156), their recall was severely compromised (0.3992 for both), indicating that these ensemble methods would miss over 60% of true CIN cases—a clinically unacceptable trade-off in screening contexts where false negatives carry severe consequences. By contrast, CIN-RiskNet maintained high precision (0.9545) without sacrificing sensitivity, reflecting robust adaptive feature interaction modeling rather than reliance on manually engineered feature interactions. SVM exhibited extreme precision–recall imbalance (0.7889 and 0.2533), yielding a low F1-score (0.3835) that confirms its unsuitability for imbalanced clinical tasks prioritizing reduced missed diagnoses. LOG and MLP delivered poor overall performance (F1-scores of 0.2213 and 0.2082), constrained by linear separability assumptions or insufficient and unstable capacity to model global feature dependencies. Collectively, these findings validate the superior sensitivity and robust classification capability of CIN-RiskNet across the entire dataset.

The proposed model's F1-score of 0.9542 highlights its exceptional balance between precision and recall—critical for clinical applications requiring simultaneous minimization of false positives (avoiding unnecessary interventions) and false negatives (preventing missed diagnoses). Further validating the effectiveness of the dynamic gating mechanism and SMOTE-Noise hybrid strategy in addressing class imbalance and noise sensitivity. These results solidify the model's potential as a robust tool for high-risk CIN prediction.

In summary, the proposed model achieved state-of-the-art performance, surpassing traditional machine learning approaches across all key metrics—accuracy, precision, recall, and F1-score—with its efficacy further demonstrated by stable training convergence and low classification error rates. These performance estimates should be interpreted in the context of internal validation within a single-center retrospective cohort. Although leakage control was strictly enforced, transportability across different institutions, assay platforms, and case-mix distributions remains to be verified.

It can be seen that the prediction ability of the model is indeed strong, while we cannot ignore the preconditions of imbalanced data samples. This is an unavoidable limitation of this study, but the observed gains from imbalance sample distribution should be interpreted in the clinical context of CIN prevalence. In real-world PCI cohorts, CIN events are relatively infrequent compared with non-CIN cases, resulting in an inherently imbalanced outcome distribution. Therefore, improvements from SMOTE-based rebalancing are expected and clinically meaningful, particularly for reducing missed high-risk cases while maintaining overall discrimination and calibration under a leakage-free evaluation protocol.

Under the five-fold cross-validation setting, we compared three augmentation strategies using identical architecture and training parameters: (1) no augmentation, (2) SMOTE only, (3) Gaussian noise only and (4) SMOTE combined with Gaussian noise. The baseline model without augmentation achieved an accuracy of 0.9370 and an F1 score of 0.9056. Applying SMOTE alone substantially improved the model's performance, increasing the accuracy to 0.9524 and the F1 score to 0.9506. Notably, the combination of SMOTE and Gaussian noise yielded the highest predictive performance, reaching an accuracy of 0.9540 and an F1 score of 0.9542. Under leakage-free five-fold cross-validation, SMOTE combined with Gaussian noise achieved the highest F1-score (0.9542) and recall (0.9540) among all augmentation configurations. SMOTE alone produced a slightly lower F1-score (0.9506), confirming that the hybrid strategy optimally balances precision and recall for CIN prediction. These results establish SMOTE combined with Gaussian noise as the optimal augmentation strategy, achieving superior performance across all key metrics. The comparative performance of the four augmentation settings is illustrated in Figure 6.

Figure 6

Importantly, these improvements were obtained under strict leakage-free stratified cross-validation, where augmentation was applied only to the training subset in each fold.

To provide a clinically interpretable benchmark, we compared CIN-RiskNet with a modified Mehran comparator under the same stratified five-fold test splits. CIN-RiskNet achieved a markedly higher AUROC (0.7594 vs. 0.5094). The near-random discrimination of the modified Mehran comparator reflects the substantial missingness of key clinical variables in our structured dataset—notably hypotension, intra-aortic balloon pump use, and congestive heart failure—rather than intrinsic inadequacy of the original Mehran score. This underscores a practical challenge in retrospective studies: structured electronic health records often lack variables required by established risk scores, limiting the direct applicability of traditional scoring systems.

Although the Brier score of CIN-RiskNet was modestly higher (0.1178 vs. 0.0639), this is expected given the trade-off between strong discrimination and marginal probability calibration: a comparator with near-random discrimination tends to predict close to the baseline event rate for most patients, yielding a deceptively low Brier score that reflects clinical non-informativeness rather than true calibration superiority. Importantly, CIN-RiskNet's superiority is independently confirmed by its consistent outperformance of all machine learning baselines (Table 4), which were evaluated under identical data and validation conditions without reliance on clinical scoring variables. Collectively, these findings support the incremental predictive utility of CIN-RiskNet for clinical risk stratification, particularly in settings where traditional risk scores cannot be fully reconstructed from available data.

To further validate the performance and feasibility of individual modules, we conducted ablation studies by systematically removing or replacing components of the proposed method. Additionally, we integrated multiple alternative modules with the baseline TabTransformer for comparative analysis. The comprehensive results of these ablation and module substitution experiments are documented in Table 5 and visually presented in Figure 5(b).

As shown in Table 5, when the original Transformer was directly applied to medical tabular data with a single random split (S1), it achieved an accuracy of 0.9355, a precision of 0.8731, and an F1-score of 0.9032. Its lack of optimization for feature independence and the single-split evaluation made it sensitive to data distribution shifts, particularly in the minority class. Through 5-fold cross-validation (S2), the model's accuracy improved to 0.9370 (+0.15%), while the F1-score increased from 0.9032 to 0.9056, demonstrating effective mitigation of data distribution bias and more stable performance estimation.

The dynamic gating mechanism (S3) further elevated the F1-score from 0.9056 to 0.9095 (+0.39%), with precision rising from 0.8762 to 0.8805. This confirms that the gating mechanism is a core contributor to model robustness, enhancing the model's ability to selectively emphasize clinically relevant features. Replacing mean pooling with a CLS token (S4) yielded marginal gains (F1-score 0.9095 to 0.9126), suggesting that the dynamic gating mechanism already captures sufficient global context.

Notably, integrating SMOTE (S5) produced the most substantial improvement: precision surged from 0.8845 to 0.9489 (+6.44%), and the F1-score improved from 0.9126 to 0.9506 (+4.16%). This indicates that class imbalance was the primary bottleneck in CIN prediction, and synthetic minority oversampling effectively compensated for the scarcity of positive cases. In contrast, noise augmentation alone (S6) improved precision to 0.9139 and F1 to 0.9281, but its effect was markedly weaker than SMOTE, suggesting that insufficient minority-class samples, rather than simple overfitting, was the dominant issue.

Combining SMOTE with Gaussian noise augmentation (S7) yielded consistent performance improvements across all key metrics. Accuracy increased from 0.9524 to 0.9540, precision from 0.9489 to 0.9545, recall from 0.9524 to 0.9540, and the corrected F1-score from 0.9506 to 0.9542. These results demonstrate that the hybrid SMOTE-Noise strategy achieves superior precision-recall balance compared to SMOTE alone, with the Gaussian noise component enhancing robustness to clinical measurement variability without compromising discriminative power. The complementary mechanisms, SMOTE addressing class imbalance through synthetic minority expansion and Gaussian noise simulating real-world measurement errors, jointly optimize model performance for CIN screening applications. Consequently, the configuration leveraging cross-validation, dynamic gating, and SMOTE combined with Gaussian noise emerges as the optimal framework for CIN prediction.

To improve interpretability, we conducted SHAP analysis on the revised CIN-RiskNet and reported both a SHAP summary bar plot and a SHAP beeswarm plot. As shown in Figure 7A (bar plot) and Figure 7B (beeswarm plot), the model's predictions were predominantly driven by renal-function-related variables and clinically relevant treatment and risk factors. All SHAP analyses were performed using the final leakage-controlled feature set containing only pre-procedural/intra-procedural variables.

Figure 7

Specifically, the top features demonstrated clear physiological relevance. Elevated preoperative serum creatinine and reduced eGFR strongly increased the predicted risk, as they directly reflect diminished baseline glomerular filtration reserve—the most established predisposing factor for CIN. Larger contrast agent doses also markedly amplified risk, consistent with the dose-dependent tubular toxicity and medullary hypoxia induced by iodinated contrast media; higher volumes prolong renal exposure and overwhelm the kidneys’ compensatory mechanisms. Height emerged as a notable predictor, likely serving as a proxy for body surface area and thereby modulating the effective contrast concentration relative to renal cortical mass.

Among the remaining influential variables, a history of myocardial infarction and the use of CCBs and ACEIs showed appreciable positive contributions to predicted risk. These associations likely reflect underlying hemodynamic instability and compromised renal microcirculation in patients with advanced cardiovascular disease, rather than direct nephrotoxicity of the medications themselves. Notably, diuretic use, reduced LVEF, and advanced age exhibited relatively modest SHAP values, suggesting limited independent predictive leverage beyond renal function and contrast exposure within this strictly preoperative prediction framework. Conversely, metabolic indicators such as lipid profiles, BMI, and preoperative BUN demonstrated minimal influence on model output.

These findings indicate that the model focuses on biologically plausible predictors and is consistent with known CIN risk mechanisms. The most influential predictors were mainly renal-function-related and clinically meaningful variables, suggesting that the model’s decision process is aligned with known CIN risk mechanisms.

4 Discussion

As a major contributor to hospital-acquired acute kidney injury, CIN poses a serious threat to patient safety, carrying substantial risks of complications and mortality for those already at high risk (27). The underlying mechanisms involve direct tubular toxicity and medullary ischemia induced by iodinated contrast media, particularly when renal functional reserve is already compromised (28, 29). Although the Mehran score provides a contemporary risk stratification tool, external validation studies report only moderate discrimination and its linear additive structure cannot capture complex nonlinear interactions among renal function, contrast volume, and hemodynamics. Moreover, as a static rule-based model, the Mehran score cannot adapt to the noise, missing data, or high dimensionality inherent in electronic health records, nor does it address the extreme class imbalance of CIN, which often results in poor sensitivity in low-prevalence settings (30, 31). Recent machine learning and deep learning approaches have shown improved predictive accuracy, however, many still depend on manual feature engineering, remain vulnerable to class imbalance, and lack robustness to heterogeneous real-world data. Crucially, modern deep learning architectures designed for tabular data, particularly those leveraging self-attention to model inter-feature dependencies without handcrafted interactions, have rarely been applied to CIN prediction. Motivated by these limitations, CIN-RiskNet is designed to learn discriminative feature representations directly from structured clinical data, with SMOTE oversampling and Gaussian noise injection employed to mitigate class imbalance and simulate clinical measurement variability, respectively, while the dynamic gating mechanism suppresses noise at the feature level.

Compared with conventional risk scoring and standard machine-learning baselines, CIN-RiskNet provides a data-driven framework for moving CIN management from reactive treatment toward earlier prevention-oriented decision support.

Going forward, we plan to collaborate with multiple healthcare institutions to build a geographically diverse dataset, allowing us to evaluate model stability across heterogeneous populations. We also aim to integrate multimodal data, combining imaging features and laboratory results to develop a comprehensive multimodal prediction system. Through the synergistic design of gating mechanisms, data augmentation, and the TabTransformer architecture, this study presents a highly accurate and interpretable CIN prediction model. While challenges remain in terms of computational complexity and cross-center generalizability, the modular innovation of our approach offers new insights for the development of medical AI models. Future efforts will focus on clinical translation and technical refinement to support the deployment of this model in real-world healthcare settings.

Several limitations should be acknowledged. First, this was a single-center retrospective study, which limits external generalizability despite the relatively large cohort (n = 1,679). As Transformers are susceptible to overfitting center-specific data distributions, we employed strict leakage-free stratified cross-validation, dynamic feature gating, and SMOTE restricted to training folds to mitigate this risk. Nevertheless, these internal safeguards cannot substitute for prospective multicenter validation, which remains essential before clinical deployment. Second, although we enforced strict temporal feature governance to avoid outcome leakage, independent replication in external datasets is still necessary to confirm transportability. Third, comparator analysis used a modified Mehran framework because several variables required for the original score were unavailable in our structured dataset; therefore, this comparison should be interpreted as a pragmatic local benchmark. The near-random AUROC (0.5094) reflects data incompleteness—a common real-world limitation that underscores the advantage of data-driven models capable of leveraging all available variables without requiring complete risk score reconstruction. Fourth, this study used structured tabular variables only; future work should incorporate multimodal data and prospective multicenter cohorts. In addition, future studies should include prospective impact analyses to determine whether model-guided prevention strategies can improve patient-centered outcomes beyond retrospective discrimination metrics.

5 Conclusion

This study proposes CIN-RiskNet, a dynamic feature-enhanced TabTransformer model with hybrid SMOTE-Noise augmentation for predicting CIN. The model significantly improves prediction performance by integrating dynamic feature selection, noise-robust data augmentation, and multi-head attention mechanisms. By unifying these components in an end-to-end architecture, CIN-RiskNet not only streamlines the workflow but also ensures that feature interactions are captured adaptively and robustly, even in the presence of noisy or imbalanced data. Experimental results show that CIN-RiskNet achieves an accuracy of 95.40%, recall of 95.40%, and F1-score of 95.42% under five-fold cross-validation, substantially outperforming both traditional machine learning models and conventional clinical scoring systems. Importantly, ablation studies confirm that each module—dynamic gating, SMOTE-Noise augmentation, and multi-head self-attention—contributes positively to overall performance gains, with the SMOTE-Noise combination achieving the optimal precision-recall balance underscoring the necessity of their combined application. Despite these encouraging internal results, routine clinical deployment is premature; prospective external multicenter validation and calibration assessment are required before broad implementation.

References

• 1.

  KhanIIftekharMFYousafzaiZAAminQKShafiqI. Clinical and angiographic predictors of acute kidney injury in ST-elevation myocardial infarction treated with primary percutaneous coronary intervention. Cureus. (2025) 17(11):e96052. 10.7759/cureus.96052

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2.

  ZhangPFuHLiuJLiuXYangSGuoZet al. Preoperative fasting blood glucose levels and the risk of contrast-induced nephropathy in patients with diabetes and pre-diabetes undergoing coronary arteriography or percutaneous coronary intervention: a cross-sectional study. Angiology. (2022) 73(7):660–7. 10.1177/00033197211061916

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3.

  ArbelYFusterVBaberUHamzaTHSiamiFSFarkouhME. Incidence, determinants and impact of acute kidney injury in patients with diabetes mellitus and multivessel disease undergoing coronary revascularization: results from the FREEDOM trial. Int J Cardiol. (2019) 293:197–202. 10.1016/j.ijcard.2019.05.064

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4.

  McCulloughPAWolynRRocherLLLevinRNO'NeillWW. Acute renal failure after coronary intervention: incidence, risk factors, and relationship to mortality. Am J Med. (1997) 103(5):368–75. 10.1016/s0002-9343(97)00150-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5.

  MehranRAymongEDNikolskyELasicZIakovouIFahyMet al. A simple risk score for prediction of contrast-induced nephropathy after percutaneous coronary intervention: development and initial validation. J Am Coll Cardiol. (2004) 44(7):1393–9. 10.1016/j.jacc.2004.06.068

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6.

  ChoiHChoiBHanSLeeMShinG-TKimHet al. Applicable machine learning model for predicting contrast-induced nephropathy based on pre-catheterization variables. Intern Med (2024) 63(6):773–80. 10.2169/internalmedicine.1459-22

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7.

  CoxMPanagidesJCDi CapuaJDuaAKalvaSKalpathy-CramerJet al. An interpretable machine learning model for the prevention of contrast-induced nephropathy in patients undergoing lower extremity endovascular interventions for peripheral arterial disease. Clin Imaging. (2023) 101:1–7. 10.1016/j.clinimag.2023.05.011

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8.

  YunDChoSKimYCKimDKOhK-HJooKWet al. Use of deep learning to predict acute kidney injury after intravenous contrast media administration: prediction model development study. JMIR Med Inform. (2021) 9(10):e27177. 10.2196/27177

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9.

  YanPDuanSBLuoXQZhangNYDengYH. Development and validation of a deep neural network-based model to predict acute kidney injury following intravenous administration of iodinated contrast media in hospitalized patients with chronic kidney disease: a multicohort analysis. Nephrol Dial Transplant. (2023) 38(2):352–61. 10.1093/ndt/gfac049

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10.

  WangFCasalinoLPKhullarD. Deep learning in medicine-promise, progress, and challenges. JAMA Intern Med. (2019) 179(3):293–4. 10.1001/jamainternmed.2018.7117

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11.

  PengMZhangQXingXGuiTHuangXJiangY-Get al. Trainable undersampling for class-imbalance learning. In: AAAI Conference on Artificial Intelligence (Palo Alto, CA, USA: AAAI Press) (2019). 10.1609/AAAI.V33I01.33014707

  • CrossRef
  • Google Scholar
  • View reference in article

• 12.

  ChawlaNVBowyerKWHallLOKegelmeyerWPJJ. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res. (2002) 16:321–57. 10.1613/jair.953

  • CrossRef
  • Google Scholar
  • View reference in article

• 13.

  ElreedyDAtiyaAF. “A novel distribution analysis for smote oversampling method in handling class imbalance”. In: Rodrigues JMF, Cardoso PJS, Monteiro J, Lam R, Krzhizhanovskaya VV, Lees MH, Dongarra JJ, Sloot PMA, editors. International Conference on Conceptual Structures. Cham, Switzerland: Springer (2019). 10.1007/978-3-030-22744-9_18

  • CrossRef
  • Google Scholar
  • View reference in article

• 14.

  XuZShenDNieTKouYYinNHanXJIS. A cluster-based oversampling algorithm combining SMOTE and k-means for imbalanced medical data. Inf Sci. (2021) 572:574–89. 10.1016/j.ins.2021.02.056

  • CrossRef
  • Google Scholar
  • View reference in article

• 15.

  KatohSPeltonenMZeniyaMSakamotoYUtsunomiyaKNishimuraRet al. Improving classification performance for a novel imbalanced medical dataset using SMOTE method. J Clin Med. (2020) 9(3):3161–72. 10.3390/jcm9103161

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16.

  HochreiterSSchmidhuberJ. Long short-term memory. Neural Comput. (1997) 9(8):1735–80. 10.1162/neco.1997.9.8.1735

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17.

  ChungJGulcehreCChoKBengioYJ. Empirical evaluation of gated recurrent neural networks on sequence modeling (2014).

  • Google Scholar
  • View reference in article

• 18.

  OruhJViririSAdegunAJIA. Long short-term memory recurrent neural network for automatic speech recognition. IEEE Access. (2022) 10:30069–79. 10.1109/ACCESS.2022.3159339

  • CrossRef
  • Google Scholar
  • View reference in article

• 19.

  JinHYunKKimE. Gating mechanism in deep neural networks for resource-efficient continual learning. IEEE Access. (2022) 10:18776–86. 10.1109/ACCESS.2022.3147237

  • CrossRef
  • Google Scholar
  • View reference in article

• 20.

  KesnerALAbourbehGMishaniEChisinRTshoriSFreedmanN. Gating, enhanced gating, and beyond: information utilization strategies for motion management, applied to preclinical PET. EJNMMI Res. (2013) 3(1):29. 10.1186/2191-219X-3-29

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21.

  GrinsztajnLOyallonEVaroquauxGJA. Why do tree based models still outperform deep learning on typical tabular data. arXiv.org. (2022) 35:507–20. 10.48550/arXiv.2207.08815

  • CrossRef
  • Google Scholar
  • View reference in article

• 22.

  GorishniyYRubachevIKhrulkovVBabenkoAJA. Revisiting deep learning models for tabular data. Neural Information Processing Systems. (2021) 34:18932–43. 10.48550/arXiv.2106.11959

  • CrossRef
  • Google Scholar
  • View reference in article

• 23.

  CoxDR. The regression analysis of binary sequences. J R Stat Soc Ser B Stat Methodol. (1958) 20(2):215–32. 10.1111/j.2517-6161.1958.tb00292.x

  • CrossRef
  • Google Scholar
  • View reference in article

• 24.

  BreimanLJM. Random forests. Mach Learn. (2001) 45:5–32. 10.1023/A:1010933404324

  • CrossRef
  • Google Scholar
  • View reference in article

• 25.

  CortesCVapnikVJM. Support-vector networks. Mach Learn. (1995) 20:273–97. 10.1007/BF00994018

  • CrossRef
  • Google Scholar
  • View reference in article

• 26.

  ChenTGuestrinC. “Xgboost: a scalable tree boosting system”. In: Krishnapuram B, Shah M, Smola AJ, Aggarwal CC, Shen D, Rastogi R, editors. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. NewYork, NY, USA: ACM (2016). p. 785–94.

  • Google Scholar
  • View reference in article

• 27.

  CarulloNTripodiLMichaelAFagaTBolignanoDCoppolinoGet al. Update on contrast-induced nephropathy: recent developments in its prevention, early diagnosis, and therapy. Medicina (Kaunas). (2026) 62(5):948. 10.3390/medicina62050948

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28.

  ÇalışkanFArıcıMAKocaPMıcılıSCKızıldağSBoraRUet al. Probenecid protects against contrast-induced nephropathy through modulation of small rho GTPases and Pannexin1 channels in Wistar rats. Clin Exp Nephrol. (2025) 29(12):1734–43. 10.1007/s10157-025-02742-4

  • CrossRef
  • Google Scholar
  • View reference in article

• 29.

  TianFLiuKTangZZhouGZhouGChenRet al. Glycyrrhizin alleviates contrast-induced acute kidney injury via inhibiting HMGB1-mediated renal tubular epithelial cells ferroptosis. Ren Fail. (2025) 47(1):2548613. 10.1080/0886022x.2025.2548613

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 30.

  ZhangJChenHYuLHanW. Machine learning prediction of contrast-induced AKI after PCI using the systemic immune-inflammation Index: insights from MIMIC-IV. Ren Fail. (2025) 47(1):2610913. 10.1080/0886022x.2025.2610913

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31.

  TangYWuTWangXWuXChenAChenGet al. Deep learning for the prediction of acute kidney injury after coronary angiography and intervention in patients with chronic kidney disease: a model development and validation study. Ren Fail. (2025) 47(1):2474206. 10.1080/0886022x.2025.2474206

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article