Prognostication of Acute Kidney Injury Following Percutaneous Coronary Intervention

Abstract

Background

Acute kidney injury (AKI) is a serious complication of percutaneous coronary intervention (PCI) associated with increased mortality and health care costs. Traditional risk scores often rely on intraprocedural variables, limiting their utility for preprocedural prophylaxis.

Objectives

We aimed to develop and validate a machine learning model to predict post-PCI AKI using strictly preprocedural electronic health record data.

Methods

This retrospective cohort study analyzed routine electronic health record data from a tertiary medical center (2004-2022). The primary outcome was AKI, defined according to Kidney Disease: Improving Global Outcomes criteria (absolute serum creatinine increase ≥0.3 mg/dL or relative increase ≥50% within 48 hours). A gradient-boosted decision tree ensemble (CatBoost) was trained on preprocedural demographic, clinical, and laboratory variables. Performance was evaluated on a held-out test set (20%) using the area under the receiver operating characteristic curve and precision-recall curve.

Results

The final cohort included 23,728 PCI procedures from 17,943 patients, with an AKI prevalence of 7.0%. On the held-out test set, the model achieved an area under the receiver operating characteristic curve of 0.85 (95% CI: 0.82-0.87) and a precision-recall curve of 0.38 (95% CI: 0.32-0.43). Calibration was excellent (integrated calibration index = 0.017). At the screening threshold (probability 0.041), sensitivity was 0.83 (95% CI: 0.80-0.87). At the rule-in threshold (probability 0.199), specificity was 0.92 (95% CI: 0.91-0.93). Key predictors included baseline creatinine, hemoglobin, uric acid, and white blood cell count.

Conclusions

In this single-center study, we developed a machine learning model using preprocedural variables that predicts post-PCI AKI. Although external validation is required, this model could support individualized risk stratification and preventive strategies.

Central Illustration

Percutaneous coronary intervention (PCI) remains the cornerstone of revascularization for coronary artery disease. However, it carries the inherent risk of periprocedural complications that can significantly impact patient prognosis.¹ Among these, contrast-associated acute kidney injury (AKI) is one of the most common and consequential, occurring in 7% to 15% of patients depending on the definition used and the population complexity.^2,3 The development of AKI after PCI is associated with prolonged hospitalization, need for dialysis, and a markedly increased risk of short- and long-term mortality.^4,5

Identifying patients at high risk for AKI before the procedure is critical for implementing preventive strategies, such as targeted hydration protocols, minimization of contrast volume, and hemodynamic optimization.^6,7 Although traditional risk assessment tools like the Mehran risk score and the National Cardiovascular Data Registry CathPCI model have provided a framework for risk stratification, they face significant limitations in the era of precision medicine.^8,9 Many of these scores rely on intraprocedural variables—most notably contrast volume—which are unknown at the time of preprocedural planning. Furthermore, traditional logistic regression models often assume linear relationships between risk factors and outcomes, potentially failing to capture the complex, nonlinear interactions characteristic of biological systems.¹⁰

The advent of artificial intelligence and machine learning (ML) offers a paradigm shift in clinical prognostication. Algorithms such as gradient-boosted decision trees can ingest high-dimensional electronic health record data to model complex relationships without explicit prespecification.¹¹ Recent studies have suggested that ML models may outperform traditional scores in predicting cardiovascular outcomes.^12,13 However, the clinical adoption of these models requires not just high discrimination, but also precise calibration (accurate probability estimation) and interpretability.¹⁴

In this study, we developed and validated a ML model to predict post-PCI AKI using a large, real-world cohort. We employed a gradient-boosted decision tree ensemble (CatBoost), utilizing strictly preprocedural variables to ensure the model's utility for decision-making in patients undergoing PCI.

Methods

Study population and data source

This retrospective study used routinely collected electronic health record data from Rabin Medical Center (Petah Tikva, Israel). The study population included patients undergoing PCI between 2004 and 2022. Each PCI procedure was treated as a separate observation.

Outcome definition

The primary outcome was postprocedural AKI. AKI was defined according to the 2012 Kidney Disease: Improving Global Outcomes Clinical Practice Guideline for Acute Kidney Injury criteria,¹⁵ based on changes in serum creatinine. A procedure was labeled as AKI-positive if either an absolute increase in serum creatinine of at least 0.3 mg/dL or a relative increase of at least 50% (≥1.5-fold from baseline) was observed within 48 hours following PCI. Procedures without available baseline or postprocedural creatinine measurements were excluded from outcome labeling. Procedures with incompatible age values were also excluded before model development. To assess potential selection bias related to creatinine availability, selected baseline characteristics were compared between procedures included in the analytic cohort and procedures excluded owing to unavailable AKI outcome labeling. Differences were summarized using absolute standardized mean differences.

Predictor variables and data preprocessing

Predictor variables were restricted to information available at or before the time of PCI. These included demographic variables, comorbid conditions, baseline laboratory measurements, echocardiographic parameters, and clinical presentation at the time of PCI. All features underwent standardized preprocessing, including inspection for implausible values and consistency checks across the study period. No postprocedural variables were used as predictors.

Model development

A gradient boosting–based decision tree ensemble was used to predict post-PCI AKI. The data set was split into a development and a held-out test set at approximately 80/20. Splitting was performed with stratification by the binary AKI label and grouping by patient identifier, ensuring that all procedures from a given patient were assigned exclusively to a single partition.

Model development was conducted exclusively on the development set. CatBoost hyperparameters were optimized within the development set using 5-fold patient-aware cross-validation, with patient-level grouping maintained across folds. The held-out test set was not used for hyperparameter tuning, model selection, probability calibration, or threshold selection. After hyperparameter selection, performance estimation within the development set used 10-fold group-aware cross-validation, again maintaining patient-level grouping across folds. Class imbalance was addressed using built-in class weighting. The selected model configuration was used consistently for the primary analysis and temporal validation analysis, and the primary held-out test predictions were used unchanged for subgroup sensitivity analyses.

Probability calibration and decision threshold selection

Predicted probabilities from cross-validation were aggregated to generate out-of-fold (OOF) predictions for the development set. These OOF probabilities were used to train a logistic regression calibration model, which was subsequently applied to all predicted probabilities.

Decision thresholds were selected exclusively on the development set using calibrated OOF predictions. For threshold selection, calibrated OOF predictions from all 10 development-set cross-validation folds were combined into a single development-set prediction set. Each operating threshold was then selected once using this combined OOF prediction set and the corresponding observed AKI labels; thresholds were not optimized separately within individual folds. Three clinically motivated operating points were defined: 1) a balanced operating point, selected to maximize balanced accuracy; 2) a screening operating point, selected to optimize precision under a minimum sensitivity constraint; and 3) a rule-in operating point, selected to optimize precision under a minimum specificity constraint. For the screening operating point, a minimum sensitivity requirement of 0.80 was imposed, whereas for the rule-in operating point, a minimum specificity requirement of 0.90 was used. Once selected, thresholds were fixed and applied to the held-out test set without change.

Model evaluation

Final model performance was evaluated on the held-out test set using calibrated probabilities and the prespecified decision thresholds. Discrimination was assessed using the area under the receiver operating characteristic curve (ROC-AUC) and the precision-recall curve (PR-AUC). Calibration was assessed using the integrated calibration index (ICI) and Brier score. Clinically, the ICI reflects how closely predicted AKI probabilities agree with observed AKI rates across the risk spectrum, with lower values indicating better agreement. The Brier score reflects the overall accuracy of predicted risk estimates at the individual-procedure level, with lower values indicating more reliable probabilistic predictions. Classification performance was summarized using sensitivity, specificity, precision (positive predictive value [PPV]), F1 score, accuracy, and balanced accuracy. CIs (95%) were estimated using nonparametric bootstrapping with 1,000 iterations. Decision curve analysis was performed as an additional assessment of potential clinical utility. Net benefit was calculated across threshold probabilities and compared with default strategies of treating all procedures as high risk or treating no procedures as high risk.

For baseline comparisons in Table 1, continuous and ordinal variables were compared using the Wilcoxon rank sum test, and categorical variables were compared using the chi-square test or Fisher exact test, as appropriate. Percentages were calculated using nonmissing observations for each variable. All P values are descriptive and were not adjusted for multiple comparisons.

Table 1.

Baseline Characteristics of the Study Population

	No AKI (n = 22,061)	AKI (n = 1,667)	P Value
Age, y	65 (57-74)	73 (64-81)	<0.001
Female	4,771 (21.6%)	547 (32.8%)	<0.001
Diabetes mellitus	10,671 (48.4%)	1,014 (60.8%)	<0.001
Hypertension	16,937 (76.8%)	1,449 (86.9%)	<0.001
Congestive heart failure	6,259 (28.4%)	993 (59.6%)	<0.001
Prior myocardial infarction	4,536 (20.6%)	407 (24.4%)	<0.001
Peripheral vascular disease	1,481 (6.7%)	208 (12.5%)	<0.001
Prior stroke	1,428 (6.5%)	199 (11.9%)	<0.001
Atrial Fibrillation (known history)	1,880 (43.8%)	159 (28.4%)	<0.001
CHA2DS2-VASc score	4 (2-5)	5 (4-6)	<0.001
Serum creatinine, mg/dL	0.92 (0.79-1.10)	1.33 (0.96-2.26)	<0.001
Hemoglobin, g/dL	13.50 (12.20-14.60)	11.80 (10.30-13.50)	<0.001
White blood cell count, ×109/L	8 (6-10)	9 (7-13)	<0.001
Platelet count, ×109/L	222 (184-268)	229 (181-285)	0.002
Uric acid, mg/dL	5.90 (4.90-7.00)	6.90 (5.50-8.60)	<0.001
Fibrinogen, mg/dL	426 (357-513)	493 (385.25-598)	<0.001
Left ventricular ejection fraction, %	55 (40-60)	45 (35-60)	<0.001
PCI for acute coronary syndrome	6,621 (30.0%)	283 (17.0%)	<0.001
Myocardial infarction-related PCI	9,039 (41.0%)	949 (56.9%)	<0.001

Values are presented as median (interquartile range) for continuous and ordinal variables and n (%) for categorical variables. Percentages were calculated using non-missing observations for each variable. P values are descriptive and were calculated using the Wilcoxon rank-sum test for continuous and ordinal variables and the χ² test or Fisher’s exact test for categorical variables, as appropriate.

AKI = acute kidney injury; PCI = percutaneous coronary intervention.

Internal temporal validation

As a supportive robustness analysis, an additional internal temporal validation was performed within the same labeled cohort. Using PCI date, the data set was repartitioned into an earlier development set and a later test set with a cutoff of January 1, 2019. To prevent information leakage, patient-level separation was maintained across partitions, such that patients contributing procedures to the later period were excluded from the earlier development set. A separate model was then trained on the earlier-period development set using the same modeling pipeline as in the primary analysis, and its performance was evaluated on the later-period test set.

Sensitivity analysis in patients with reduced baseline kidney function

As a separate sensitivity analysis based on the primary held-out test set, model performance was additionally evaluated among procedures with reduced baseline kidney function. Estimated glomerular filtration rate (eGFR) was calculated from baseline serum creatinine, age, and sex using the 2021 Chronic Kidney Disease Epidemiology Collaboration creatinine equation, and the subgroup was defined as baseline eGFR <45 mL/min/1.73 m². Discrimination in this subgroup was assessed using ROC-AUC and PR-AUC, and calibration was assessed using ICI and Brier score calculated from calibrated held-out test-set probabilities.

Model explainability

Model explainability was assessed using complementary approaches. SHapley Additive exPlanations (SHAP) values were used as the primary method to quantify the direction and magnitude of individual predictor contributions to model output at the sample level. SHAP values were computed from the trained CatBoost model before Platt calibration and therefore reflect predictor contributions to the model's underlying output rather than to calibrated probabilities. In addition, feature importance based on relative influence on model predictions was extracted from the trained CatBoost model as a supporting analysis.

Ethical considerations

This study was conducted in accordance with institutional policies governing retrospective analyses of de-identified clinical data.

Results

Patient characteristics

The initial data set included 31,370 PCI procedures. After applying the inclusion and exclusion criteria, including exclusion procedures without sufficient creatinine data for AKI outcome labeling, the final analytic cohort consisted of 23,728 PCI procedures performed on 17,943 unique patients (Figure 1). The observed prevalence of AKI was 7.0% (1,667 events). Patients who developed AKI were significantly older (median 73 vs 65 years, P < 0.001) and had a higher burden of comorbidities, including diabetes (60.8% vs 48.4%, P < 0.001) and congestive heart failure (59.6% vs 28.4%, P < 0.001). AKI patients had higher baseline inflammatory markers, including WBC count, uric acid (median 6.9 vs 5.9 mg/dL, P < 0.001), and fibrinogen (493 vs 426 mg/dL, P < 0.001). Further information about baseline characteristics is available in Table 1.

Figure 1.

Patient Flow Diagram

AKI = acute kidney injury.

Selected baseline characteristics of procedures included in the analytic cohort and procedures excluded owing to unavailable AKI outcome labeling, along with their standard mean deviation differences, are shown in Supplemental Table 1.

Overall model performance

Using 10-fold patient-grouped cross-validation on the development set, the model demonstrated stable discrimination and precision-recall performance across folds. The mean cross-validated ROC-AUC was 0.84 ± 0.02, and mean PR-AUC was 0.37 ± 0.05. Calibration was assessed using cross-validated predictions, yielding an ICI of 0.017 and a Brier score of 0.053 (Figure 2). Decision thresholds were selected using calibrated OOF predictions from the development set. The balanced operating point, chosen to maximize balanced accuracy, corresponded to a decision threshold of 0.056. On the held-out test set, the model achieved a ROC-AUC of 0.85 (95% CI: 0.82-0.87, Figure 3) and a PR-AUC of 0.38 (95% CI: 0.32-0.43, Figure 3). At the balanced operating point (threshold = 0.056), sensitivity was 0.78 (95% CI: 0.73-0.82), specificity was 0.77 (95% CI: 0.76-0.79), and PPV was 0.21 (95% CI: 0.19-0.23), with a balanced accuracy of 0.78 (95% CI: 0.75-0.80).

Figure 2.

Calibration of Predicted Acute Kidney Injury Risk

Calibration plot comparing predicted probabilities with observed event rates based on out-of-fold predictions from the development set obtained using 10-fold cross-validation. The dashed line indicates perfect calibration.

Figure 3.

Discrimination Performance on the Held-Out Test Set

(A) Receiver operating characteristic (ROC) curve and (B) precision-recall (PR) curve for prediction of post-PCI acute kidney injury on the test set. PR-AUC = precision–recall curve; ROC-AUC = area under the receiver operating characteristic curve.

Performance of the screening and rule-in operating points

The screening operating point, chosen to optimize precision under a minimum sensitivity constraint of 0.80, corresponded to a threshold of 0.041. The rule-in operating point, chosen to optimize precision under a minimum specificity constraint of 0.90, corresponded to a threshold of 0.199. These thresholds were fixed and applied unchanged for evaluation on the held-out test set.

At the screening operating point (threshold = 0.041), sensitivity increased to 0.83 (95% CI: 0.80-0.87), with a corresponding PPV of 0.18 (95% CI: 0.16-0.20) and specificity of 0.70 (95% CI: 0.69-0.72). At the rule-in operating point (threshold = 0.199), specificity reached 0.92 (95% CI: 0.91-0.93), with a PPV of 0.33 (95% CI: 0.29-0.38) and sensitivity of 0.52 (95% CI: 0.47-0.57). Performance metrics across the 3 operating points, reflecting trade-offs between sensitivity and precision aligned with their intended clinical use, are summarized in Table 2.

Table 2.

Model Performance at Selected Clinical Thresholds

Operating Point	Threshold	Sensitivity	Specificity	Precision (PPV)	F1 Score
Balanced	0.056	0.78 (0.73-0.82)	0.77 (0.76-0.79)	0.21 (0.19-0.23)	0.33 (0.30-0.36)
Screening	0.041	0.83 (0.80-0.87)	0.70 (0.69-0.72)	0.18 (0.16-0.20)	0.29 (0.27-0.32)
Rule-in	0.199	0.52 (0.47-0.57)	0.92 (0.91-0.93)	0.33 (0.29-0.38)	0.41 (0.36-0.45)

Values are presented as point estimates with 95% confidence intervals obtained by nonparametric bootstrapping. Thresholds were selected using calibrated out-of-fold predictions from the development set and applied unchanged to the held-out test set.

PPV = positive predictive value.

Decision curve analysis supported potential clinical utility of the model, with greater net benefit than default treat-all or treat-none strategies across the evaluated range of threshold probabilities, including the selected screening, balanced, and rule-in operating points (Supplemental Figure 1).

Internal temporal validation

In a supportive internal temporal validation, the cohort was repartitioned using the prespecified temporal split, yielding 17,370 procedures in the earlier development set and 4,965 procedures in the later test set. AKI prevalence was 6.9% in the earlier-period development set and 8.5% in the later-period test set. In the later-period test set, the temporally trained model achieved a ROC-AUC of 0.85 (95% CI: 0.83-0.87) and a PR-AUC of 0.41 (95% CI: 0.37-0.46), indicating preserved discriminative performance under a temporally separated evaluation design. Performance at the selected operating points remained clinically meaningful and is summarized in Table 3.

Table 3.

Model Performance at Selected Clinical Thresholds in the Internal Temporal Validation Analysis

Operating Point	Threshold	Sensitivity	Specificity	Precision (PPV)	F1 Score
Balanced	0.056	0.86 (0.83-0.89)	0.63 (0.61-0.64)	0.18 (0.16-0.19)	0.29 (0.27-0.31)
Screening	0.041	0.92 (0.89-0.94)	0.46 (0.45-0.48)	0.14 (0.12-0.15)	0.24 (0.22-0.26)
Rule-in	0.199	0.64 (0.59-0.68)	0.88 (0.87-0.89)	0.33 (0.30-0.37)	0.44 (0.40-0.47)

Internal temporal validation was performed using a strict patient-aware split based on PCI date, as described in Methods. A separate model was trained on the earlier-period development set using the same pipeline as in the primary analysis, and performance was evaluated on the later-period test set. Values are presented as point estimates with 95% CIs obtained by nonparametric bootstrapping.

Abbreviation as in Table 2.

Sensitivity analysis in patients with reduced baseline kidney function

In a sensitivity analysis based on the primary held-out test set, the subgroup with baseline eGFR <45 mL/min/1.73 m² included 570 procedures and had an AKI prevalence of 27.9%. In this subgroup, the model achieved a ROC-AUC of 0.73 (95% CI: 0.68-0.78) and a PR-AUC of 0.53 (95% CI: 0.46-0.61). Calibration assessment using calibrated held-out test-set probabilities yielded an ICI of 0.069 and a Brier score of 0.175.

Discussion

In this study, we developed and validated a ML model for predicting AKI after PCI. The model, trained on over 23,000 procedures, achieved an area under the curve (AUC) of 0.85 (Central Illustration). By using strictly preprocedural data, this tool enables risk stratification before the patient enters the catheterization laboratory, thereby supporting true preventive medicine. We evaluated a single calibrated prediction model across multiple operating thresholds corresponding to distinct clinical objectives: one for screening patients at high risk and another for rule-in use during hospitalization.

Central Illustration.

Machine Learning Prediction of Acute Kidney Injury After PCI

This central illustration summarizes the development, performance, and intended clinical application of a machine learning model for predicting acute kidney injury after PCI using preprocedural EHR data. AKI = acute kidney injury; EHR = electronic health record; HbA1c = glycosylated hemoglobin; ICI = integrated calibration index; LDL = low-density lipoprotein; PCI = percutaneous coronary intervention; PR-AUC = precision–recall curve; ROC-AUC = area under the receiver operating characteristic curve; SHAP = SHapley Additive exPlanations.

As a supportive robustness analysis, we also performed an internal temporal validation using a strict patient-aware split based on PCI date. Under this temporally separated validation design, model discrimination remained stable, with a ROC-AUC of 0.85 and a PR-AUC of 0.41 in the later-period test set. Although this analysis was performed within the same institutional data set, the preservation of performance under temporal separation supports the robustness of the model to changes in practice over time.

In a sensitivity analysis restricted to patients with baseline eGFR <45 mL/min/1.73 m², model discrimination was lower than in the overall test set, with a ROC-AUC of 0.73 and a PR-AUC of 0.53. This finding is not unexpected, given the substantially higher AKI prevalence and narrower clinical profile of this subgroup. Nevertheless, the model retained measurable discriminative ability in a clinically important population with impaired baseline kidney function.

When comparing our base model to other commonly used ones like the Mehran risk score or the National Cardiovascular Data Registry CathPCI Model, direct head-to-head comparison was not feasible because these models rely on intraprocedural variables and covariates not consistently available in our data set; therefore, we did not perform a formal comparative analysis. Reported AUC values for these models in prior studies are typically in the range of ∼0.71 to 0.75,16, 17, 18 whereas our model achieved an AUC of 0.85. However, these comparisons should be interpreted cautiously given differences in study design, populations, and available variables. We believe one of the key drivers of this performance is the use of the CatBoost algorithm, which effectively models nonlinear interactions between variables. Most logistic regression models assume a linear correlation between the variables and the outcome, but physiological systems are rarely linear. For instance, the impact of systolic blood pressure on renal risk is likely nonlinear and dependent on baseline renal function—nuances that simple integer-based scores fail to capture.¹⁶ Furthermore, our focus on calibration (ICI 0.017) ensures that predicted risk probabilities correspond to observed event rates, building clinician trust in the tool. The model is applicable in the preprocedural setting, as it does not depend on intraprocedural variables, thereby enabling advance risk stratification and implementation of preventive measures.

Several findings from our model are noteworthy. Variables such as uric acid and fibrinogen, which are not routinely incorporated into traditional AKI risk scores, emerged as prominent predictors in the SHAP (Figure 4) analysis. Hyperuricemia is increasingly recognized as a marker of oxidative stress and endothelial dysfunction, potentially contributing to renal vasoconstriction and tubular injury in the setting of contrast exposure.^19,20 Similarly, elevated fibrinogen reflects systemic inflammation and increased blood viscosity, both of which may impair renal microcirculation and increase susceptibility to nephrotoxic injury. These observations are consistent with prior studies identifying these markers as independent predictors of contrast-associated AKI.^21,22 At the same time, several leading predictors, including baseline creatinine and hemoglobin, are well-established risk factors. The added value of the ML approach therefore lies not in identifying entirely novel variables, but in its ability to model complex, nonlinear interactions among predictors and generate well-calibrated, individualized risk estimates.

Figure 4.

Feature Attribution Using SHAP Values

Summary plot of Shapley Additive exPlanations (SHAP) values for the test set, illustrating the direction and magnitude of each feature's contribution to predicted AKI risk. Positive SHAP values indicate increased predicted risk, whereas negative values indicate decreased predicted risk. Points are colored according to the corresponding feature value. HbA1c = glycosylated hemoglobin; LDL = low-density lipoprotein; PCI = percutaneous coronary intervention.

We further evaluated the calibrated model across 2 additional operating thresholds corresponding to screening and rule-in clinical objectives. In our testing, the screening operating threshold demonstrated strong discrimination, achieving a sensitivity of 0.83, and the rule-in operating threshold achieved a specificity of 0.92. In a cohort with a 7% AKI prevalence, the observed PPVs across operating points (0.18-0.33) represent a substantial enrichment of risk relative to baseline, and the PR-AUC of 0.38 markedly exceeds the prevalence baseline, indicating meaningful precision-recall performance in this low-event-rate setting. To quantify the clinical utility, we did an exploratory decision curve analysis, showing positive net benefit across relevant thresholds compared to treat-all and treat-none, highlighting the potential value of model-guided risk stratification. Once validated, these modes can support 2 management pathways: patients below the screening threshold may be safely discharged earlier, or they may undergo less aggressive hydration, thereby sparing resources. Conversely, patients exceeding the rule-in threshold (>20% predicted risk) are candidates for resource-intensive preventive measures, such as the RenalGuard²³ or ultralow contrast PCI guided by intravascular ultrasound,²⁴ thereby enabling truly personalized medicine. To enhance clinical interpretability, consider an elderly patient undergoing PCI with reduced baseline renal function, anemia, and elevated inflammatory markers (eg, fibrinogen, uric acid). In such a profile, the model assigns a higher predicted risk of AKI by integrating multiple risk factors into a single individualized estimate. In patients identified as high risk, this may support targeted preventive strategies, including preprocedural hospitalization for aggressive hydration, cessation of nephrotoxic agents when feasible, contrast minimization, and postprocedural hydration with close laboratory monitoring of renal function.

Study Limitations

This study has several limitations. First, it is a retrospective single-center study; external validation in diverse populations is necessary to confirm generalizability. Although a temporally separated internal validation was performed as a supportive robustness analysis, this does not substitute for external validation. Second, AKI was defined using serum creatinine–based Kidney Disease: Improving Global Outcomes criteria, and urine output data were not available, which may have led to an underestimation of AKI incidence. Finally, patients with pre-existing renal conditions, including chronic dialysis, kidney transplantation, or preprocedural AKI, were not excluded, reflecting real-world clinical practice but potentially introducing clinical heterogeneity and affecting outcome specificity in certain subgroups.

Conclusions

In this single-center study, we developed a ML model that predicts post-PCI AKI using readily available preprocedural variables. The model demonstrates strong discriminative performance in a purely preprocedural setting, with potential advantages over traditional scores that rely on intraprocedural variables. By providing calibrated risk estimates and clinically relevant decision thresholds, this approach may support individualized risk stratification and targeted preventive strategies.

Funding support and author disclosures

The authors have reported that they have no relationships relevant to the contents of this paper to disclose.