Predicting Chronic Kidney Disease After Cisplatin Treatment Using Population-Level Data

Key Points

Question

Can we predict chronic kidney disease after cisplatin treatment?

Findings

In this prognostic study including 9521 patients, 1228 of 9010 patients (13.6%) developed chronic kidney disease, and the estimated glomerular filtration rate decreased by a mean of 8.1 mL/min/1.73 m²; a simple spline-based regression model based solely on the pretreatment estimated glomerular filtration rate achieved a similar predictive performance (area under the curve, 0.80) as complex machine learning models incorporating many clinical features. The model was validated in an external cohort.

Meaning

These findings suggest that cisplatin treatment is associated with a predictable decrease in the estimated glomerular filtration rate and that patients with lower baseline kidney function are most likely to develop chronic kidney disease.

Abstract

Importance

Cisplatin is a widely used treatment for cancer that can permanently damage the kidneys. Treatment modifications and other strategies may prevent chronic kidney disease (CKD) in patients at risk; however, the incidence and predictability of CKD following cisplatin treatment remain poorly understood.

Objective

To characterize the incidence of CKD after cisplatin treatment and evaluate prediction models.

Design, Setting, and Participants

In this population-based prognostic study, prediction models were developed based on a retrospective cohort study of patients who received cisplatin chemotherapy for nonhematologic cancer in an outpatient setting between July 1, 2014, and June 30, 2017. Models were tested on a temporal-test cohort of patients from Ontario, Canada, who started treatment between July 1, 2017, and June 30, 2020, and an external-test cohort of patients from a single center in the United States. Data were analyzed from May 1, 2021 to May 7, 2025.

Exposures

Predictive features included demographics, cancer diagnosis, cisplatin dose and schedule, comorbidities, laboratory testing, and patient-reported symptoms.

Main Outcomes and Measures

The outcomes were CKD (estimated glomerular filtration rate [eGFR] <60 mL/min/1.73 m²) and the eGFR after cisplatin treatment. Measures included the area under the receiver operating characteristic curve and the mean absolute error (MAE).

Results

The population-level cohort included 9521 patients (median age, 63 years [IQR, 56-70 years]; 4841 men [50.8%]). Among the 9010 patients without pretreatment CKD, 1228 (13.6%) developed CKD, 81 (0.9%) developed grade 4 or worse CKD, and 16 (0.18%) required dialysis. The eGFR decreased by a mean of 8.1 mL/min/1.73 m² (95% CI, 7.8-8.4 mL/min/1.73 m²). A simple spline-based regression model based solely on the pretreatment eGFR predicted posttreatment CKD in the temporal-test cohort (area under the curve, 0.80 [95% CI, 0.78-0.82]) and the external-test cohort (area under the curve, 0.73 [95% CI, 0.66-0.78]). Similarly, the posttreatment eGFR was predicted by a spline regression based solely on the pretreatment eGFR (temporal-test MAE, 12.6 mL/min/1.73 m² [95% CI, 12.3-13.0 mL/min/1.73 m²]; external-test MAE, 14.3 mL/min/1.73 m² [95% CI, 13.2-15.5 mL/min/1.73 m²]). Complex machine learning systems incorporating all features failed to improve predictions over the univariable models.

Conclusions and Relevance

This study found that cisplatin treatment was followed by a predictable decrease in the eGFR, placing patients with a lower baseline eGFR at the highest risk of CKD. A simple model based on the pretreatment eGFR predicts CKD risk and could guide clinical decision-making.

This prognostic study characterizes the incidence of chronic kidney disease after cisplatin treatment and evaluates prediction models.

Introduction

Cisplatin has long been a cornerstone of cancer treatment, used across cancers of the lung, biliary tract, stomach, anus, bladder, cervix, testes, and more.¹ Although cisplatin has extended countless lives, some patients are left with irreversible kidney damage through tubular cell apoptosis, inflammation, and ischemic injury.² Personalized interventions—such as avoiding nephrotoxins and hypotension, hydration, specialist consultations, and close monitoring—can help protect patients at risk.³ Furthermore, alternatives to cisplatin are often available for patients at risk; however, these strategies may be associated with inferior outcomes.

Despite their clinical importance, the incidence and predictors of chronic kidney disease (CKD) after cisplatin treatment remain poorly understood. Studies on the incidence of CKD after cisplatin treatment to date have focused on pediatric populations or single-center experiences.^4,5,6,7 Although several tools have demonstrated encouraging results for predicting acute kidney injury (AKI) during cisplatin treatment,^8,9,10,11 none were evaluated on a population level or predicted CKD. Importantly, although cisplatin-related AKI may often be transient, CKD outcomes may have greater implications for cancer survivorship and subsequent cardiorenal risk. Furthermore, whether machine learning, which demonstrates an impressive predictive performance for AKI in general,¹² is applicable to patients treated with cisplatin remains untested.

In this study, we analyzed population-level administrative data to determine the incidence of CKD after cisplatin treatment. We then trained simple regressions and complex machine learning models to predict CKD. Finally, we validated our models in a temporally distinct test population-based cohort and an external-test cohort at a large academic center.

Methods

This study adhered to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) + AI statement¹³ and reporting guideline and was approved by the Sunnybrook Research Ethics Board and the Memorial Sloan Kettering Cancer Center’s Institutional Review Board. Our code is publicly available at GitHub, Inc.¹⁴

Population-Level Cohort

We extracted data from population-level administrative health datasets on residents of Ontario. We analyzed data at ICES, a research institute with special legal status that permits using these data without consent to evaluate and improve the health system. Access to these data is available by applying to ICES.¹⁵ ICES administrative datasets capture all health activities covered by the universal Ontario Health Insurance Plan.¹⁶

From the Activity Level Reporting dataset,¹⁷ we identified adults with a valid Ontario Health Insurance Plan number who started cisplatin treatment between July 1, 2014, and June 30, 2020. We excluded treatments for hematologic cancers (International Classification of Diseases for Oncology, 3rd Edition¹⁸ topography codes C42 and C77 or morphology codes including and above 959) and treatments administered in clinical trials.

We divided the cohort by the date of treatment initiation (a temporal split) to mirror prospective model deployment, providing a more robust validity assessment than a random split.¹⁹ Patients who received their first treatment after July 1, 2017, formed the temporal-test cohort, which we used only for model evaluation and sensitivity analysis; this cohort was divided approximately in half (eFigure 1 in Supplement 1). Earlier patients formed the development cohort, which we randomly split further into training and tuning cohorts (80%:20% ratio).

External-Test Cohort

We evaluated the generalizability of our models in a previously described external-test cohort.⁵ In brief, this cohort comprised patients aged 18 years or older who received cisplatin treatment at Memorial Sloan Kettering Cancer Center between January 1, 2000, and September 21, 2011, and survived at least 5 years from treatment. We required an outcome measurement, resulting in a smaller cohort than originally reported.

Outcome Measures

The primary outcome was CKD after cisplatin treatment, which was defined following the Kidney Disease: Improving Global Outcomes²⁰ and Kidney Disease Outcomes Quality Initiative guidelines.²¹ Specifically, CKD was identified when the estimated glomerular filtration rate (eGFR) was less than 60 mL/min/1.73 m², calculated using the 2021 Chronic Kidney Disease Epidemiology Collaboration creatinine equation, which does not require information on race and ethnicity and is robust across different races and ethnicities.²² We used the earliest creatinine measurement at least 90 days after completion of cisplatin chemotherapy. Sensitivity analysis considered the mean of 2 creatinine measurements. Secondary outcomes included CKD stages²⁰ G3b (eGFR <45 mL/min/1.73 m²), G4 (<30 mL/min/1.73 m²), and G5 (<15 mL/min/1.73 m² or dialysis). Dialysis initiation data were extracted from physician billing records based on whether dialysis fee codes were billed 90 days to 6 months after completion of therapy.²³ Information on other markers of kidney damage, such as histology and imaging, was unavailable. Other secondary end points included the posttreatment eGFR and the change in eGFR before and after treatment.

To isolate the association of cisplatin treatment with CKD, we conducted a matched case-control analysis. In brief, we applied coarsened exact matching, in which each patient was matched 1:1 to another individual receiving cancer treatment during the same period in Ontario, Canada, with a regimen that did not include cisplatin based on age, sex, baseline eGFR, and time to posttreatment creatinine measurement.

To understand trajectories of kidney function preceding CKD, we also considered AKI, defined per Kidney Disease: Improving Global Outcomes²⁰ and the Acute Kidney Injury Network²⁴ as an increase in peak creatinine level greater than or equal to 0.3 mg/dL (to convert to micromoles per liter, multiply by 88.4) or at least 1.5 times the baseline measurement. AKI was further categorized into stage 2 (≥2.0 times baseline) and stage 3 (≥3.0 times baseline or an absolute value of 4.0 mg/dL [≥353.6 µmol/L] or initiation of kidney replacement therapy).

Statistical Analysis

Predictive Features

Features extracted from administrative data were selected based on expert and literature review and spanned demographics, cancer diagnosis and treatment details, patient-reported symptoms²⁵ and functional status,²⁶ comorbidities (specifically hypertension and diabetes obtained from billing codes²⁷), acute care utilization history, and laboratory values (eTable 1 in Supplement 1). Feature preprocessing is described in the eAppendix in Supplement 1.

Model Development and Evaluation

We trained models using features measured before the first cisplatin treatment to predict the primary and secondary outcomes. As a simple baseline, we estimated logistic regressions for CKD and a linear regression for the eGFR, applying a restricted cubic spline of 5 degrees with 2 knots to the pretreatment eGFR, hereafter called univariable regression. We also trained complex machine learning systems and an ensemble system that combined estimates from all systems. Additional information on the machine learning systems is available in the eAppendix in Supplement 1. Data were analyzed from May 1, 2021 to May 7, 2025. We used the Wald test for computing odds ratios and the Welch t test for comparing decreases in the eGFR. P values are 2-sided. We used P < .05 to indicate statistical significance.

We evaluated the model performance for CKD in both the temporal- and external-test cohorts using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). The model performance for the posttreatment eGFR was evaluated using the mean absolute error and R². The mean absolute error is calculated as the absolute value of the difference between the predicted and true posttreatment eGFR, with smaller values indicating more accurate predictions. Only a binned version of the univariable regression, which we made publicly available in our code base, was evaluated in the external-test cohort to adhere to privacy restrictions because access to machine learning models are not permitted outside of ICES. We generated 95% CIs and conducted model comparisons using 10 000 bootstrap samples. Calibration was assessed using calibration plots and the integrated calibration index and the Harrell E_max index, which is the maximum absolute difference between a smooth calibration curve and the diagonal line of perfect calibration.²⁸ To determine the clinical utility of our predictive models, we performed a decision-curve analysis.²⁹ We assessed the importance of the features in the machine learning models by measuring the association between these features and performance when randomly permuting each feature in the validation cohort. We also conducted a multivariate logistic regression without penalization, including key covariates, to identify factors associated with CKD. To identify whether any features were important in the context of specific cancer types and regimens beyond the baseline eGFR, we retrained specific models for the most common cancer type and regimen combinations and recalculated the importance of the features. Finally, we evaluated model performance within subgroups. Subgroup analysis by race and ethnicity was not performed because data on race and ethnicity were not collected in the temporal-test cohort from Ontario, Canada, and because the sample sizes of the subgroups in the external-test cohort were small.

Results

Cohort Characteristics

After applying eligibility criteria (eFigure 2 in Supplement 1), the population-level cohort included 9521 patients (median age, 63 years [IQR, 56-70 years]; 4841 men [50.8%]). The development cohort included 4914 patients (Table). The median age was 63 years (IQR, 55-70 years), 2513 (51.1%) were male, and 221 (4.5%) had baseline CKD. The most prevalent cancer type was bronchus and lung cancer (1773 patients [36.1%]). Weekly cisplatin treatment with radiotherapy was the most common treatment regimen (851 patients [17.3%]). The first cisplatin treatment was reduced by at least 10% of the ideal dose for 1750 patients (21.5%) without pretreatment CKD and 275 patients (60.0%) with pretreatment CKD (eFigure 3 in Supplement 1). Missing values ranged from 0.2% to 59.6% across features. The temporal-test cohort included 4607 patients with similar characteristics, although temporal shifts in cisplatin treatment were observed (eg, a lower frequency of stomach cancer). The external-test cohort included 443 patients (eTable 2 in Supplement 1) with a comparable baseline rate of CKD (34 patients [7.7%]). Race and ethnicity data were unavailable in the population-based cohorts, whereas 32 patients (7.2%) in the external-test cohort self-identified as Asian, 27 (6.1%) as Black, and 370 (83.5%) as White.

Table. Characteristics of the Patients in the Population-Level Development and Temporal-Test Cohorts^a.

Characteristic at baseline	Development, July 1, 2014 to June 30, 2017 (n = 4914)	Temporal test, July 1, 2017 to June 30, 2020 (n = 4607)
Age, median (IQR), y	63 (55-70)	64 (56-70)
Sex
Female	2401 (48.9)	2279 (49.5)
Male	2513 (51.1)	2328 (50.5)
Recently immigrated	582 (11.8)	534 (11.6)
Rural residence	662 (13.5)	712 (15.5)
Chronic kidney disease	221 (4.5)	290 (6.3)
Diabetes	809 (16.5)	723 (15.7)
Hypertension	2342 (47.7)	2213 (48.0)
Cancer topography (top 3)
Bronchus and lung	1773 (36.1)	1462 (31.7)
Cervix	353 (7.2)	403 (8.7)
Stomach	415 (8.4)	180 (3.9)
Treatment regimen (top 3)
Weekly cisplatin with radiotherapy	851 (17.3)	985 (21.4)
Cisplatin and etoposide over 3 d	847 (17.2)	581 (12.6)
Cisplatin every 3 wk with radiotherapy	560 (11.4)	537 (11.7)

^a Data are presented as the No. (%) of patients unless indicated otherwise.

Incidence of CKD

eFigures 4 and 5 in Supplement 1 illustrate the transitions from pretreatment CKD grades, AKI stages during treatment, and posttreatment CKD grades. In the overall cohort of 9010 patients, 1228 (13.6%) without pretreatment CKD developed CKD after cisplatin treatment, with 380 (4.2%) progressing to grade 3b or worse and 81 (0.9%) progressing to grade 4 or worse. Dialysis was required for 16 patients (0.18%). Compared with matched controls receiving other cancer treatments (eTable 3 in Supplement 1), cisplatin treatment was associated with an odds ratio of 2.46 (95% CI, 2.24-2.70; P < .001) for CKD and 2.63 (95% CI, 2.25-3.09; P < .001) for stage 3b or worse CKD. The mean decrease in the eGFR after cisplatin treatment was 8.1 mL/min/1.73 m² (95% CI, 7.8-8.4 mL/min/1.73 m²) compared with 1.1 mL/min/1.72 m² (95% CI, 0.8-1.4 mL/min/1.73 m²) in controls (P < .001). AKI affected 2360 (26.2%) patients, of whom 548 (23.2%) subsequently developed CKD. Similar patterns were observed in the development and temporal-test cohorts separately and in the external-test cohort (eTable 4 in Supplement 1). Redefining CKD based on 2 measurements yielded consistent results. Posttreatment CKD and a change in the eGFR were not associated with how long after completing cisplatin treatment the creatinine level was measured for the first time (eTable 5 and eFigure 6 in Supplement 1), suggesting that cisplatin treatment was not associated with further deterioration beyond 3 months after treatment.

Prediction Model Development for CKD

Univariable regression mapped the pretreatment eGFR to posttreatment CKD risk and severity (Figure, A), showing a steep increase in CKD risk with a lower baseline eGFR. In the tuning cohort, the AUROC was 0.81 (95% CI, 0.76-0.85) for CKD prediction, 0.78 (95% CI, 0.69-0.87) for grade 3b or worse CKD, and 0.84 (95% CI, 0.71-0.94) for grade 4 or worse CKD (eTable 6 in Supplement 1). The best-performing machine learning model, an ensemble of all models, achieved AUROCs of 0.86 (95% CI, 0.83-0.89) for CKD, 0.89 (95% CI, 0.84-0.93) for grade 3b or worse CKD, and 0.91 (95% CI, 0.83-0.98) for grade 4 or worse CKD.

Figure. Univariable Regression Model.

A, Mapping of pretreatment estimated glomerular filtration rate (eGFR) to the probability of posttreatment chronic kidney disease (CKD) and grade 3b or worse and grade 4 CKD. B, Receiver operating characteristic (ROC) curve in the temporal-test set. C, ROC curve in the external-test set. AUROC indicates area under the receiver operating characteristic curve.

Model Evaluation in the Temporal-Test Cohort

In the temporal-test cohort, the AUROC of the univariable regression was 0.80 (95% CI, 0.78-0.82) for CKD, 0.78 (95% CI, 0.75-0.82) for grade 3b or worse CKD, and 0.72 (95% CI, 0.63-0.80) for grade 4 or worse CKD (Figure, B). The AUPRC was 0.55 (95% CI, 0.51-0.59) for CKD, 0.33 (95% CI, 0.27-0.40) for grade 3b or worse CKD, and 0.10 (95% CI, 0.05-0.20) for grade 4 or worse CKD (eFigure 7A in Supplement 1). The predictions were well calibrated (eFigure 7B in Supplement 1), and the model would provide a net benefit over a wide range of threshold probabilities (eFigure 8 in Supplement 1). Performance was consistent across subgroups, with variation associated with differences in the rate of posttreatment CKD (eFigure 9 in Supplement 1). Results were unchanged if CKD was defined using 2 measurements (eFigure 10 in Supplement 1).

More complex ensemble machine learning models provided no meaningful gains over univariable regression, with AUROCs of 0.82 (95% CI, 0.80-0.84) for CKD, 0.80 (95% CI, 0.77-0.84) for grade 3b or worse CKD, and 0.76 (95% CI, 0.70-0.83) for grade 4 or worse CKD (eFigure 11 and eTable 7 in Supplement 1). AUPRCs for ensemble models were worse than the univariable regression. Permutation-based analysis identified the pretreatment eGFR as the most important feature (eFigure 12 in Supplement 1). Similarly, the pretreatment eGFR was the most important feature in models trained and tested within specific cancer type and treatment regimen combinations (eTable 8 in Supplement 1). A logistic regression selecting key covariates confirmed that the baseline eGFR has the strongest association with posttreatment CKD (eTable 9 in Supplement 1).

Model Evaluation in the External-Test Cohort

In the external-test cohort, AUROCs for the univariable regression were 0.73 (95% CI, 0.66-0.78) for CKD, 0.87 (95% CI, 0.80-0.92) for grade 3b or worse CKD, and 0.81 (95% CI, 0.61-1.00) for grade 4 or worse CKD (Figure, C). AUPRCs were 0.50 (95% CI, 0.39-0.61) for CKD, 0.37 (95% CI, 0.22-0.57) for grade 3b or worse CKD, and 0.27 (95% CI, 0.01-1.00) for grade 4 or worse CKD (eFigure 13A in Supplement 1). Within the limitations of the smaller sample size, the predictions were generally calibrated, although the model tended to underestimate higher probabilities (eFigure 13B in Supplement 1).

Models for Posttreatment eGFR

Because the definition and grades of CKD are based on thresholds, we also predicted the posttreatment eGFR as a continuous value. In the development cohort, the posttreatment eGFR was associated with the pretreatment eGFR (eFigure 14A in Supplement 1). Univariable regression mapped the pretreatment eGFR to the posttreatment eGFR (eFigure 14B in Supplement 1), indicating that for patients without pretreatment CKD (ie, eGFR <60 mL/min/1.73 m²), the eGFR decreased by 8.6 mL/min/1.73 m². In contrast, patients with pretreatment CKD were not predicted to have a worsening posttreatment eGFR. Predictions in the temporal-test cohort showed an MAE of 12.6 mL/min/1.73 m² and an R² of 0.422 (eFigure 14C in Supplement 1). Machine learning models incorporating a broad range of features only marginally improved performance, with the ensemble having an MAE of 12.2 mL/min/1.73 m² and an R² of 0.460 (eTable 10 in Supplement 1). In the external-test cohort, the univariable regressions had an MAE of 14.3 and an R² of 0.283 (eFigure 14D and eTable 10 in Supplement 1).

Discussion

This prognostic study used population-level data to define the incidence of CKD after cisplatin treatment and developed prediction models for CKD and the posttreatment eGFR. Overall, 13.6% of patients developed CKD after cisplatin treatment, with a mean eGFR decrease of 8.1 mL/min/1.73 m². Most patients who experienced AKI did not develop CKD. A simple model based solely on the pretreatment eGFR predicted CKD and the posttreatment eGFR. Notably, complex machine learning models incorporating a wide range of clinical features provided minimal improvements in predictive performance. This simple model is included in eTable 11 in Supplement 1, which clinicians can use to inform the consent process and guide decision-making around cisplatin treatment and preventative interventions.

These data provide, to our knowledge, the first population-based estimates for CKD after cisplatin treatment among adults. Our findings for CKD align with our single-institution study of kidney function after cisplatin treatment,⁵ where fewer than 3% of patients progressed to stage G5 CKD and none required dialysis. Importantly, given that many patients with AKI may not develop CKD, our data on CKD prediction may provide additional clinical utility over existing AKI prediction models. This is particularly relevant with respect to cancer survivorship, as CKD has known implications for cardiovascular outcomes and progressive kidney dysfunction.^30,31

The predictive models generated by this study could personalize and improve the care of patients receiving cisplatin. First, informed consent requires disclosing risks that might affect treatment decisions,³² including CKD after cisplatin treatment. Second, personalized estimates of CKD can guide preventive strategies,³ particularly those that may compromise treatment effectiveness, such as substituting alternative treatments for cisplatin. Finally, these models could select patients for future research into preventing CKD after cisplatin treatment.

Despite a large sample size and extensive availability of predictive features, machine learning models did not outperform the simpler model based on the pretreatment eGFR alone. Analysis of the importance of the features of the machine learning models confirmed that most of the predictive ability came from the pretreatment eGFR, with other clinical features playing a small role. Together, these results suggest that novel predictors beyond clinical characteristics such as genetic variation¹⁰ will be required to improve the prediction of CKD after cisplatin treatment, using the baseline eGFR as a benchmark for comparison.

Broadly, these results demonstrate that patients typically experience a modest decrease in the eGFR after cisplatin treatment (ie, mean, 8.1 mL/min/1.73 m²). Consequently, patients with a lower pretreatment eGFR are most likely to develop CKD.

Limitations

This study has limitations. First, several predictors that may improve model performance were unavailable in the dataset, such as concomitant medications, which are not covered under universal insurance in Ontario. Similarly, we did not have details on radiotherapy fields, which has been shown in previous studies to affect kidney function, although to a lesser extent than cisplatin treatment.³³ Second, our definition of CKD was based on the eGFR for pragmatic reasons. Future work could consider a broader perspective on CKD, including albuminuria and electrolyte abnormalities.²¹ Furthermore, measurement error in the estimation of the eGFR may partially account for prediction errors, although results were robust to considering single or multiple measurements. Third, the results were validated in a previously published separate single-center cohort,⁵ in which selection criteria included only individuals who survived at least 5 years from treatment, which could introduce survivorship bias. Despite these differences, the validation provides additional robustness to our core findings. Although the system was temporally and externally validated, any predictive model should be evaluated prospectively before deployment.

Conclusions

In conclusion, this prognostic study provided population-based estimates of CKD after cisplatin treatment and a prediction model based solely on the pretreatment eGFR. These results can be used in clinical practice to inform consent and guide prevention strategies.

Supplement 1.

eAppendix. Supplementary Text

eFigure 1. Monthly Distribution of the First Treatment of Cisplatin Received by Patients in the Study

eFigure 2. CONSORT Diagram

eFigure 3. Dose Reduction From Ideal Dosing for Cisplatin According to Baseline eGFR

eFigure 4. Alluvial Plot Displaying the Trajectories of Pre-Treatment Chronic Kidney Disease (CKD), Acute Kidney Injury (AKI) During Cisplatin Treatment, and Post-Treatment CKD in the Development Cohort

eFigure 5. Alluvial Plot Displaying the Trajectories of Pre-Treatment Chronic Kidney Disease (CKD), Acute Kidney Injury (AKI) During Cisplatin Treatment, and Post-Treatment CKD in the Temporal Test Cohort

eFigure 6. The Proportion of Patients With Post-Treatment Chronic Kidney Disease (CKD) (a-c) and the Distribution of the Post-Treatment Estimated Glomerular Filtration Rate (eGFR) (d-f), According to the Days Between the Last Cisplatin Treatment and the Earliest Creatinine Measurement, in the Development (a, d), Temporal Test (b, e), and External Test (c, f) Cohorts

eFigure 7. The Performance of the Ensemble Machine Learning Model Predicting Chronic Kidney Disease (CKD) in the Temporal Test Set

eFigure 8. Decision Curve Analysis of the Spline-Based Regression Models Based on Pre-Treatment Estimated Glomerular Filtration Rate for a) CKD, b) Grade 3b or Worse CKD, and c) Grade 4 or Worse CKD

eFigure 9. The Univariable Regression Model Performance Predicting Chronic Kidney Disease (CKD) Within Subgroup Populations in the Temporal Test Set

eFigure 10. Performances Across Different Models and Targets in the Temporal Test Cohort When CKD Is Defined Using Two Measurements

eFigure 11. Performances Across Different Models and Targets in the Temporal Test Cohort

eFigure 12. Feature Importance for Predicting Chronic Kidney Disease, as Measured Using the Permutation Importance Score, Which Is the Change in the Area Under The Receiver Operating Characteristic Curve (AUROC) After Randomly Permuting Features Across Patients and Time in the Test Cohort

eFigure 13. The Performance of the Univariable Regression Model Predicting Chronic Kidney Disease (CKD) in the External Test Set

eFigure 14. Predicting Estimated Glomerular Filtration Rate (eGFR) After Treatment With Cisplatin From Pre-Treatment eGFR

eTable 1. Complete List of Features Used in the Machine Learning Models, With the Mean (Standard Deviation) and Missing Proportion Provided Based on the Development Cohort

eTable 2. Characteristics of the Patients in the External Test Cohort (N=443)

eTable 3. Characteristics of the Patients Who Received Systemic Therapies for Cancer Besides Cisplatin Selected Using Coarsened Exact Matching 1:1. 242 (2.5%) of Cisplatin-Treated Patients Could Not Be Matched and Were Excluded From This Analysis

eTable 4. Incidence of Chronic Kidney Disease (CKD) After Cisplatin Treatment in the Temporal and External Test Sets Among Patients Without CKD at Baseline

eTable 5. Days From the Last Cisplatin Treatment Until the Earliest Creatinine Measurement for Determining Post-Treatment Chronic Kidney Disease and Estimated Glomerular Filtration Rate

eTable 6. Performance of the Baseline Spline-Based Regression on Estimated Glomerular Filtration Rate and the Machine Learning Models in the Tuning Cohort

eTable 7. Performance of the Baseline Spline-Based Regression on Estimated Glomerular Filtration Rate and the Machine Learning Models in the Temporal Test Cohort

eTable 8. Area Under the Receiver Operating Characteristic Curve (AUROC) for Predicting Chronic Kidney Disease (CKD) When Model Training and Evaluation Is Restricted to Specific Combinations of Cancer Type and Treatment Regimen (Using Cancer Care Ontario Regimen Names)

eTable 9. Logistic Regression in the Development Cohort Selecting Key Covariates

eTable 10. Performance of the Baseline Spline-Based Regression to Predict Post-Treatment Estimated Glomerular Filtration Rate

eTable 11. Nomogram for Mapping Pre-Treatment eGFR to the Predicted Probability of Post-Treatment CKD, Based on the Univariable Regression Model

Supplement 2.

Data Sharing Statement