Machine Learning Prediction and Causal Forest Analysis of Severe Acute Kidney Injury in ICU Patients with COPD: A MIMIC-IV Study Authors: LiFeng Fang, Fei Wang, Junkang Chen, Yulong Zheng Published: 2026 ABSTRACT BackgroundSevere acute kidney injury (AKI) is frequent in critically ill patients with chronic obstructive pulmonary disease (COPD), but predictive importance does not establish causality.ObjectiveTo model severe AKI (KDIGO stage 3), evaluate temporal robustness at a 24-hour ICU landmark, and explore adjusted exposure-effect estimates.MethodsThis retrospective MIMIC-IV v3.1 cohort included 4,705 adults with COPD. The original modeling dataset was split into training (n = 3,294) and testing (n = 1,411) sets. Consensus features from LASSO, Boruta, and random-forest importance informed 15 classifiers; the highest observed test-set AUC was interpreted with SHAP. A 24-hour landmark analysis excluded earlier stage 3 AKI and used pre-landmark predictors for incident stage 3 AKI thereafter. CausalForestDML estimated exploratory adjusted effects of the most recent valid creatinine within 365 days before ICU admission.ResultsSevere AKI occurred in 1,085 patients (23.06%). The original six-feature CatBoost model included total ICU length of stay and had an AUC of 0.848 (95% CI, 0.825–0.872); it is interpreted as a retrospective trajectory model. The landmark cohort included 4,171 patients and 841 events. The temporal five-feature AUC was 0.679 (95% CI, 0.641–0.717); without pre-ICU creatinine it was 0.686 (95% CI, 0.648–0.724; paired difference, −0.0067; P = 0.293). The adjusted creatinine risk difference was 0.0547 per 1 mg/dL (95% CI, −0.0021 to 0.1115) and 0.0705 (95% CI, 0.0022–0.1387) after 1st-99th percentile trimming.ConclusionThis study establishes a COPD-specific benchmark showing that a parsimonious model can characterize severe-AKI trajectories and that temporal separation materially changes performance and interpretation. Integrating model comparison, SHAP, adjusted-effect estimation, and landmark validation provides a rigorous framework for distinguishing prognostic importance from causal relevance and defines priorities for external validation. Introduction Acute kidney injury (AKI) is common in intensive care unit (ICU) patients and is associated with mortality, prolonged hospitalization, and progression to chronic kidney disease.1–8 Patients with chronic obstructive pulmonary disease (COPD) may be particularly vulnerable because acute respiratory failure, hypoxemia, hypercapnia, systemic inflammation, right-heart dysfunction, hemodynamic instability, infection, mechanical ventilation, diuretics, vasoactive drugs, and potentially nephrotoxic antimicrobials can converge during critical illness.9–15 For pulmonologists and critical-care clinicians, identifying patients at high renal risk may support closer renal monitoring and medication review, but prediction alone cannot identify modifiable causes. Machine-learning models have improved AKI risk prediction in general ICU and disease-specific cohorts, with gradient-boosting algorithms often achieving strong discrimination.16–27 However, existing studies differ substantially in prediction horizon, endpoint definition, feature burden, interpretability, and validation strategy. A model that uses variables measured after or during outcome development can show high retrospective discrimination without serving as an admission-time prediction tool. Similarly, a variable that predicts AKI may reflect severity or incorporation bias rather than a modifiable causal pathway. Causal machine-learning methods, including generalized random forests and double/debiased machine learning, can estimate adjusted and heterogeneous exposure-effect relationships under assumptions of temporal ordering, conditional exchangeability, positivity, and consistency.28–31 These methods reduce bias from flexible nuisance-model estimation, but they do not automatically solve unmeasured confounding, reverse causality, or overlap between predictor and outcome definitions. Their results in observational ICU data should therefore be interpreted as exploratory unless supported by temporally valid designs and external validation. Evidence combining prediction and causal machine-learning analysis in ICU patients with COPD remains limited. The COPD-specific value of the present study lies in evaluating a parsimonious severe-AKI model in this high-risk population and in testing whether the variables with the greatest predictive importance also yield non-zero adjusted exposure-effect estimates. This study had two objectives: first, to compare 15 algorithms and develop an interpretable model for severe AKI in ICU patients with COPD; and second, to use causal forest analysis exploratorily to assess whether predictive importance and adjusted effect estimates converged. We hypothesized that predictive importance and adjusted effect estimates would not necessarily coincide. Methods Study Design and Data Source This retrospective cohort study used the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, version 3.1.32 MIMIC-IV is a credentialed-access, deidentified electronic health record database maintained by the Massachusetts Institute of Technology and hosted on PhysioNet. It contains ICU and hospital data from Beth Israel Deaconess Medical Center between 2008 and 2019. The collection of patient information and creation of the MIMIC-IV research resource were reviewed by the Institutional Review Board at Beth Israel Deaconess Medical Center, which approved the data-sharing initiative and granted a waiver of informed consent. Database access was obtained by LiFeng Fang after completion of the required CITI Data or Specimens Only Research refresher course (completion record 78374867; completed 16 July 2026 and valid through 16 July 2029) and acceptance of the PhysioNet data use agreement. The present study used deidentified data only. This study was approved by the Medical Ethics Committee of Huai’an Second People’s Hospital under approval no. HBYLL202572. The investigators had no direct interaction with participants and no access to identifiable private information. This study followed the Declaration of Helsinki, the PhysioNet data use agreement, and STROBE reporting guidance.33 Study Population The study population consisted of adult ICU patients with COPD. COPD was identified by ICD-10 codes in the J44.x family during the index hospitalization. For exploratory phenotype analyses, patients were stratified as J44.1-coded acute exacerbation of COPD (AECOPD), J44.0-coded COPD with acute lower respiratory infection, or other J44.x-coded COPD. Absence of a J44.1 code was not interpreted as clinically stable COPD. The inclusion criteria were age 18 years or older, first ICU stay for the hospitalization, COPD diagnosis, and ICU length of stay of at least 24 hours. Patients with pre-existing end-stage renal disease or chronic dialysis were excluded. The final reproduced cohort comprised 4,705 patients (Supplementary Table S1). Outcome Definition The primary outcome in the original analysis was severe AKI, defined as KDIGO stage 3 during the ICU stay.34 Staging used serum creatinine criteria and renal replacement therapy; urine-output criteria were not used because urine-output documentation was inconsistent. For the 24-hour landmark analysis, time zero was ICU admission and the landmark was ICU admission plus 24 hours. Baseline serum creatinine was defined as the most recent valid value (0.1–20 mg/dL) recorded during the 365 days before ICU admission. Incident severe AKI was defined as the first KDIGO stage 3 observation at or after the landmark and before ICU discharge. Patients with stage 3 AKI before the landmark or no post-landmark KDIGO observation were excluded (Supplementary Tables S1 and S2). Data Extraction and Variable Selection Data extraction was performed using SQL in PostgreSQL interfaced with MIMIC-IV v3.1. The original retrospective trajectory analysis used demographic characteristics, vital signs, laboratory tests, disease-severity scores, comorbidities, interventions recorded during the ICU stay, and outcomes including total ICU length of stay, hospital length of stay, 7-day mortality, and 28-day mortality. The temporal sensitivity analysis used re-extracted time-stamped source tables. Predictors for the landmark model were restricted to information available before the 24-hour landmark: most recent pre-ICU creatinine, first-24-hour anion gap, first-24-hour heart rate, SOFA, and SAPS II. Total ICU length of stay was excluded from the landmark model because it is partly measured after AKI onset (Supplementary Table S2). Detailed extraction logic and analytic procedures are provided in the Supplementary Methods. Temporal model performance and paired comparisons are reported in Supplementary Tables S3 and S4, and ICD-coded COPD phenotype distributions and subgroup discrimination estimates are reported in Supplementary Tables S5 and S6. Missing Data Handling Variables with more than 30% missing values were excluded from the analysis to avoid introducing excessive imputation bias. For the remaining variables with less than 30% missing data, multiple imputation by chained equations (MICE) was employed to address missing values.35 The imputation model incorporated all candidate predictor variables and the outcome variable to preserve the underlying data structure and reduce potential bias introduced by missingness. Five imputed datasets were generated, and the pooled estimates were used for subsequent analyses. The completeness and plausibility of the imputed data were verified through diagnostic checks, including comparison of distributions between observed and imputed values. Feature Selection The original modeling dataset was split into a training set (70%, n = 3,294) and a testing set (30%, n = 1,411) using stratified sampling to maintain the severe-AKI event proportion. Feature selection was performed exclusively on the training set. LASSO regression with 10-fold cross-validation identified 8 candidate variables.36 Boruta confirmed 34 features as important after 100 iterations.37 Random forest importance ranked by Mean Decrease Gini selected the top 10 variables.38 The intersection of all three methods yielded six consensus predictors: total ICU length of stay, serum creatinine, SAPS II, SOFA score, anion gap, and heart rate. Because total ICU length of stay is measured over the full stay, this six-feature model is interpreted as a retrospective trajectory model rather than an admission-time model. Machine Learning Model Development and Comparison Using the six consensus features, 15 machine-learning classifiers were trained on the training set and evaluated on the held-out testing set. The models were random forest, XGBoost tree booster, radial-basis support vector machine, logistic regression, k-nearest neighbors, partial least squares discriminant model, gradient boosting machine, single-hidden-layer neural network, naive Bayes, linear discriminant analysis, elastic-net logistic regression, AdaBoost. M1, XGBoost linear booster, CatBoost, and LightGBM.39–41 Hyperparameters were optimized by cross-validation on the training set. Model performance was assessed using AUC with 95% confidence intervals, accuracy, sensitivity, specificity, PPV, NPV, F1 score, Youden’s index, decision curve analysis, residual analysis, calibration curves, and clinical impact curves. Feature-removal sensitivity models excluded total ICU length of stay, creatinine, or both variables (Supplementary Table S7). Paired model comparisons are reported in Supplementary Table S8. Model Interpretability The optimal machine learning model was interpreted using SHapley Additive exPlanations (SHAP),22 a game theory–based framework that quantifies the marginal contribution of each feature to individual predictions. Global feature importance was measured by the mean absolute SHAP value across all samples, providing an overall ranking of predictor relevance. A beeswarm plot was generated to visualize the distribution of SHAP values for each feature, illustrating the directionality and magnitude of feature effects. SHAP dependence plots were constructed for each predictor to examine the nonlinear relationship between feature values and their contribution to the model output, with interaction effects color-coded by a secondary feature. Individual-level prediction explanations were provided through waterfall plots and force plots for representative patient samples, demonstrating the cumulative and directional contributions of each feature to the final predicted probability. Causal Inference Analysis We used CausalForestDML from the EconML Python library to estimate exploratory adjusted exposure-effect relationships.30 The estimator combines generalized random forests with double/debiased machine learning and cross-fitting.28,29 In this observational application, estimates require temporal ordering, positivity, consistency, and no unmeasured confounding. These assumptions cannot be guaranteed in MIMIC-IV; therefore, results are described as exploratory adjusted effect estimates rather than definitive causal effects. The original CausalForestDML analysis estimated average treatment effects for five grouped interventions and six continuous exposures, but several original variables were vulnerable to temporal overlap or outcome-definition concerns. We corrected the population mean-effect confidence intervals and added exposure counts for grouped interventions. We also performed a restricted-adjustment sensitivity analysis that excluded total ICU length of stay, creatinine, CRRT, survival-derived coding variables, and other variables with clear temporal or outcome-definition concerns. For the temporally ordered effect analysis, the exposure was the most recent valid pre-ICU creatinine and the outcome was incident severe AKI after the 24-hour landmark. Adjustment was restricted to variables available before the exposure or at admission, including age, sex, ICD-coded COPD phenotype, admission type, Charlson comorbidities, and Charlson comorbidity index. Post-exposure SOFA, SAPS II, anion gap, heart rate, and total ICU length of stay were deliberately excluded. CATE distributions were examined descriptively (Supplementary Table S9). Survival Analysis Kaplan–Meier survival curves were constructed to compare survival probabilities between patients with and without severe AKI at 7-day and 28-day time points. Between-group survival differences were assessed using the Log rank test. Hazard ratios (HRs) with 95% CIs were estimated using the Cox proportional hazards model to quantify the association between severe AKI and mortality risk. Statistical Analysis Baseline characteristics were compared between patients with and without severe AKI. Continuous variables with normal distributions were expressed as mean ± standard deviation and compared using independent-samples t tests. Continuous variables with non-normal distributions were presented as median (interquartile range) and compared using the Mann–Whitney U-test. Categorical variables were expressed as frequencies and percentages and compared using chi-square or Fisher’s exact tests. AUC confidence intervals and paired AUC comparisons used DeLong methods. CatBoost was the prespecified reference for comparisons against the other 14 models, and Holm correction was applied (Supplementary Tables S4 and S8). Python, R, scikit-learn, CatBoost, LightGBM, XGBoost, EconML, SHAP, pROC, and lifelines were used for analysis. Two-sided P < 0.05 was considered statistically significant. Results Study Population and Baseline Characteristics A total of 4,705 adult ICU patients with an ICD-10 J44.x COPD diagnosis recorded during the index hospitalization were included in the final analysis. Among them, 1,085 (23.06%) developed severe AKI (KDIGO stage 3), while 3,620 (76.94%) did not. Baseline characteristics stratified by severe AKI status are presented in Table 1. The mean age was 71.81 ± 10.92 years, and 53.09% were male. No significant differences were observed in age (P = 0.25) or sex (P = 0.65) between the two groups. Patients with severe AKI had significantly longer ICU stays (8.55 ± 9.58 vs 3.65 ± 4.05 days, P < 0.0001) and hospital stays (17.12 ± 15.93 vs 11.57 ± 11.13 days, P < 0.0001). Both 7-day mortality (15.12% vs 3.51%, P < 0.0001) and 28-day mortality (25.35% vs 4.48%, P < 0.0001) were substantially higher in the severe AKI group.Table 1Baseline Characteristics of ICU Patients with COPD Stratified by Severe AKI StatusVariableOverall(n = 4,705)No Severe AKI(n = 3,620)Severe AKI(n = 1,085)StatisticP valueDemographicsSex0.200.65 Female2207(46.91)1705(47.10)502(46.27) Male2498(53.09)1915(52.90)583(53.73)Age, years71.81 ± 10.9271.71 ± 11.0172.13 ± 10.60−1.140.25Clinical outcomesHospital length of stay, days12.85 ± 12.6211.57 ± 11.1317.12 ± 15.93−10.73<0.0001Total ICU length of stay, days4.78 ± 6.173.65 ± 4.058.55 ± 9.58−16.43<0.00017-day vital status191.83<0.0001 Alive4414(93.82)3493(96.49)921(84.88) Death291(6.18)127(3.51)164(15.12)28-day vital status429.10<0.0001 Alive4268(90.71)3458(95.52)810(74.65) Death437(9.29)162(4.48)275(25.35)Vital signsHeart rate, beats/min104.21 ± 21.38103.20 ± 20.45107.58 ± 23.93−5.46<0.0001Systolic blood pressure, mmHg148.99 ± 22.33148.80 ± 21.76149.62 ± 24.13−1.000.32Diastolic blood pressure, mmHg91.86 ± 21.5791.89 ± 20.7891.75 ± 24.010.180.86Mean blood pressure, mmHg108.21 ± 25.94107.63 ± 24.03110.16 ± 31.41−2.450.01Respiratory rate, breaths/min29.27 ± 6.5929.12 ± 6.6029.79 ± 6.50−3.00<0.01Temperature, degrees C37.29 ± 0.5937.28 ± 0.5737.34 ± 0.67−2.64<0.01SpO2, %99.30 ± 1.3299.30 ± 1.2499.29 ± 1.560.360.72ComorbiditiesIschemic stroke5.950.01 No4474(95.09)3458(95.52)1016(93.64) Yes231(4.91)162(4.48)69(6.36)Chronic kidney disease103.53<0.0001 No3273(69.56)2654(73.31)619(57.05) Yes1432(30.44)966(26.69)466(42.95)Myocardial infarction3.910.05 No3556(75.58)2761(76.27)795(73.27) Yes1149(24.42)859(23.73)290(26.73)Congestive heart failure41.20<0.0001 No2347(49.88)1899(52.46)448(41.29) Yes2358(50.12)1721(47.54)637(58.71)Diabetes6.430.01 No2957(62.85)2311(63.84)646(59.54) Yes1748(37.15)1309(36.16)439(40.46)Malignant cancer0.340.56 No3974(84.46)3051(84.28)923(85.07) Yes731(15.54)569(15.72)162(14.93)Laboratory valuesHemoglobin10.26 ± 1.9110.30 ± 1.8810.11 ± 1.992.79<0.01Platelet count207.26 ± 94.93207.69 ± 90.53205.84 ± 108.370.510.61RDW15.78 ± 2.4615.61 ± 2.4516.32 ± 2.41−8.41<0.0001WBC count13.08 ± 10.2012.62 ± 8.7914.63 ± 13.81−4.52<0.0001Anion gap14.30 ± 4.2713.65 ± 3.6316.48 ± 5.40−16.18<0.0001Bicarbonate24.70 ± 5.1525.05 ± 4.9923.54 ± 5.498.14<0.0001Blood urea nitrogen30.79 ± 23.7127.68 ± 21.0041.19 ± 28.76−14.38<0.0001Total calcium8.53 ± 0.718.53 ± 0.698.52 ± 0.780.370.71Creatinine1.53 ± 1.341.27 ± 0.982.41 ± 1.90−18.99<0.0001Glucose, mg/dL151.48 ± 74.84149.46 ± 75.25158.23 ± 73.10−3.44<0.001Sodium138.79 ± 4.86138.79 ± 4.74138.82 ± 5.26−0.170.87Potassium4.45 ± 0.704.39 ± 0.644.64 ± 0.84−8.84<0.0001INR1.59 ± 0.971.53 ± 0.851.80 ± 1.27−6.69<0.0001PT17.36 ± 11.0016.65 ± 9.6419.71 ± 14.42−6.57<0.0001APTT46.35 ± 31.7245.16 ± 30.6750.30 ± 34.73−4.39<0.0001PO2154.04 ± 117.87154.38 ± 120.65152.93 ± 108.120.380.71PCO255.50 ± 17.4355.12 ± 17.3056.79 ± 17.81−2.73<0.01pH7.39 ± 0.067.40 ± 0.067.38 ± 0.075.86<0.0001Severity scoresCharlson Comorbidity Index6.97 ± 2.746.79 ± 2.727.59 ± 2.74−8.40<0.0001SAPS II39.02 ± 12.8036.74 ± 11.0746.62 ± 15.05−20.07<0.0001SOFA score5.05 ± 3.454.39 ± 2.977.24 ± 4.02−21.70<0.0001Glasgow Coma Scale13.71 ± 2.5113.82 ± 2.3413.34 ± 2.984.89<0.0001Clinical interventionsDobutamine36.55<0.0001 No4621(98.21)3579(98.87)1042(96.04) Yes84(1.79)41(1.13)43(3.96)Any antibiotic62.74<0.0001 No837(17.79)732(20.22)105(9.68) Yes3868(82.21)2888(79.78)980(90.32)Vasoactive agent150.87<0.0001 No3076(65.38)2536(70.06)540(49.77) Yes1629(34.62)1084(29.94)545(50.23)Recorded diuretic exposure3.770.05 No1388(29.50)1094(30.22)294(27.10) Yes3317(70.50)2526(69.78)791(72.90)Mechanical ventilation208.08<0.0001 No2406(51.14)2060(56.91)346(31.89) Yes2299(48.86)1560(43.09)739(68.11)Notes: Data are presented as mean ± standard deviation for continuous variables and n (%) for categorical variables. P values compare patients without versus with severe AKI using independent-samples t tests for continuous variables and chi-square tests for categorical variables, as appropriate.Abbreviations: AKI, acute kidney injury; APTT, activated partial thromboplastin time; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; DBP, diastolic blood pressure; GCS, Glasgow Coma Scale; ICU, intensive care unit; INR, international normalized ratio; MBP, mean blood pressure; PCO2, partial pressure of carbon dioxide; PO2, partial pressure of oxygen; PT, prothrombin time; RDW, red-cell distribution width; SAPS II, Simplified Acute Physiology Score II; SBP, systolic blood pressure; SOFA, Sequential Organ Failure Assessment; SpO2, peripheral oxygen saturation; WBC, white blood cell. Table 1Baseline Characteristics of ICU Patients with COPD Stratified by Severe AKI StatusVariableOverall(n = 4,705)No Severe AKI(n = 3,620)Severe AKI(n = 1,085)StatisticP valueDemographicsSex0.200.65 Female2207(46.91)1705(47.10)502(46.27) Male2498(53.09)1915(52.90)583(53.73)Age, years71.81 ± 10.9271.71 ± 11.0172.13 ± 10.60−1.140.25Clinical outcomesHospital length of stay, days12.85 ± 12.6211.57 ± 11.1317.12 ± 15.93−10.73<0.0001Total ICU length of stay, days4.78 ± 6.173.65 ± 4.058.55 ± 9.58−16.43<0.00017-day vital status191.83<0.0001 Alive4414(93.82)3493(96.49)921(84.88) Death291(6.18)127(3.51)164(15.12)28-day vital status429.10<0.0001 Alive4268(90.71)3458(95.52)810(74.65) Death437(9.29)162(4.48)275(25.35)Vital signsHeart rate, beats/min104.21 ± 21.38103.20 ± 20.45107.58 ± 23.93−5.46<0.0001Systolic blood pressure, mmHg148.99 ± 22.33148.80 ± 21.76149.62 ± 24.13−1.000.32Diastolic blood pressure, mmHg91.86 ± 21.5791.89 ± 20.7891.75 ± 24.010.180.86Mean blood pressure, mmHg108.21 ± 25.94107.63 ± 24.03110.16 ± 31.41−2.450.01Respiratory rate, breaths/min29.27 ± 6.5929.12 ± 6.6029.79 ± 6.50−3.00<0.01Temperature, degrees C37.29 ± 0.5937.28 ± 0.5737.34 ± 0.67−2.64<0.01SpO2, %99.30 ± 1.3299.30 ± 1.2499.29 ± 1.560.360.72ComorbiditiesIschemic stroke5.950.01 No4474(95.09)3458(95.52)1016(93.64) Yes231(4.91)162(4.48)69(6.36)Chronic kidney disease103.53<0.0001 No3273(69.56)2654(73.31)619(57.05) Yes1432(30.44)966(26.69)466(42.95)Myocardial infarction3.910.05 No3556(75.58)2761(76.27)795(73.27) Yes1149(24.42)859(23.73)290(26.73)Congestive heart failure41.20<0.0001 No2347(49.88)1899(52.46)448(41.29) Yes2358(50.12)1721(47.54)637(58.71)Diabetes6.430.01 No2957(62.85)2311(63.84)646(59.54) Yes1748(37.15)1309(36.16)439(40.46)Malignant cancer0.340.56 No3974(84.46)3051(84.28)923(85.07) Yes731(15.54)569(15.72)162(14.93)Laboratory valuesHemoglobin10.26 ± 1.9110.30 ± 1.8810.11 ± 1.992.79<0.01Platelet count207.26 ± 94.93207.69 ± 90.53205.84 ± 108.370.510.61RDW15.78 ± 2.4615.61 ± 2.4516.32 ± 2.41−8.41<0.0001WBC count13.08 ± 10.2012.62 ± 8.7914.63 ± 13.81−4.52<0.0001Anion gap14.30 ± 4.2713.65 ± 3.6316.48 ± 5.40−16.18<0.0001Bicarbonate24.70 ± 5.1525.05 ± 4.9923.54 ± 5.498.14<0.0001Blood urea nitrogen30.79 ± 23.7127.68 ± 21.0041.19 ± 28.76−14.38<0.0001Total calcium8.53 ± 0.718.53 ± 0.698.52 ± 0.780.370.71Creatinine1.53 ± 1.341.27 ± 0.982.41 ± 1.90−18.99<0.0001Glucose, mg/dL151.48 ± 74.84149.46 ± 75.25158.23 ± 73.10−3.44<0.001Sodium138.79 ± 4.86138.79 ± 4.74138.82 ± 5.26−0.170.87Potassium4.45 ± 0.704.39 ± 0.644.64 ± 0.84−8.84<0.0001INR1.59 ± 0.971.53 ± 0.851.80 ± 1.27−6.69<0.0001PT17.36 ± 11.0016.65 ± 9.6419.71 ± 14.42−6.57<0.0001APTT46.35 ± 31.7245.16 ± 30.6750.30 ± 34.73−4.39<0.0001PO2154.04 ± 117.87154.38 ± 120.65152.93 ± 108.120.380.71PCO255.50 ± 17.4355.12 ± 17.3056.79 ± 17.81−2.73<0.01pH7.39 ± 0.067.40 ± 0.067.38 ± 0.075.86<0.0001Severity scoresCharlson Comorbidity Index6.97 ± 2.746.79 ± 2.727.59 ± 2.74−8.40<0.0001SAPS II39.02 ± 12.8036.74 ± 11.0746.62 ± 15.05−20.07<0.0001SOFA score5.05 ± 3.454.39 ± 2.977.24 ± 4.02−21.70<0.0001Glasgow Coma Scale13.71 ± 2.5113.82 ± 2.3413.34 ± 2.984.89<0.0001Clinical interventionsDobutamine36.55<0.0001 No4621(98.21)3579(98.87)1042(96.04) Yes84(1.79)41(1.13)43(3.96)Any antibiotic62.74<0.0001 No837(17.79)732(20.22)105(9.68) Yes3868(82.21)2888(79.78)980(90.32)Vasoactive agent150.87<0.0001 No3076(65.38)2536(70.06)540(49.77) Yes1629(34.62)1084(29.94)545(50.23)Recorded diuretic exposure3.770.05 No1388(29.50)1094(30.22)294(27.10) Yes3317(70.50)2526(69.78)791(72.90)Mechanical ventilation208.08<0.0001 No2406(51.14)2060(56.91)346(31.89) Yes2299(48.86)1560(43.09)739(68.11)Notes: Data are presented as mean ± standard deviation for continuous variables and n (%) for categorical variables. P values compare patients without versus with severe AKI using independent-samples t tests for continuous variables and chi-square tests for categorical variables, as appropriate.Abbreviations: AKI, acute kidney injury; APTT, activated partial thromboplastin time; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; DBP, diastolic blood pressure; GCS, Glasgow Coma Scale; ICU, intensive care unit; INR, international normalized ratio; MBP, mean blood pressure; PCO2, partial pressure of carbon dioxide; PO2, partial pressure of oxygen; PT, prothrombin time; RDW, red-cell distribution width; SAPS II, Simplified Acute Physiology Score II; SBP, systolic blood pressure; SOFA, Sequential Organ Failure Assessment; SpO2, peripheral oxygen saturation; WBC, white blood cell. Baseline Characteristics of ICU Patients with COPD Stratified by Severe AKI Status Notes: Data are presented as mean ± standard deviation for continuous variables and n (%) for categorical variables. P values compare patients without versus with severe AKI using independent-samples t tests for continuous variables and chi-square tests for categorical variables, as appropriate. Abbreviations: AKI, acute kidney injury; APTT, activated partial thromboplastin time; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; DBP, diastolic blood pressure; GCS, Glasgow Coma Scale; ICU, intensive care unit; INR, international normalized ratio; MBP, mean blood pressure; PCO2, partial pressure of carbon dioxide; PO2, partial pressure of oxygen; PT, prothrombin time; RDW, red-cell distribution width; SAPS II, Simplified Acute Physiology Score II; SBP, systolic blood pressure; SOFA, Sequential Organ Failure Assessment; SpO2, peripheral oxygen saturation; WBC, white blood cell. Among laboratory parameters, patients with severe AKI had significantly higher serum creatinine (2.41 ± 1.90 vs 1.27 ± 0.98 mg/dL, P < 0.0001), BUN (41.19 ± 28.76 vs 27.68 ± 21.00 mg/dL, P < 0.0001), anion gap (16.48 ± 5.40 vs 13.65 ± 3.63 mEq/L, P < 0.0001), and potassium (4.64 ± 0.84 vs 4.39 ± 0.64 mEq/L, P < 0.0001), while bicarbonate was significantly lower (23.54 ± 5.49 vs 25.05 ± 4.99 mEq/L, P < 0.0001). Disease severity scores were significantly elevated in the severe AKI group: SOFA (7.24 ± 4.02 vs 4.39 ± 2.97, P < 0.0001) and SAPS II (46.62 ± 15.05 vs 36.74 ± 11.07, P < 0.0001). A higher proportion of patients in the severe AKI group received mechanical ventilation (68.11% vs 43.09%, P < 0.0001) and vasoactive agents (50.23% vs 29.94%, P < 0.0001). Kaplan–Meier survival analysis revealed that patients with severe AKI had significantly worse survival at both 7 days (HR = 2.31, 95% CI: 1.82–2.92; P < 0.001) and 28 days (HR = 2.22, 95% CI: 1.81–2.72; P < 0.001) compared with those without severe AKI (Figure 1). Figure 1Kaplan–Meier survival curves for ICU patients with COPD stratified by severe AKI status. (A) Seven-day survival. Patients with severe AKI had significantly lower survival probability compared with those without severe AKI (HR = 2.31, 95% CI: 1.82–2.92; log-rank P < 0.001). (B) Twenty-eight-day survival (HR = 2.22, 95% CI: 1.81–2.72; log-rank P < 0.001). Shaded areas represent 95% confidence intervals.Image A: Kaplan Meier graph shows survival probability by acute kidney injury status. Two lines, ′no′ and ′yes′, with step-shaped curves and censor marks. X-axis: Days (0-7), Y-axis: Survival probability (0.00-1.00). Text: p < 0.001, HR = 2.31 (95% CI: 1.82-2.92). Both lines start near 1.00, decline over time; ′yes′ line stays below ′no′ line. Risk table: ′no′ - 0: 3620, 1: 3608, 2: 2147, 3: 1364, 4: 940, 5: 681, 6: 494, 7: 368; ′yes′ - 0: 1085, 1: 1080, 2: 949, 3: 805, 4: 679, 5: 582, 6: 494, 7: 423. Image B: Kaplan Meier graph with shaded confidence bands. X-axis: Days (0-30), Y-axis: Survival probability (0.00-1.00). Text: p < 0.001, HR = 2.22 (95% CI: 1.81-2.72). Dashed lines at survival probability 0.50 and day 20. ′Yes′ line drops faster, reaches 0.50 at day 20; ′no′ line stays above 0.50 through day 30. Risk table: ′no′ - 0: 3620, 5: 681, 10: 188, 15: 86, 20: 41, 25: 19, 30: 0; ′yes′ - 0: 1086, 5: 582, 10: 280, 15: 163, 20: 92, 25: 57, 30: 0.Two line graphs showing Kaplan Meier survival by severe acute kidney injury status over 7 and 30 days.Abbreviations: HR, hazard ratio; CI, confidence interval; AKI, acute kidney injury. Kaplan–Meier survival curves for ICU patients with COPD stratified by severe AKI status. (A) Seven-day survival. Patients with severe AKI had significantly lower survival probability compared with those without severe AKI (HR = 2.31, 95% CI: 1.82–2.92; log-rank P < 0.001). (B) Twenty-eight-day survival (HR = 2.22, 95% CI: 1.81–2.72; log-rank P < 0.001). Shaded areas represent 95% confidence intervals. Feature Selection The dataset was split into a training set (n = 3,294) and testing set (n = 1,411) with stratified sampling. Three feature selection methods were applied to the training set (Figure 2A–G). LASSO regression with 10-fold cross-validation at lambda_1se (0.0073) identified 8 variables. Boruta confirmed 34 features as important after 100 iterations, and random forest importance selected the top 10 variables. The intersection yielded six consensus predictors: total ICU length of stay, creatinine, SAPS II, SOFA score, anion gap, and heart rate. These variables were used for the original retrospective trajectory model. Figure 2Feature selection using three complementary methods. (A) LASSO coefficient path plot. (B) Ten-fold cross-validation error curve for LASSO regression. (C) Random forest out-of-bag error curve. (D) Random forest variable importance ranked by Mean Decrease Gini. (E) Boruta algorithm importance trajectory over 100 iterations. (F) Boruta feature importance boxplot. (G) Venn diagram displaying the intersection of variables selected by LASSO (8), Boruta (34), and random forest (10). Six consensus variables were identified.The image A showing a line graph of coefficient paths. The horizontal axis label is minus log left parenthesis lambda right parenthesis, with ticks at 2, 3, 4, 5, 6. The vertical axis label is Coefficients, with ticks at 0, 5, 10, 15. Multiple lines start near 0 at minus log left parenthesis lambda right parenthesis about 2 and diverge; the highest line rises to about 17 by about 6. The image B showing a cross validation curve. The horizontal axis label is minus log left parenthesis lambda right parenthesis, range about 2 to 8. The vertical axis label is Binomial Deviance, range about 0.80 to 1.10. Points with error bars drop from about 1.08 near 2 to a minimum about 0.82 near 5, then rise slightly to about 0.83 to 0.84 by 8. The image C showing a random forest error curve. The horizontal axis label is trees, range 0 to 500. The vertical axis label is Error, range 0.0 to 0.6. Three curves: one starts about 0.25 and declines to about 0.17 by 50 then stays near 0.17; one starts about 0.15 and declines to about 0.05 by 50 then stays near 0.05; one stays near 0.60 across all trees. The image D showing a dot chart of variable importance. The horizontal axis label is Mean decrease in Gini, range 0 to 120. The vertical axis lists variables including Total ICU length of stay, Creatinine, SAPS II, SOFA score, Blood urea nitrogen, Anion gap, PT, Heart rate, WBC count, RDW, PCO2, Sodium, Potassium, Platelet count, Bicarbonate, Age, Diastolic blood pressure, Glucose, Hemoglobin, pH, Glucose, aPTT, Systolic blood pressure, Mean blood pressure, Total bilirubin, Sodium, Respiratory rate, Temperature, INR, Charlson Comorbidity Index. The largest dot is Total ICU length of stay near 120; Creatinine near 60; SAPS II near 45; SOFA score near 35; remaining dots cluster mostly below about 20. The image E showing Boruta importance trajectories. The horizontal axis label is Classifier run, range 0 to 100. The vertical axis label is Importance, range about negative 10 to 50. Many lines fluctuate; several top lines stay around 35 to 45; many mid lines stay around 0 to 15; a few low lines stay around negative 5 to 0. The image F showing a Boruta feature importance boxplot. The horizontal axis lists many variables from left to right, including Shocked minimum, Shocked maximum, Shocked mean, Mechanical ventilation, Myocardial infarction, CHF, PVD, CVA, Dementia, COPD, Rheumatic disease, Peptic ulcer disease, Mild liver disease, Diabetes, Diabetes with complications, Hemiplegia, Renal disease, Malignancy, Moderate or severe liver disease, Metastatic solid tumor, AIDS, Respiratory rate, Temperature, INR, Recorded date of death, Total bilirubin, Charlson Comorbidity Index, Age, Sodium, Potassium, Bicarbonate, Platelet count, WBC count, RDW, Hemoglobin, Glucose, pH, PCO2, PT, Anion gap, Blood urea nitrogen, Heart rate, SOFA score, SAPS II, Creatinine, Total ICU length of stay. The vertical axis label is Importance, range 0 to 50. Boxplots rise gradually, with the highest at Total ICU length of stay around 45 to 50, followed by Creatinine around 35 to 40, SAPS II around 20 to 25, SOFA score around 15 to 20 and Heart rate around 10 to 15. The image G showing a Venn diagram with circles labeled LASSO, Boruta and RFE. Counts shown: LASSO only 0 with 0.0 percent; Boruta only 22 with 64.7 percent; RFE only 0 with 0.0 percent; LASSO and Boruta overlap 2 with 5.9 percent; Boruta and RFE overlap 4 with 11.8 percent; LASSO and RFE overlap 0 with 0.0 percent; all three overlap 6 with 17.6 percent.A multi plot figure showing LASSO, random forest, Boruta feature selection and a Venn overlap summary. Feature selection using three complementary methods. (A) LASSO coefficient path plot. (B) Ten-fold cross-validation error curve for LASSO regression. (C) Random forest out-of-bag error curve. (D) Random forest variable importance ranked by Mean Decrease Gini. (E) Boruta algorithm importance trajectory over 100 iterations. (F) Boruta feature importance boxplot. (G) Venn diagram displaying the intersection of variables selected by LASSO (8), Boruta (34), and random forest (10). Six consensus variables were identified. Machine Learning Model Development and Comparison Fifteen machine-learning algorithms were trained on the six selected features, with hyperparameters optimized by cross-validation (Table 2). Model performance was evaluated on the independent testing set using AUC, accuracy, sensitivity, specificity, F1 score, PPV, NPV, and Youden’s index (Figure 3A–F and Table 3). CatBoost had the highest observed test-set AUC (0.848, 95% CI, 0.825–0.872), followed by the XGBoost tree booster (0.844, 95% CI, 0.820–0.868) and LightGBM (0.844, 95% CI, 0.820–0.868). Paired DeLong tests using CatBoost as the prespecified reference showed that its advantage was not statistically significant after Holm correction versus the XGBoost tree booster (AUC difference 0.0042; adjusted P = 0.155), LightGBM (difference 0.0045; adjusted P = 0.077), logistic regression (adjusted P = 0.055), or elastic-net regression (adjusted P = 0.055). Complete paired comparisons are shown in Supplementary Figure S1 and Supplementary Table S8.Table 2Optimal Hyperparameters for the 15 Machine-Learning Models Selected by Cross-Validation in the Training SetModelHyperparameterSelected ValueRandom forestmtry2XGBoost tree boosternrounds100XGBoost tree boostermax_depth3XGBoost tree boostereta0.1XGBoost tree boostergamma0XGBoost tree boostercolsample_bytree0.7XGBoost tree boostermin_child_weight1XGBoost tree boostersubsample0.7Radial-basis support vector machinesigma0.1Radial-basis support vector machineC2Logistic regressionparameternoneK-nearest neighborsk9Partial least squares discriminant modelncomp5Gradient boosting machinen.trees50Gradient boosting machineinteraction.depth1Gradient boosting machineshrinkage0.1Gradient boosting machinen.minobsinnode10Single-hidden-layer neural networksize7Single-hidden-layer neural networkdecay0.01Naive BayesfL0Naive BayesusekernelTRUENaive Bayesadjust1.5Linear discriminant analysisparameternoneElastic-net logistic regressionalpha0.5Elastic-net logistic regressionlambda0.001AdaBoost.M1mfinal100AdaBoost.M1maxdepth1AdaBoost.M1coeflearnBreimanXGBoost linear boosternrounds50XGBoost linear boosterlambda1XGBoost linear boosteralpha0XGBoost linear boostereta0.01CatBoostdepth4CatBoostlearning_rate0.05CatBoostl2_leaf_reg1CatBoostiterations146LightGBMnum_leaves15LightGBMlearning_rate0.01LightGBMmin_data_in_leaf10LightGBMlambda_l20LightGBMnrounds207Notes: Hyperparameters were selected using the training data only. Model names correspond to the executable implementations used in the analysis; the XGBoost tree and linear boosters and elastic-net logistic regression are reported separately.Abbreviations: AdaBoost, adaptive boosting; LightGBM, Light Gradient-Boosting Machine; XGBoost, Extreme Gradient Boosting. Table 3Performance of 15 Machine-Learning Models in the Held-Out Testing SetModelAUC (95% CI)AccuracySensitivitySpecificityPPVNPVBrier ScoreCatBoost0.848 (0.825–0.872)0.8280.4340.9470.7090.8480.121XGBoost tree booster0.844 (0.820–0.868)0.8360.4680.9470.7240.8560.122LightGBM0.844 (0.820–0.868)0.8260.3570.9660.7580.8340.125Gradient boosting machine0.839 (0.815–0.863)0.8240.3630.9610.7380.8350.126Single-hidden-layer neural network0.835 (0.810–0.860)0.8240.4550.9340.6730.8510.124AdaBoost.M10.835 (0.810–0.859)0.8150.3480.9550.6980.8300.135Logistic regression0.832 (0.806–0.859)0.8260.3970.9550.7250.8410.126Elastic-net logistic regression0.832 (0.806–0.859)0.8260.3910.9570.7300.8400.126Random forest0.831 (0.806–0.856)0.8260.4310.9440.6970.8470.126Linear discriminant analysis0.828 (0.802–0.855)0.8250.3970.9530.7170.8410.128Naive Bayes0.823 (0.797–0.849)0.8120.5140.9010.6090.8610.141Partial least squares discriminant model0.820 (0.793–0.847)0.8150.3260.9610.7160.8270.165XGBoost linear booster0.810 (0.783–0.836)0.8190.4580.9270.6540.8510.136K-nearest neighbors0.778 (0.748–0.808)0.8110.3540.9480.6690.8310.141Radial-basis support vector machine0.770 (0.736–0.804)0.8260.3630.9650.7560.8350.134Notes: All metrics were calculated in the held-out testing set (n = 1,411). AUC confidence intervals were calculated using DeLong’s method. Accuracy, sensitivity, specificity, PPV, and NPV used a probability threshold of 0.5.Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Figure 3Comparison of predictive performance across 15 machine learning models. (A, B) Performance metrics for training (A) and testing (B) sets. (C and D) Forest plots of AUC with 95% CIs for training (C) and testing (D) sets. (E and F) Decision curve analysis for training (E) and testing (F) sets.The image A showing a multi line chart of model performance metrics for the training set. The x axis labels are Accuracy, F1, NPV, PPV, Sensitivity, Specificity, Youden. The y axis label is Metric value, ranging from 0.0 to 1.0. Multiple model lines cluster near Accuracy about 0.8 to 0.85, dip at F1 about 0.45 to 0.6, rise at NPV about 0.8 to 0.85, vary at PPV about 0.65 to 0.85, dip at Sensitivity about 0.3 to 0.55, peak at Specificity about 0.9 to 1.0 and end at Youden about 0.25 to 0.45. The image B showing a multi line chart of model performance metrics for the testing set. The x axis labels are Accuracy, F1, NPV, PPV, Sensitivity, Specificity, Youden. The y axis label is Metric value, ranging from 0.0 to 1.0. Lines cluster near Accuracy about 0.8 to 0.85, F1 about 0.45 to 0.6, NPV about 0.8 to 0.85, PPV about 0.6 to 0.75, Sensitivity about 0.35 to 0.55, Specificity about 0.9 to 1.0 and Youden about 0.25 to 0.4. The image C showing a forest plot of AUC with 95 percent confidence interval for the training set. The x axis label is AUC (95 percent CI), ranging from 0.750 to 0.900. The y axis lists models with point estimates: LightGBM 0.875, CatBoost 0.860, XGBoost tree booster 0.835, Gradient boosting machine 0.833, AdaBoost. M1 0.830, Neural network 0.828, Random forest 0.825, XGBoost linear booster 0.818, Logistic regression 0.816, Elastic net regression 0.815, Naive Bayes 0.814, Linear discriminant analysis 0.810, Part ial least squares 0.802, Radial basis SVM 0.776, K nearest neighbors 0.766. The image D showing a forest plot of AUC with 95 percent confidence interval for the testing set. The x axis label is AUC (95 percent CI), ranging from 0.725 to 0.900. The y axis lists models with point estimates: CatBoost 0.848, XGBoost tree booster 0.844, LightGBM 0.844, Gradient boosting machine 0.839, Neural network 0.835, AdaBoost. M1 0.835, Elastic net regression 0.832, Logistic regression 0.832, Random forest 0.831, Linear discriminant analysis 0.828, Naive Bayes 0.823, Part ial least squares 0.820, XGBoost linear booster 0.810, K nearest neighbors 0.778, Radial basis SVM 0.770. The image E showing decision curve analysis for the training set. The x axis label is Threshold probability, ranging from 0.0 to 0.9. The y axis label is Net benefit, ranging from negative 0.05 to 0.20. Two reference lines are labeled Treat all and Treat none. Multiple model curves start near 0.20 at threshold 0.0 and decline toward about 0.0 by thresholds about 0.8 to 0.9, with one curve dipping below 0.0 around threshold about 0.25 to 0.35. The image F showing decision curve analysis for the testing set. The x axis label is Threshold probability, ranging from 0.0 to 0.9. The y axis label is Net benefit, ranging from negative 0.05 to 0.20. Two reference lines are labeled Treat all and Treat none. Multiple model curves start near 0.20 at threshold 0.0 and decline toward about 0.0 by thresholds about 0.8 to 0.9, with one curve dipping below 0.0 around threshold about 0.25 to 0.35.A composite of 6 plots comparing 15 machine learning models on performance metrics, AUC and net benefit.Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval. Table 2Optimal Hyperparameters for the 15 Machine-Learning Models Selected by Cross-Validation in the Training SetModelHyperparameterSelected ValueRandom forestmtry2XGBoost tree boosternrounds100XGBoost tree boostermax_depth3XGBoost tree boostereta0.1XGBoost tree boostergamma0XGBoost tree boostercolsample_bytree0.7XGBoost tree boostermin_child_weight1XGBoost tree boostersubsample0.7Radial-basis support vector machinesigma0.1Radial-basis support vector machineC2Logistic regressionparameternoneK-nearest neighborsk9Partial least squares discriminant modelncomp5Gradient boosting machinen.trees50Gradient boosting machineinteraction.depth1Gradient boosting machineshrinkage0.1Gradient boosting machinen.minobsinnode10Single-hidden-layer neural networksize7Single-hidden-layer neural networkdecay0.01Naive BayesfL0Naive BayesusekernelTRUENaive Bayesadjust1.5Linear discriminant analysisparameternoneElastic-net logistic regressionalpha0.5Elastic-net logistic regressionlambda0.001AdaBoost.M1mfinal100AdaBoost.M1maxdepth1AdaBoost.M1coeflearnBreimanXGBoost linear boosternrounds50XGBoost linear boosterlambda1XGBoost linear boosteralpha0XGBoost linear boostereta0.01CatBoostdepth4CatBoostlearning_rate0.05CatBoostl2_leaf_reg1CatBoostiterations146LightGBMnum_leaves15LightGBMlearning_rate0.01LightGBMmin_data_in_leaf10LightGBMlambda_l20LightGBMnrounds207Notes: Hyperparameters were selected using the training data only. Model names correspond to the executable implementations used in the analysis; the XGBoost tree and linear boosters and elastic-net logistic regression are reported separately.Abbreviations: AdaBoost, adaptive boosting; LightGBM, Light Gradient-Boosting Machine; XGBoost, Extreme Gradient Boosting. Optimal Hyperparameters for the 15 Machine-Learning Models Selected by Cross-Validation in the Training Set Notes: Hyperparameters were selected using the training data only. Model names correspond to the executable implementations used in the analysis; the XGBoost tree and linear boosters and elastic-net logistic regression are reported separately. Abbreviations: AdaBoost, adaptive boosting; LightGBM, Light Gradient-Boosting Machine; XGBoost, Extreme Gradient Boosting. Table 3Performance of 15 Machine-Learning Models in the Held-Out Testing SetModelAUC (95% CI)AccuracySensitivitySpecificityPPVNPVBrier ScoreCatBoost0.848 (0.825–0.872)0.8280.4340.9470.7090.8480.121XGBoost tree booster0.844 (0.820–0.868)0.8360.4680.9470.7240.8560.122LightGBM0.844 (0.820–0.868)0.8260.3570.9660.7580.8340.125Gradient boosting machine0.839 (0.815–0.863)0.8240.3630.9610.7380.8350.126Single-hidden-layer neural network0.835 (0.810–0.860)0.8240.4550.9340.6730.8510.124AdaBoost.M10.835 (0.810–0.859)0.8150.3480.9550.6980.8300.135Logistic regression0.832 (0.806–0.859)0.8260.3970.9550.7250.8410.126Elastic-net logistic regression0.832 (0.806–0.859)0.8260.3910.9570.7300.8400.126Random forest0.831 (0.806–0.856)0.8260.4310.9440.6970.8470.126Linear discriminant analysis0.828 (0.802–0.855)0.8250.3970.9530.7170.8410.128Naive Bayes0.823 (0.797–0.849)0.8120.5140.9010.6090.8610.141Partial least squares discriminant model0.820 (0.793–0.847)0.8150.3260.9610.7160.8270.165XGBoost linear booster0.810 (0.783–0.836)0.8190.4580.9270.6540.8510.136K-nearest neighbors0.778 (0.748–0.808)0.8110.3540.9480.6690.8310.141Radial-basis support vector machine0.770 (0.736–0.804)0.8260.3630.9650.7560.8350.134Notes: All metrics were calculated in the held-out testing set (n = 1,411). AUC confidence intervals were calculated using DeLong’s method. Accuracy, sensitivity, specificity, PPV, and NPV used a probability threshold of 0.5.Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Performance of 15 Machine-Learning Models in the Held-Out Testing Set Notes: All metrics were calculated in the held-out testing set (n = 1,411). AUC confidence intervals were calculated using DeLong’s method. Accuracy, sensitivity, specificity, PPV, and NPV used a probability threshold of 0.5. Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Comparison of predictive performance across 15 machine learning models. (A, B) Performance metrics for training (A) and testing (B) sets. (C and D) Forest plots of AUC with 95% CIs for training (C) and testing (D) sets. (E and F) Decision curve analysis for training (E) and testing (F) sets. Model Evaluation and Clinical Utility Residual analysis and classification metrics are shown in Figure 4A–F. The CatBoost model achieved an accuracy of 84.9% in the training set and 82.8% in the testing set. Clinical impact curves and calibration curves showed acceptable agreement between predicted and observed risk across clinically relevant threshold ranges (Figure 5A–D). Because the model included total ICU length of stay, these results should be interpreted as retrospective trajectory discrimination rather than admission-time prediction. Figure 4Residual analysis and classification performance of CatBoost. (A and B) Reverse cumulative distribution of absolute residuals for training (A) and testing (B) sets. (C and D) Boxplots of absolute residuals for training (C) and testing (D) sets. Red dots indicate root mean square of residuals. (E and F) Confusion matrices for training (E; accuracy = 84.9%) and testing (F; accuracy = 82.8%) sets.Image A and B depict line graphs of reverse cumulative distribution of absolute residuals, with descending curves from 80% to 10% as residuals increase. Image C and D show horizontal boxplots of absolute residuals by model, ranging from 0.00 to 1.00, with models like RandomForest, XGBoost and NeuralNet listed vertically. Red dots indicate root mean square of residuals. Image E and F present confusion matrices with axes labeled Predicted and True Label, categories Control and DN. Image E cell values: DN predicted Control 392, DN predicted DN 368, Control predicted Control 2427, Control predicted DN 107. Image F cell values: DN predicted Control 184, DN predicted DN 141, Control predicted Control 1028, Control predicted DN 58.A multi plot figure showing residual curves, residual boxplots and confusion matrices for CatBoost. Figure 5Clinical utility evaluation of CatBoost. (A and B) Clinical impact curves for training (A) and testing (B) sets. (C and D) Calibration curves for training (C) and testing (D) sets. The dashed diagonal line represents perfect calibration.The image contains four graphs labeled A, B, C and D. Graph A is a line graph showing clinical impact for the training set with the x-axis labeled High Risk Threshold from 0 to 1 and the y-axis labeled Number of Cases. The observed cases are shown with a dashed line and predicted cases with a solid line, both decreasing as the threshold increases. Graph B is similar to A but for the testing set, showing a similar trend. Graph C is a calibration plot for the training set with the x-axis labeled Mean Predicted Probability from 0 to 1 and the y-axis labeled Observed Event Rate. A dashed diagonal line represents perfect calibration and data points closely follow this line. Graph D is a calibration plot for the testing set, showing more deviation from the diagonal line compared to C. The graphs illustrate the CatBoost model′s performance, with calibration plots indicating how well predicted probabilities match observed outcomes.Four graphs showing clinical impact and calibration curves for CatBoost model evaluation. Residual analysis and classification performance of CatBoost. (A and B) Reverse cumulative distribution of absolute residuals for training (A) and testing (B) sets. (C and D) Boxplots of absolute residuals for training (C) and testing (D) sets. Red dots indicate root mean square of residuals. (E and F) Confusion matrices for training (E; accuracy = 84.9%) and testing (F; accuracy = 82.8%) sets. Clinical utility evaluation of CatBoost. (A and B) Clinical impact curves for training (A) and testing (B) sets. (C and D) Calibration curves for training (C) and testing (D) sets. The dashed diagonal line represents perfect calibration. Model Interpretability SHAP analysis was performed to interpret CatBoost predictions (Figure 6A–E). The global feature importance ranking by mean absolute SHAP value was total ICU length of stay (0.774), creatinine (0.467), SAPS II (0.219), anion gap (0.095), SOFA score (0.089), and heart rate (0.076). Higher values of total ICU length of stay, creatinine, SAPS II, and anion gap were associated with higher predicted severe-AKI risk. These associations describe model behavior and should not be read as evidence of modifiability. Figure 6SHAP-based interpretability analysis of CatBoost. (A) Global feature importance ranked by mean absolute SHAP values. (B) SHAP beeswarm plot. Color indicates feature value from low (purple) to high (Orange). (C) SHAP dependence plots for each predictor, with color indicating interaction with SAPS II. (D and E) Individual prediction explanation: waterfall plot (D) and force plot (E) for a representative sample.Image A: Bar chart showing mean absolute SHAP values by feature. Values: ICU LOS 0.774, Creatinine 0.467, SAPS II 0.219, Anion gap 0.095, SOFA 0.089, Heart rate 0.076. Image B: Beeswarm plot with SHAP value on x-axis, features on y-axis: ICU LOS, Creatinine, SAPS II, Anion gap, SOFA, Heart rate. Image C: Six scatter plots with feature on x-axis, SHAP value on y-axis. Titles: ICU LOS, Creatinine, SAPS II, Anion gap, SOFA, Heart rate. Includes color scale labeled SAPS II. Image D: Waterfall plot showing prediction contributions: SAPS II < 60 (+0.414), Creatinine > 2.4 (+0.517), ICU LOS > 19.4 (+1.91), Heart rate < 60 (-0.172), SOFA < 5 (-0.077), Anion gap > 14.2 (-0.06). Labels: E f(x) = 1.6, f(x) = 0.892. Image E: Force plot with prediction segments: +0.414, +0.517, +1.91, -0.172, -0.077, -0.06. Feature labels: SAPS II < 60, Creatinine > 2.4, ICU LOS > 19.4, Heart rate < 60, SOFA < 5, Anion gap > 14.2. Endpoints: E f(x) = 1.6, f(x) = 0.892.A multi plot figure showing SHAP feature importance, beeswarm, dependence plots and prediction explanations.Abbreviation: SHAP, SHapley Additive exPlanations. SHAP-based interpretability analysis of CatBoost. (A) Global feature importance ranked by mean absolute SHAP values. (B) SHAP beeswarm plot. Color indicates feature value from low (purple) to high (Orange). (C) SHAP dependence plots for each predictor, with color indicating interaction with SAPS II. (D and E) Individual prediction explanation: waterfall plot (D) and force plot (E) for a representative sample. Causal Inference Analysis The original CausalForestDML analysis is now presented as exploratory adjusted exposure-effect estimation (Table 4, Figure 7A and B). Among grouped interventions, all 95% confidence intervals crossed zero: mechanical ventilation (exposed n = 2,299), vasoactive agents (n = 1,629), recorded diuretic exposure (n = 3,317), any antibiotic (n = 3,868), and dobutamine (n = 84). The dobutamine interval was especially wide, and non-significance should not be interpreted as evidence of renal safety. The antibiotic variable combined agents with different nephrotoxicity profiles and does not support class-specific conclusions.Table 4Exploratory Adjusted Exposure-Effect Estimates for Severe AKI from CausalForestDMLExposureExposure TypeExposed, nUnexposed, nAdjusted Risk Difference95% CIMechanical ventilationGrouped intervention2,2992,4060.0256−0.0549 to 0.1061Vasoactive agentGrouped intervention1,6293,076−0.0019−0.0843 to 0.0806Recorded diuretic exposureGrouped intervention3,3171,388−0.0332−0.1089 to 0.0424Any antibioticGrouped intervention3,868837−0.0082−0.1014 to 0.0850DobutamineGrouped intervention844,6210.0241−0.3188 to 0.3669CreatinineContinuous exposure0.07230.0227 to 0.1219SOFA scoreContinuous exposure0.0071−0.0098 to 0.0241SAPS IIContinuous exposure0.0040−0.0009 to 0.0089Anion gapContinuous exposure0.0028−0.0075 to 0.0131Heart rateContinuous exposure0.0005−0.0013 to 0.0022Total ICU length of stayContinuous exposure0.02000.0079 to 0.0321Notes: Binary exposures compare exposed with unexposed patients; counts are not applicable to continuous exposures. Continuous-exposure estimates are per one-unit increase. Confidence intervals are population mean-effect intervals from CausalForestDML. Because temporal ordering and the no-unmeasured-confounding assumption cannot be guaranteed, these estimates are exploratory and do not establish causal effects.Abbreviations: AKI, acute kidney injury; CI, confidence interval; ICU, intensive care unit; SAPS II, Simplified Acute Physiology Score II; SOFA, Sequential Organ Failure Assessment. Figure 7Exploratory adjusted exposure-effect analysis using CausalForestDML. (A) Grouped intervention estimates with 95% confidence intervals. (B) Continuous-exposure estimates per unit with 95% confidence intervals. Estimates are adjusted risk differences and should not be interpreted as confirmed causal effects because temporal ordering and no-unmeasured-confounding assumptions cannot be guaranteed.Image A shows a forest plot of interventions with the x-axis labeled ′Adjusted risk-difference estimate′ and ticks at -0.2, 0.0, 0.2, 0.4, 0.6. The y-axis lists interventions: Mechanical ventilation (0.0256, interval -0.0549 to 0.1061), Vasoactive agent (-0.0019, interval -0.0843 to 0.0806), Recorded diuretic exposure (-0.0332, interval -0.1089 to 0.0424), Any antibiotic (-0.0082, interval -0.1014 to 0.0850), Dobutamine (0.0241, interval -0.3188 to 0.3669). Image B shows a forest plot of continuous exposures per unit with the x-axis labeled ′Adjusted risk-difference estimate per unit′ and ticks at 0.000, 0.025, 0.050, 0.075, 0.100, 0.125, 0.150. The y-axis lists: Creatinine (0.0723, interval 0.0227 to 0.1219), SOFA score (0.0071, interval -0.0098 to 0.0241), SAPS II (0.0040, interval -0.0009 to 0.0089), Anion gap (0.0028, interval -0.0075 to 0.0131), Heart rate (0.0005, interval -0.0013 to 0.0022), Total ICU length of stay (0.0200, interval 0.0079 to 0.0321).Two forest plots of adjusted risk difference estimates, showing mostly near zero effects and wide intervals.Abbreviations: CI, confidence interval; DML, double machine learning. Table 4Exploratory Adjusted Exposure-Effect Estimates for Severe AKI from CausalForestDMLExposureExposure TypeExposed, nUnexposed, nAdjusted Risk Difference95% CIMechanical ventilationGrouped intervention2,2992,4060.0256−0.0549 to 0.1061Vasoactive agentGrouped intervention1,6293,076−0.0019−0.0843 to 0.0806Recorded diuretic exposureGrouped intervention3,3171,388−0.0332−0.1089 to 0.0424Any antibioticGrouped intervention3,868837−0.0082−0.1014 to 0.0850DobutamineGrouped intervention844,6210.0241−0.3188 to 0.3669CreatinineContinuous exposure0.07230.0227 to 0.1219SOFA scoreContinuous exposure0.0071−0.0098 to 0.0241SAPS IIContinuous exposure0.0040−0.0009 to 0.0089Anion gapContinuous exposure0.0028−0.0075 to 0.0131Heart rateContinuous exposure0.0005−0.0013 to 0.0022Total ICU length of stayContinuous exposure0.02000.0079 to 0.0321Notes: Binary exposures compare exposed with unexposed patients; counts are not applicable to continuous exposures. Continuous-exposure estimates are per one-unit increase. Confidence intervals are population mean-effect intervals from CausalForestDML. Because temporal ordering and the no-unmeasured-confounding assumption cannot be guaranteed, these estimates are exploratory and do not establish causal effects.Abbreviations: AKI, acute kidney injury; CI, confidence interval; ICU, intensive care unit; SAPS II, Simplified Acute Physiology Score II; SOFA, Sequential Organ Failure Assessment. Exploratory Adjusted Exposure-Effect Estimates for Severe AKI from CausalForestDML Notes: Binary exposures compare exposed with unexposed patients; counts are not applicable to continuous exposures. Continuous-exposure estimates are per one-unit increase. Confidence intervals are population mean-effect intervals from CausalForestDML. Because temporal ordering and the no-unmeasured-confounding assumption cannot be guaranteed, these estimates are exploratory and do not establish causal effects. Abbreviations: AKI, acute kidney injury; CI, confidence interval; ICU, intensive care unit; SAPS II, Simplified Acute Physiology Score II; SOFA, Sequential Organ Failure Assessment. Exploratory adjusted exposure-effect analysis using CausalForestDML. (A) Grouped intervention estimates with 95% confidence intervals. (B) Continuous-exposure estimates per unit with 95% confidence intervals. Estimates are adjusted risk differences and should not be interpreted as confirmed causal effects because temporal ordering and no-unmeasured-confounding assumptions cannot be guaranteed. In the original corrected analysis, creatinine had an adjusted risk-difference estimate of 0.0723 (95% CI, 0.0227–0.1219) and total ICU length of stay had an estimate of 0.0200 (95% CI, 0.0079–0.0321). These intervals do not repair temporal overlap. Removing total ICU length of stay from the original CatBoost model reduced the AUC from 0.848 to 0.761 (paired difference 0.0872; Holm-adjusted P < 0.001), and removing creatinine reduced the AUC to 0.833 (paired difference 0.0156; Holm-adjusted P = 0.008; Supplementary Table S7 and Supplementary Figure S2). In the temporally separated landmark analysis, 4,171 patients were eligible and 841 developed incident severe AKI (Supplementary Table S1). The five-feature landmark model had an AUC of 0.679 (95% CI, 0.641–0.717), and removing pre-ICU creatinine yielded an AUC of 0.686 (95% CI, 0.648–0.724; paired difference −0.0067; P = 0.293; Supplementary Table S3, S4 and Supplementary Figure S3). The adjusted pre-ICU creatinine risk difference was 0.0547 per 1 mg/dL (95% CI, −0.0021 to 0.1115) and 0.0705 (95% CI, 0.0022–0.1387) after 1st-99th percentile trimming (Supplementary Table S9). The restricted-adjustment sensitivity estimates are shown in Supplementary Figure S4. In the landmark cohort, 801 patients had J44.1-coded AECOPD, 223 had J44.0-coded COPD with acute lower respiratory infection, and 3,147 had other J44.x-coded COPD; the event-rate difference between J44.1-coded AECOPD and patients without J44.1 was not statistically significant (P = 0.155; Supplementary Tables S5 and S6). Discussion This study establishes a COPD-specific benchmark for severe AKI across 15 machine-learning algorithms and, more importantly, shows how temporal design changes the meaning of model performance. In 4,705 ICU patients with COPD, the six-feature CatBoost model achieved a test-set AUC of 0.848 for retrospective trajectory discrimination. In the strictly ordered 24-hour landmark cohort, discrimination was 0.679 after excluding total ICU length of stay and restricting predictors to pre-landmark information. The difference between these analyses is a central result: models built with variables collected across the ICU course can characterize high-risk trajectories well, whereas admission- or landmark-time prediction requires explicit control of feature availability. This empirical comparison adds a temporal-validity dimension that is often absent from AKI model benchmarks.16–21,25–27 CatBoost had the highest observed test-set AUC among the 15 algorithms. Paired DeLong testing showed that its advantage over the XGBoost tree booster, LightGBM, logistic regression, and elastic-net regression was not statistically significant after Holm correction; this does not diminish the benchmark. Instead, it shows that strong performance arose from a compact consensus feature set and rigorous evaluation rather than from algorithm choice alone. Previous AKI studies also reported strong performance from gradient-boosting methods, usually with different endpoints and larger predictor sets.19–21,25–27 Demonstrating an AUC of 0.848 with six variables, while formally testing model differences, provides a credible and reproducible basis for parsimonious COPD-AKI model development. CatBoost’s ordered boosting and the accompanying SHAP analysis further support transparent interpretation of this compact model.22,41 The COPD-specific importance of these findings extends beyond algorithm ranking. AKI during COPD exacerbation is associated with mortality and adverse renal outcomes, and impaired renal function is common across the COPD spectrum.10,11,13–15 The six-variable signature captures renal function, global illness severity, acid-base disturbance, hemodynamic stress, and the evolving ICU course. The prominence of creatinine and the ICU trajectory accords with general critical-care evidence linking AKI and creatinine changes to mortality, health-care utilization, and persistent renal dysfunction.1,3,42–44 By evaluating this compact signature in a COPD ICU cohort, the study translates general AKI predictors into a population in which hypoxemia, hypercapnia, infection, right-heart dysfunction, mechanical ventilation, and medication exposure converge.9–15 This population-specific evaluation is clinically relevant because renal risk in COPD must be interpreted within the respiratory and hemodynamic context. A second major contribution is the direct demonstration that predictive importance and adjusted exposure-effect estimates are not interchangeable. SHAP explains how a fitted model forms predictions, whereas generalized random forests and double/debiased machine learning estimate exposure-effect relationships under assumptions about temporal order, exchangeability, positivity, and consistency.22–24,28–31 The nonconcordance among SHAP rankings and adjusted estimates is therefore informative rather than a null result. It shows that SOFA, SAPS II, anion gap, and heart rate can support risk stratification without automatically becoming intervention targets. This distinction is essential for translating machine-learning models into clinical research because interventions directed at non-causal severity markers are unlikely to change renal outcomes.23,24 The temporal analyses sharpen, rather than weaken, the study’s contribution. Strictly pre-ICU creatinine did not improve 24-hour landmark discrimination, and its adjusted estimate crossed zero in the full complete-case cohort but excluded zero after trimming. Because creatinine contributes to KDIGO staging and total ICU length of stay can be prolonged by AKI, temporally overlapping definitions can create incorporation bias and reverse causality.3,34,42–44 The landmark analysis provides direct empirical evidence that temporal eligibility can change both predictive performance and causal interpretation. Likewise, the grouped intervention confidence intervals define the current precision of the evidence: they do not establish absence of effect, but they provide population estimates, exposure counts, and uncertainty bounds for better-powered time-varying analyses.23,24,31 The findings should be interpreted in light of several limitations. MIMIC-IV is a single-center resource, no external cohort was available, and pre-ICU laboratory capture depends on records available within the database.32 Urine-output criteria were not used for AKI staging, and potentially important factors such as outpatient medications, contrast exposure, fluid balance, detailed hemodynamic trajectories, infection severity, and treatment timing were unavailable or incompletely measured.34 The intervention variables were necessarily broad, and novel renal biomarkers were not available.45 The renal replacement therapy component of KDIGO stage 3 may also vary with practice-dependent initiation thresholds, adding outcome heterogeneity.46,47 These constraints limit transportability and causal identification, but they do not negate the internal 15-model benchmark, the corrected paired comparisons, or the temporal audit that identifies which performance claims are supportable. Taken together, this study has direct clinical and methodological importance. For clinicians, the six-feature model identifies COPD ICU trajectories associated with severe AKI and supports prioritization of serial renal-function monitoring, medication review, contrast stewardship, hemodynamic assessment, and dose adjustment. The decision-curve and clinical-impact analyses further quantify potential utility across risk thresholds.48 For model developers, the contrast between AUCs of 0.848 and 0.679 demonstrates why time zero, prediction horizon, and feature availability must be specified before a model is labeled clinically deployable.16–18 For causal machine-learning research, the combined SHAP, adjusted-effect, and landmark analyses provide a reproducible framework for separating prognostic importance from intervention relevance.22–24,28–31 Prospective multicenter validation and time-varying modeling can now build on this benchmark rather than starting from an untested algorithmic claim. Conclusion This study provides a COPD-specific benchmark for severe AKI across 15 machine-learning algorithms and shows that a parsimonious six-variable CatBoost model achieves strong retrospective trajectory discrimination. By integrating SHAP interpretation, corrected adjusted-effect estimation, and a strictly ordered 24-hour landmark analysis, the study demonstrates that predictive strength, temporal availability, and causal relevance are distinct properties. This distinction is the central contribution: retrospective models can characterize high-risk ICU trajectories, while clinically deployable prediction requires temporally eligible features. The results provide a rigorous foundation for multicenter validation and the development of time-aware COPD-AKI models.