1. Introduction
The relationship between cancer and kidney disease has become increasingly important as advances in oncology continue to improve survival and expand treatment options. Modern cancer care relies on therapies that have transformed outcomes across many malignancies but frequently expose patients to substantial nephrotoxic risk. Platinum-based chemotherapy, targeted therapies, immune checkpoint inhibitors (ICIs), cellular therapies, contrast-enhanced imaging, and increasingly complex multimodal treatment strategies expose patients to diverse and often overlapping nephrotoxic risks throughout the cancer care continuum [ 1 , 2 , 3 ]. Preserving kidney function is therefore central not only for reducing morbidity but also for maintaining access to effective anticancer treatment [ 4 , 5 ].
Acute kidney injury (AKI) is particularly consequential because even modest deterioration in kidney function may lead to dose reductions, treatment delay, regimen modification, or discontinuation of anticancer therapy [ 4 , 5 , 6 ]. Treatment delays are associated with increased cancer mortality [ 7 ], while AKI may also lead to recurrent injury, chronic kidney disease (CKD), prolonged hospitalization, and poorer long-term outcomes [ 8 , 9 , 10 ].
The burden of AKI varies substantially by malignancy, treatment, and clinical setting. In a Danish cohort of 37,267 patients with cancer, the one-year cumulative incidence of AKI was 17.5%, with particularly high risk in multiple myeloma, liver cancer, and kidney cancer [ 11 ]. Approximately one in ten patients receiving systemic anticancer therapy developed AKI in a population-based Canadian study [ 12 ]. Among 482,016 older US adults with newly diagnosed cancer, the five-year cumulative incidence of AKI was 19.8%, with marked variation across cancer types [ 13 ]. Contemporary Danish data further demonstrate treatment-specific heterogeneity, ranging from less than 1% after breast cancer surgery to more than 50% after radical nephrectomy or allogeneic hematopoietic stem cell transplantation (HSCT), with AKI consistently associated with increased mortality [ 14 ].
Risk is even greater among hospitalized and critically ill patients, where sepsis, tumor lysis syndrome, hemodynamic instability, intensive chemotherapy, and nephrotoxic supportive care frequently coexist [ 9 , 15 ]. Many patients also enter treatment with reduced renal reserve. The IRMA study established that impaired kidney function is common at treatment initiation [ 16 ], and more recent oncology cohorts confirm a substantial prevalence of CKD across malignancies [ 17 ].
Cancer-associated AKI is often multifactorial, and may arise from the malignancy itself, urinary tract obstruction, volume depletion, infection, contrast exposure, direct treatment toxicity, or immune-mediated injury [ 1 , 2 , 3 ]. These mechanisms can occur simultaneously or sequentially as patients move between ambulatory therapy, imaging, surgery, and hospitalization. This complexity has contributed to the growth of onco-nephrology as a subspecialty focused on preserving kidney health throughout the cancer journey [ 18 ].
Current practice nevertheless remains largely reactive, typically occurring after a rise in serum creatinine or decline in urine output. For research and standardized classification, AKI is most commonly defined and staged according to Kidney Disease: Improving Global Outcomes (KDIGO) criteria [ 19 ]. Although KDIGO provides the accepted framework for AKI definition and severity classification, explicit staging is not consistently applied in routine clinical practice, and clinically significant AKI may remain underrecognized. By the time functional changes become apparent, meaningful kidney injury may already be established [ 20 ]. Advances in AI, ML, and electronic health record (EHR) analytics have created opportunities to identify elevated risk hours to days earlier in general hospital populations [ 21 , 22 ]. The March 2026 KDIGO AKI/AKD public-review draft emphasizes the potential role of externally validated clinical risk models when linked to kidney-protective care pathways [ 23 ].
The key question is therefore not simply whether AKI can be predicted, but whether prediction systems can capture the dynamic nature of kidney risk in cancer and translate earlier recognition into actions that preserve kidney function and treatment continuity. This review synthesizes conventional and AI/ML-based AKI prediction evidence across major oncology settings, places these studies within the broader AKI prediction literature, and identifies priorities for validation, implementation, and precision onco-nephrology.
2. Literature Search and Review Approach
This state-of-the-art narrative review was informed by iterative anchor searches of PubMed and Google Scholar conducted through 23 July 2026, supplemented by backward and forward citation searching and targeted searches across major oncology settings. Search concepts combined “acute kidney injury” or “AKI” with “cancer” or “oncology” and terms related to prediction, clinical risk scores, artificial intelligence, and machine learning. Additional targeted searches addressed cisplatin, contrast-enhanced imaging, immune checkpoint inhibitor, hematopoietic stem cell transplantation, CAR T-cell therapy, targeted anticancer therapy, and oncologic surgery. Searches were restricted to English-language publications, with no restriction on publication date. We prioritized peer-reviewed studies that developed or evaluated conventional or AI/ML-based AKI prediction models in cancer populations. Where dedicated models were unavailable, we considered cancer-specific observational studies reporting clearly defined AKI outcomes and clinically relevant risk factors. Studies were prioritized by author consensus according to clinical relevance, methodological contribution, population and treatment specificity, and contribution beyond evidence already represented. Complete search strategies are provided in Supplementary Tables S1 and S2 . Consistent with the narrative-review design, no formal protocol, exhaustive duplicate screening, risk-of-bias assessment, or meta-analysis was undertaken. Evidence maturity was categorized descriptively as “emerging” when evidence consisted primarily of observational studies or initial model development without independent validation; “promising” when a dedicated prediction model had undergone internal, temporal, or limited external validation; and “mature” when evidence included independent external validation, prospective evaluation, and progress toward clinical implementation. These categories are descriptive author assessments and do not constitute a formal evidence-grading system ( Supplementary Methods S1 ).
3. Why AKI Prediction Matters in Oncology
The rationale for oncology-specific AKI prediction is the possibility that some episodes may be prevented or mitigated when increased risk is recognized before clinically overt injury. Earlier identification may support individualized kidney health assessment, nephrotoxin stewardship, optimization of hydration and drug dosing, intensified monitoring, and timely nephrology involvement. These measures are most useful when initiated before injury is established, but prediction alone is unlikely to improve outcomes unless it is linked to a clearly defined clinical response pathway [ 24 ].
The clinical importance of prediction also varies according to severity and trajectory of AKI. Predicting a first episode of creatinine-defined AKI addresses a different clinical question from predicting recurrent AKI, severe AKI, kidney replacement therapy, persistent kidney dysfunction, or AKI that disrupts cancer treatment. Recurrent AKI may reflect cumulative treatment exposure, incomplete renal recovery, persistent vulnerability, or repeated clinical insults and may therefore require different predictors and intervention strategies from a first AKI event. Severe or recurrent AKI may be particularly important targets for risk stratification and may justify different thresholds for preventive intervention. AKI is associated with mortality, CKD, prolonged hospitalization, and greater healthcare utilization [ 24 , 25 , 26 ]; in oncology, it can also determine eligibility for nephrotoxic or renally cleared therapies. Cisplatin is a clear example, where even modest kidney dysfunction may trigger dose reduction, delay, or discontinuation and reduce cumulative treatment intensity [ 6 , 27 ]. Similar challenges occur with ICIs, targeted agents, cellular therapies, and multimodal regimens, although the relevant injury mechanisms and preventive options differ by treatment.
Risk stratification could therefore help direct limited resources toward patients most likely to benefit while avoiding unnecessary surveillance among lower-risk individuals. The most useful systems will likely update risk over time as patients accumulate treatment exposures, undergo imaging or surgery, develop infection, and receive supportive medications [ 22 , 28 , 29 ]. The aim is not to replace clinical judgement, but to provide timely and actionable information within precision cancer care [ 30 ].
3.1. Unique Features of AKI in Patients with Cancer
AKI in cancer is not simply a more frequent form of hospital-associated AKI. It is a heterogeneous syndrome shaped by cancer biology, treatment toxicity, immune dysregulation, recurrent infection, supportive care, structural complications, and pre-existing renal vulnerability. Treatment-related intrinsic kidney injury may occur alongside hemodynamic or infectious insults, while malignancy itself or cancer-directed procedures may cause postrenal AKI through urinary tract obstruction. These phenotypes have different risk factors, trajectories, and opportunities for prevention or intervention and should not necessarily be pooled within a single prediction framework.
Multiple insults often occur concurrently or sequentially. A patient receiving cisplatin may also have vomiting or diarrhea, receive nephrotoxic antibiotics, undergo contrast-enhanced imaging, or develop obstruction related to tumor burden or urinary tract involvement. Such combinations may produce overlapping prerenal, intrinsic, and postrenal mechanisms. Volume depletion is particularly common yet difficult to assess, and conventional indices such as fractional excretion of sodium have limited diagnostic performance in the presence of CKD or diuretic use [ 31 , 32 ].
The malignancy itself may directly impair kidney function through urinary tract obstruction, hypercalcemia, paraprotein-associated disease, tumor lysis syndrome, thrombotic microangiopathy, or paraneoplastic glomerular disorders [ 2 , 3 , 33 ]. Hematologic malignancies add further complexity through intensive chemotherapy, severe infection, HSCT, and disease-specific processes such as cast nephropathy [ 34 , 35 ].
Baseline kidney function is also frequently uncertain. Older age, hypertension, diabetes, previous AKI, and CKD reduce renal reserve, while cancer-related sarcopenia and cachexia can make serum creatinine misleadingly low. Recent guidance supports CKD-EPI-based estimates, creatinine-cystatin C equations when creatinine is unreliable and cystatin C when available and measured glomerular filtration rate in selected high-risk or borderline cases [ 4 , 36 ]. In practice, cystatin C and measured GFR remain inconsistently available, so some patients begin therapy with incompletely characterized kidney vulnerability.
The defining challenge for prediction is the dynamic relationship between kidney risk and cancer treatment. Successive chemotherapy cycles, immunotherapy, targeted therapies, recurrent imaging, surgery, supportive medications, infections, and hospitalizations all contribute to a continuously evolving risk profile. Combination and sequential regimens may generate additive, cumulative, or interacting effects that cannot be assigned to a single exposure. Population-based evidence showing wide variation across malignancies, treatments, and clinical settings supports viewing cancer-associated AKI as a spectrum of overlapping etiologic and treatment-related phenotypes rather than a single disease process [ 14 ].
Most existing models nevertheless estimate risk at one baseline time point, before one treatment course, procedure, or imaging encounter. Such static estimates cannot capture cumulative dose, regimen sequencing, changing kidney function, or emerging complications. Effective prediction will require longitudinal treatment histories, laboratory trajectories, concurrent nephrotoxins, and evolving clinical status. Figure 1 summarizes the principal patient-, disease-, and treatment-related drivers of AKI risk and highlights why dynamic risk trajectories are central to precision onco-nephrology.
3.2. Evolution of AKI Prediction and Lessons for Oncology
Clinical AKI prediction has evolved from regression-based scores using a small number of predefined variables to ML approaches capable of analyzing high-dimensional and longitudinal EHR data. Early models offered transparency and ease of implementation but were derived in single-center cohorts and lacked external validation. A systematic review by Hodgson et al. identified substantial heterogeneity in predictor selection, baseline creatinine definitions, AKI outcomes, and validation strategies [ 37 ].
The expansion of structured EHR data enabled random forests, gradient-boosted trees, support vector machines, and neural networks to model nonlinear relationships and interactions across laboratory, medication, physiologic, and demographic variables. Wang et al. showed that ensemble learning could improve discrimination in imbalanced EHR datasets [ 38 ]. However, systematic reviews have consistently found that algorithmic sophistication does not compensate for weak design, heterogeneous outcomes, or inadequate validation [ 39 , 40 ].
Longitudinal modeling was a major conceptual advance. Tomašev et al. trained recurrent neural networks on more than 700,000 adults and predicted AKI up to 48 h before clinical recognition, demonstrating that temporal trajectories contain information unavailable from isolated measurements [ 22 ]. Churpek et al. subsequently showed that a simpler gradient-boosted model could achieve strong performance and external validity across nearly 500,000 admissions from three health systems [ 21 ]. These studies illustrate that scalable implementation may depend as much on parsimonious design and transportability as on model complexity.
More recent reviews have similarly concluded that complex ML methods do not consistently outperform carefully developed regression or ensemble models when trained on comparable data [ 39 , 41 ]. The principal determinants of clinical usefulness are high-quality data, appropriate prediction horizons, standardized outcome labels, calibration, external validation, interpretability, and integration into workflows. These lessons are directly relevant to oncology, where general inpatient variables such as vasopressor use may be informative during critical illness but not capture cumulative chemotherapy, cancer stage, treatment sequencing, or ambulatory nephrotoxic exposures.
General AKI models therefore provide a methodological roadmap rather than a ready-made solution for cancer care. Oncology-specific systems must incorporate treatment context and changing risk while preserving calibration and clinical actionability. Figure 2 summarizes the transition from static clinical scores to longitudinal and explainable prediction, alongside representative milestones in oncology-specific research.
4. AKI Prediction Evidence Across Oncology Settings
Table 1 summarizes the populations, outcome and baseline creatinine definitions, prediction time points and horizons, predictor availability, modeling approaches, validation designs, and reported performance of the included prediction studies. To assess clinical actionability, predictors were classified according to whether they were available before treatment or surgery, at treatment initiation or hospital admission, during treatment, or only after the principal opportunity for preventive intervention. Where available, we report calibration, AUPRC, sensitivity, specificity, positive predictive value, confidence intervals, and decision-curve or net-benefit analyses in addition to AUROC. Measures not provided by the original study are identified as “not reported.”
4.1. Hospitalized Cancer Patients
Oncology-specific prediction initially developed in hospitalized and routinely monitored cancer populations, where repeated laboratory testing provides dense data, but clinical trajectories remain irregular. Park et al. created a framework capable of incorporating heterogeneous EHR observations collected at variable intervals, explicitly addressing missingness and changing patient status [ 42 ]. Two studies by Li et al. used Bayesian networks to model AKI in hospitalized patients with hematologic malignancies and gastrointestinal cancers, illustrating how interpretable probabilistic methods can represent dependencies among clinical risk factors [ 48 , 49 ].
Scanlon et al. developed one of the largest oncology-specific models using data from more than 48,000 patients and nearly 600,000 laboratory observations. A random forest model identified patients at increased risk before clinical recognition using routinely available data. The study prospectively evaluated model performance but did not prospectively implement the model or test a prediction-triggered intervention [ 29 ]. A subsequent study externally applied this model to a general, non-cancer critical-care population using the MIMIC-IV database ( n = 28,498), reporting an AUROC of 0.821 (vs. 0.881 in the original Christie test set) and identifying 67.3% of AKI events at a “medium or higher” risk threshold; discrimination was somewhat lower among patients with a cancer diagnosis (AUROC 0.801) than without (0.836) [ 62 ]. Nevertheless, although this evaluation was conducted in a non-oncology population and does not establish oncology-specific transportability, it remains one of the few AKI models originally derived in cancer patients to undergo external, cross-setting validation.
More recently, Guo et al. developed an explainable XGBoost model to predict AKI within 72 h of ICU admission among critically ill patients with digestive-system tumors. The model was developed using MIMIC-IV data and externally validated in three independent US and Chinese cohorts, with external AUROCs ranging from 0.616 to 0.769. Calibration was acceptable in the internal test cohort but poor across all three external cohorts, indicating a likely need for local recalibration. Decision-curve analysis showed internal net benefit across a range of thresholds, but this benefit was attenuated in the external cohorts and was largely limited to lower thresholds. The model relied primarily on general critical-illness variables, including SOFA score, sepsis, mechanical ventilation, and vasoactive-drug use, rather than cancer stage or longitudinal treatment exposures, and it has not undergone prospective clinical evaluation [ 50 ]. This study represents progress in external validation while demonstrating that transportability cannot be inferred from discrimination alone.
Collectively, hospitalized-cancer studies demonstrate the feasibility of dynamic AKI risk estimation from routine clinical data. Their next step is prospective workflow evaluation and deeper integration of cancer characteristics, treatment history, cumulative nephrotoxic exposure, and post-discharge or ambulatory data.
4.2. Cisplatin-Associated AKI
Cisplatin has received particularly substantial attention because nephrotoxicity is a principal dose-limiting toxicity and a defined treatment cycle creates a clear prediction window. Approximately 20–30% of patients receiving cisplatin-based chemotherapy develop AKI, depending on the population and definition used [ 6 , 27 ]. Common predictors include older age, pre-existing CKD, lower baseline estimated glomerular filtration rate (eGFR), higher cisplatin dose, hypoalbuminemia, hypertension, concurrent nephrotoxins, and prior kidney injury.
Motwani et al. developed and validated the first dedicated multivariable clinical score for AKI after the first cisplatin course using 4481 patients at two cancer centers. The score used four routinely available variables, including age, cisplatin dose, hypertension, and serum albumin, and achieved moderate discrimination across cancer types [ 43 ]. Subsequent work showed that the eGFR equation used at baseline can alter risk stratification, reinforcing the importance of standardized kidney-function assessment in oncology prediction models [ 63 ].
Gupta et al. later derived and externally validated a parsimonious clinical score for severe AKI following the first intravenous cisplatin administration across geographically diverse academic cancer centers. By targeting a twofold creatinine increase or kidney replacement therapy, the model focused on a less frequent but more clinically consequential outcome and represents the strongest multicenter validation among current cisplatin scores [ 44 ].
ML studies have expanded the range of variables and functional relationships considered. Okawa et al. combined artificial neural-network and gradient-boosted decision-tree methods using routinely available information from patients receiving their first cisplatin course. The model was developed in 1014 patients and evaluated in a temporally separated same-center cohort of 226 patients. Its strongest performance was observed among patients aged 65 years or older, with an AUROC of 0.78; serum albumin, body-surface area, and maximum daily cisplatin dose were among the most influential predictors. The outcome was based on Common Terminology Criteria for Adverse Events-based serum creatinine outcomes rather than KDIGO AKI, and the model lacked independent external validation [ 51 ].
Ambe et al. subsequently developed an interpretable CatBoost model using 29 clinical and laboratory measurements and cisplatin exposure variables in 1253 patients receiving their first cisplatin course, of whom 119 developed AKI. Cisplatin-associated AKI was defined using the serum creatinine component of the KDIGO criteria within 14 days after the final cisplatin administration in the first treatment course. The model achieved an AUROC of 0.78, and SHAP analysis was used to characterize the contribution of clinical and treatment-related variables to individual predictions. Nevertheless, the model did not incorporate urine-output criteria and lacked independent external validation [ 52 ].
These studies address complementary rather than directly competing objectives. Motwani and Gupta prioritize parsimonious clinical scores and stronger multicenter validation, whereas Okawa and Ambe illustrate the ability of ML to incorporate broader treatment and medication data and model nonlinear relationships. Their AUROCs should not be compared directly, however, because definitions, outcome severity, prediction windows, populations, and validation designs differ. Current evidence does not establish that ML outperforms well-developed conventional scores. Most models also remain static around the first cisplatin exposure and do not account for repeated cycles, cumulative dose, changing kidney function, or evolving nephrotoxic burden. Prospective studies are needed to determine whether risk-guided care reduces AKI while maintaining cancer-treatment intensity.
4.3. Contrast-Associated AKI
Contrast-enhanced computed tomography (CT) is central to cancer diagnosis, staging, response assessment, and surveillance. Patients may undergo serial contrast exposures while receiving nephrotoxic therapies or experiencing volume depletion. The term contrast-associated AKI (CA-AKI) is important because AKI after imaging is not necessarily caused by contrast, particularly when competing causes are common.
Risk tools have historically been derived in non-oncology populations, most notably the Mehran score for percutaneous coronary intervention which incorporates variables including baseline kidney function, diabetes, heart failure, and contrast volume [ 64 ]. Their applicability to CT in oncology is uncertain because cancer populations differ in treatment exposure, nutritional status, supportive medications, and imaging frequency.
Gupta et al. developed the most comprehensive oncology-specific score using 46,593 contrast-enhanced CT examinations across 25,184 patients at three academic centers. Because patients could contribute multiple CT examinations, generalized estimating equations accounted for within-patient correlation. A point score based on cancer type, kidney function, albumin, proteinuria, comorbidities, medications, and contrast volume achieved a C-statistic of 0.74 in both the development and validation samples and observed CA-AKI risk ranged from 2.2% in the lowest risk group to 32.7% in the highest-risk group [ 53 ].
This setting has a practical implementation advantage because risk can be estimated before imaging and linked to medication review, volume assessment, contrast-volume minimization, and post-imaging monitoring. However, the model was retrospectively developed and internally validated within the same health system. Although repeated examinations were handled statistically, the score remained encounter-based and did not model the number, timing, or cumulative dose of prior contrast exposures. Independent external validation and prospective evaluation of score-guided preventive pathways are therefore needed.
4.4. Hematopoietic Stem Cell Transplantation
HSCT is among the highest-risk oncology settings for AKI, with contemporary incidences approaching or exceeding 50% [ 14 ]. Earlier studies reported AKI in approximately 40–70% of allogeneic HSCT recipients, depending on conditioning, transplant characteristics, and the definition used [ 65 ]. Kidney injury may result from multiple overlapping mechanisms, including conditioning toxicity, sepsis, hypovolemia, sinusoidal obstruction syndrome, graft-versus-host disease, calcineurin-inhibitor toxicity, thrombotic microangiopathy, infection, and nephrotoxic antimicrobials [ 66 ].
Dedicated prediction evidence remains limited. Most studies have focused on identifying individual risk factors rather than developing and validating tools capable of supporting individualized care [ 65 ]. One notable exception is Rodrigues et al., who developed a conventional pre-transplant score incorporating pre-existing CKD, a hematopoietic cell transplantation comorbidity index of at least 2, a diagnosis of leukemia or lymphoma, and the platelet-to-lymphocyte ratio at admission [ 54 ]. The score was developed and evaluated in the same single-center retrospective cohort and has not undergone external or prospective validation. A static pre-transplantation score also cannot capture the rapidly changing post-transplantation risk.
Future models will likely require longitudinal approaches that incorporate evolving laboratory values, immunosuppressive therapy, nephrotoxic antimicrobial exposure, infectious complications, graft-versus-host disease, thrombotic microangiopathy, and other transplant-specific events. Such dynamic models could continuously update risk, facilitate earlier nephrology involvement, and support more proactive kidney-protective strategies in one of the highest-risk populations within onco-nephrology.
4.5. Chimeric Antigen Receptor (CAR) T-Cell Therapy-Associated AKI
CAR T-cell therapy has transformed the treatment of relapsed or refractory hematologic malignancies but is associated with a pattern of kidney injury driven largely by systemic inflammation rather than direct nephrotoxicity. Reported AKI incidence ranges from approximately 5% to 34%, and severe cytokine release syndrome (CRS), elevated baseline lactate dehydrogenase, and reduced kidney function are recurrent risk factors [ 67 , 68 ]. Because risk follows dynamic inflammatory changes, prediction may benefit from serial biomarkers, hemodynamics, and CRS grading. However, dedicated AI-based AKI models after CAR T-cell therapy remain unavailable, representing a clear research gap as cellular therapies expand across hematologic malignancies and solid tumors.
4.6. Immune Checkpoint Inhibitor-Associated AKI
ICI-associated AKI is an important but difficult-to-define complication. Reported incidence is approximately 2–5% during monotherapy and may be higher with combination regimens, but attribution varies because kidney biopsy is uncommon and competing causes are frequent [ 69 , 70 ]. Meraz-Muñoz et al. illustrated this distinction: all-cause AKI occurred in 16.5% of 309 patients receiving ICI, whereas only 12 cases were attributed to ICI-related nephrotoxicity by biopsy or nephrologist assessment [ 71 ].
Several studies have evaluated individualized AKI prediction. Yu et al. compared ML algorithms using random-forest recursive feature elimination [ 55 ]. Sakuragi et al. developed a LightGBM model that continuously predicted all-cause AKI within seven days and used longitudinal SHAP patterns to identify clinically interpretable clusters related to hypovolemia, drug-associated injury, and cancer cachexia [ 45 ]. More recently, Liu et al. compared nine algorithms and evaluated a SHAP-interpretable gradient-boosting model using a single-center training/test split among patients receiving PD-1/PD-L1 inhibitors [ 56 ]. Lou et al. additionally developed and externally validated a multicenter nomogram for severe AKI during anti-PD-1/PD-L1 therapy [ 57 ].
These studies demonstrate the feasibility of individualized risk estimation and explainable modeling during ICI therapy. However, models often predict all-cause AKI rather than immune-mediated nephritis, and outcome adjudication differs substantially. Future work should distinguish ICI-attributable injury from competing causes and test models prospectively across institutions and treatment combinations.
4.7. Targeted Therapy-Associated AKI
Despite widespread use of targeted anticancer therapies, dedicated validated AI/ML models for targeted therapy-associated AKI remain scarce. Evidence is predominantly mechanistic, pharmacovigilance-based, or observational, reflecting heterogeneous renal phenotypes across drug classes [ 24 ]. VEGF-pathway inhibitors may cause hypertension, proteinuria, and thrombotic microangiopathy; EGFR-directed therapies may cause electrolyte wasting, particularly hypomagnesemia and hypocalcemia [ 24 ]; and BRAF or MEK inhibitors have been associated with tubular injury and interstitial nephritis. Several agents, including CDK4/6, PARP, ALK, and MET inhibitors, can inhibit tubular creatinine secretion and produce pseudo-AKI without a true reduction in GFR [ 72 ].
In a real-world study of osimertinib, Lin et al. identified older age and pre-existing renal impairment as risk factors for renal injury but used conventional multivariable regression and defined injury by eGFR below 60 mL/min/1.73 m 2 and/or proteinuria rather than KDIGO AKI [ 58 ]. The diversity of mechanisms, treatment durations, combination regimens, and pseudo-AKI makes a single pooled prediction framework unlikely to be sufficient. Future studies should use drug-specific phenotypes, longitudinal treatment data, and kidney-function measures capable of distinguishing true injury from altered creatinine handling.
4.8. Surgery-Associated AKI in Oncology
Major oncologic surgery is an important but procedure-specific setting for AKI prediction. Nephrectomy, cystectomy, hepatectomy, cytoreductive surgery with hyperthermic intraperitoneal chemotherapy (HIPEC), and major colorectal surgery expose patients to nephron loss, ischemia–reperfusion injury, blood loss, hemodynamic instability, dehydration, nephrotoxins, and postoperative complications [ 73 ]. Population-based data show marked heterogeneity, from less than 1% following breast cancer surgery to more than 50% after radical nephrectomy [ 14 ].
General perioperative scores were developed mainly in cardiac or nononcologic surgical populations [ 46 , 47 ] and do not capture systemic therapy, cancer characteristics, or nephron loss. Oncology-specific studies are emerging across several procedures. Lei et al. compared ML algorithms for AKI after liver cancer resection [ 59 ]. Lee et al. evaluated 4104 patients undergoing unilateral nephrectomy for renal cell carcinoma, where LightGBM achieved a test-set AUROC of 0.810, compared with 0.775 for an internally developed logistic-regression score and 0.626 for the SPARK index. The improvement over the internal logistic model was modest and the new models lacked external validation [ 60 ].
He et al. recently developed an interpretable XGBoost model in 1916 patients undergoing mesocolic excision across five centers. Hospitals were separated into development and independent external-validation cohorts, and the external AUROC was 0.922; SHAP highlighted tumor, operative, inflammatory, blood-loss, and transfusion-related predictors [ 61 ]. Although the He et al. model demonstrated useful discrimination, it incorporated postoperative variables that would not be available at the time of preoperative risk assessment. It may therefore support postoperative surveillance or early recognition, but its performance should not be interpreted as evidence of utility for purely preoperative prevention.
Surgical oncology offers a defined opportunity for prediction because risk can be estimated around a planned procedure. Accurate preoperative risk stratification could support individualized hydration protocols, optimization of nephrotoxic medications, early nephrology consultation, perioperative hemodynamic management, and enhanced postoperative surveillance for patients at greatest risk. Future studies should distinguish preoperative from intraoperative or postoperative prediction, validate models across health systems, and determine whether adding longitudinal cancer-treatment history improves performance and clinical utility for this population.
Figure 3 synthesizes the maturity, methods, validation status, and principal gaps across these oncology settings. The figure intentionally distinguishes dedicated prediction models from observational or risk-factor evidence so that settings without validated AI/ML tools are not overstated.
5. Methodological Challenges and Clinical Translation
Despite promising discrimination in several settings, oncology-specific AKI prediction models have not yet been widely implemented in routine care or prospectively shown to improve kidney outcomes, cancer-treatment continuity, or survival. Most remain retrospective, single-center or health-system-specific, and few have been evaluated as part of clinical intervention. Clinical AI research more broadly shows that accurate prediction is insufficient when models are not transparently reported, externally validated, monitored, and integrated into decision-making [ 74 , 75 , 76 , 77 , 78 ].
5.1. Outcome Definitions and Baseline Kidney Function
Studies differ in their use of serum creatinine and urine-output criteria, staging thresholds, outcome windows, inpatient versus outpatient events, and attribution to a specific treatment. Baseline kidney function may be defined using the most recent outpatient creatinine, admission creatinine, a nadir value, or a back-calculated estimate. These choices can substantially alter AKI incidence within the same populations [ 79 ]. In supervised learning, inconsistent outcome labels introduce systematic noise and reduce comparability. Treatment-specific research must also distinguish all-cause AKI from causal phenotypes such as ICI-mediated nephritis, true contrast-associated injury, and pseudo-AKI from altered creatinine secretion.
5.2. Longitudinal Data, Missingness, and Treatment Context
Cancer care is distributed across hospitals, ambulatory infusion units, imaging centers, and laboratories, so EHRs may not capture complete treatment history, nephrotoxic medications, or laboratory trajectories. Monitoring is irregular and clinically driven; patterns of missingness may themselves contain information. Models that treat missingness as random may therefore introduce bias. Recurrent architectures can encode measurement intervals and missingness patterns [ 80 ], while attention-based methods can represent irregularly sampled time series [ 81 ]. Distributed multicenter approaches may also support model development without centralizing patient-level data [ 82 ]. These methods remain underused in oncology-specific AKI prediction.
A further challenge is the substantial real-world heterogeneity in pre-treatment health, cancer characteristics, treatment regimens, and care settings, which limits direct comparison across prediction studies. In addition to using narrower inclusion criteria or prespecified clinical strata, computational phenotyping may help identify more homogeneous patient groups. For example, community-detection algorithms could be applied to patient-similarity networks constructed from longitudinal clinical and treatment data to identify groups with shared risk trajectories. When combined with temporal models of evolving AKI risk, this approach could identify subgroups defined by factors such as baseline kidney reserve, treatment regimen, cumulative nephrotoxic exposure, prior AKI, or care setting. Prediction algorithms could then be evaluated within more comparable subgroups rather than across highly heterogeneous oncology populations. This approach remains generally exploratory in oncology-specific AKI prediction and requires validation, but it may provide a framework for determining when direct model comparison, recalibration, or pooling is appropriate.
5.3. Generalizability, Calibration, Explainability, and Fairness
External validation should assess more than just AUROC, and work to guide comparison of studies. Discrimination does not indicate whether predicted probabilities are accurate, whether performance is acceptable at clinically relevant thresholds, or whether use of the model would improve decision-making. In this way, Table 1 summarizes calibration measures, AUPRC, sensitivity, specificity, positive predictive value, confidence intervals, and decision-curve or net-benefit results wherever these were reported. Calibration may deteriorate substantially when event prevalence, treatment practice, laboratory monitoring, or patient mix differs across institutions [ 83 ]. Models should therefore report discrimination, calibration, sensitivity, positive predictive value, and net benefit in the population and workflow for which they are intended [ 83 , 84 ]. Interpretability may improve clinical trust, particularly when models identify patient-specific contributors to risk, but post hoc explanations do not establish causality. Fairness evaluation is similarly uncommon. Performance should be examined across age, sex, race or ethnicity where legally and ethically appropriate, socioeconomic context, cancer type, treatment regimen, and baseline kidney function to prevent models from amplifying existing disparities.
5.4. Implementation and Clinical Response Pathways
The value of prediction depends on what happens after a high-risk result. Alerts that are poorly timed, nonspecific, or disconnected from clinical responsibility may increase workload without improving care. The March 2026 KDIGO AKI/AKD public-review draft emphasizes external validation and standardized response pathways rather than isolated electronic alerts [ 23 ]. In oncology, actions may include medication review, hydration or hemodynamic optimization, dose assessment, contrast minimization, intensified laboratory monitoring, biomarker testing, or early nephrology consultation. The appropriate response will differ by treatment and prediction horizon.
Prospective implementation studies should therefore evaluate a complete sociotechnical intervention including data availability, model performance, alert timing, clinical acceptance, response fidelity, unintended consequences, and patient-centered outcomes. Success should be judged by prevention or mitigation of AKI, preserved kidney function, maintained cancer-treatment intensity, reduced hospitalization, and equitable benefit, not by discrimination alone. Figure 4 compares methodological features of representative oncology-specific AKI prediction studies and selected general-AKI benchmark models and illustrates persistent gaps in prospective evaluation, fairness, and real-world workflow integration.
6. Future Directions: Toward Precision Onco-Nephrology
Precision onco-nephrology aims to move beyond treating established kidney injury toward continuously adapting kidney-protective care as patients undergo successive treatments and accumulate risk. AI and ML may support this transition by integrating laboratory trajectories, medication exposures, imaging, treatment history, comorbidities, and evolving clinical status into updated risk estimates. Because cancer care is inherently iterative, with repeated chemotherapy cycles, immunotherapy, surgery, imaging, and supportive treatments, AI and ML-based predictive models offer an opportunity to move beyond static baseline risk assessment toward prediction that evolves alongside the patient’s clinical care. The goal is not an isolated probability, but a system that identifies why risk is changing and connects that information to an evidence-based response.
Four priorities should guide future work. First, AKI definitions, baseline kidney function ascertainment, treatment attribution, and prediction horizons should be standardized. Second, prediction should become both phenotype- and treatment-aware. Models should distinguish, where feasible, treatment-related intrinsic kidney injury from hemodynamic, infectious, structural/postrenal, and other cancer-associated mechanisms rather than treating all oncology-associated AKI as a single outcome. Longitudinal treatment data should include cumulative dose, combination and sequential regimens, nephrotoxic supportive care, procedures, evolving cancer-related complications, and recovery after previous AKI. Models should also consider whether the clinically relevant target is any AKI, recurrent AKI, severe AKI, kidney replacement therapy, or kidney injury that alters cancer treatment. Biomarkers such as cystatin C or tubular injury markers may add value where they are measured reliably but should be evaluated for incremental benefit over routine data.
Third, multicenter validation should include calibration, clinical utility, subgroup performance, and temporal updating as oncology practice changes. Explainability and fairness should be designed into evaluation rather than added after model development. Fourth, prospective trials should test whether model-guided pathways improve outcomes compared with usual care and should measure treatment continuity, patient burden, resource use, and unintended effects alongside kidney outcomes.
Figure 5 illustrates an interactive framework in which baseline and cycle-specific data inform risk stratification, prevention, monitoring, AKI detection, and adaptation of cancer treatment. Clinical judgement remains central; prediction should augment multidisciplinary decisions by oncology, nephrology, pharmacy, radiology, surgery, nursing, and data-science teams. A learning health-system approach could then recalibrate models, monitor safety, and refine response pathways as new data accumulate.
7. Conclusions
AKI remains a common and clinically important complication throughout the cancer care continuum. Current evidence demonstrates the feasibility of individualized prediction in hospitalized cancer populations and in selected treatment settings, particularly cisplatin exposure, contrast-enhanced CT, ICI therapy, and procedure-specific oncologic surgery. Prediction evidence remains limited in HSCT, while evidence for CAR T-cell therapy and targeted therapies remains predominantly observational. Across settings, direct comparison is limited by heterogeneous outcomes, prediction windows, populations, and validation strategies.
Most oncology-specific models remain retrospective, and few have undergone independent external validation, rigorous calibration assessment, fairness evaluation, or prospective implementation. No model has yet shown in an interventional study that prediction-guided care improves kidney outcomes, preserves cancer-treatment continuity, or improves survival. Progress will depend on standardized definitions, high-quality longitudinal oncology data, treatment-aware and interpretable models, multicenter validation, and integration with clearly assigned clinical response pathways.
The value of AI-enabled precision onco-nephrology should ultimately be judged by outcomes that matter to patients and clinicians: preserved kidney function, safer delivery of effective cancer therapy, fewer avoidable treatment disruptions, and equitable care. The vision of precision onco-nephrology, in which kidney risk is continuously assessed and addressed throughout the cancer journey, is achievable, but will become a clinical reality only through rigorous, multidisciplinary, and clinically grounded research.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Supplementary Materials
The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/cancers18182937/s1 , Supplementary Methods S1: Literature Identification and Evidence Prioritization; Table S1: Principal PubMed Anchor Search; Table S2: Targeted PubMed Search Concepts by Oncology Setting.
Author Contributions
Conceptualization, B.F.S., A.U. and C.I.B.; methodology, B.F.S., A.H., A.U. and C.I.B.; investigation and literature review, B.F.S., A.H., C.F., A.V.d.P., B.R., A.U. and C.I.B.; writing—original draft preparation, B.F.S.; writing—review and editing, A.H., C.F., A.V.d.P., B.R., A.U. and C.I.B.; visualization, B.F.S. and A.U.; supervision, A.U. and C.I.B.; project administration, B.F.S. and C.I.B. All authors have read and agreed to the published version of the manuscript.
Institutional Review Board Statement
Not applicable.
Informed Consent Statement
Not applicable.
Data Availability Statement
No new data were created or analyzed in this study. Data sharing is not applicable to this article.
Conflicts of Interest
The authors declare no conflicts of interest.
References
1. Gork I. Xiong F. Kitchlu A. Cancer Drugs and Acute Kidney Injury: New Therapies and New Challenges Curr. Opin. Nephrol. Hypertens. 2024 33 474 485 10.1097/MNH.0000000000001001 38712677
2. Liu S. Zhao J. Wang F. Acute Kidney Injury in Cancer Patients Clin. Exp. Nephrol. 2022 26 103 112 10.1007/s10157-021-02131-7 34499266
3. Rosner M.H. Perazella M.A. Acute Kidney Injury in the Patient with Cancer Kidney Res. Clin. Pract. 2019 38 295 308 10.23876/j.krcp.19.042 PMC6727896 31284363
4. Sandhu G. Adattini J. Gordon E.A. O’Neill N. Bagnis C. Chambers P. Martin J.H. Flynn A. Ibrahim K. Jardine M.J. Aligning Kidney Function Assessment in Patients with Cancer to Global Practices in Internal Medicine eClinicalMedicine 2025 82 103102 10.1016/j.eclinm.2025.103102 PMC12034077 40290845
5. Cosmai L. Porta C. Foramitti M. Perrone V. Mollica L. Gallieni M. Capasso G. Preventive Strategies for Acute Kidney Injury in Cancer Patients Clin. Kidney J. 2021 14 70 83 10.1093/ckj/sfaa127 PMC7857811 33564407
6. Lyrio R.M.D.C. Rocha B.R.A. Corrêa A.L.R.M. Mascarenhas M.G.S. Santos F.L. Maia R.D.H. Segundo L.B. De Almeida P.A.A. Moreira C.M.O. Sassi R.H. Chemotherapy-Induced Acute Kidney Injury: Epidemiology, Pathophysiology, and Therapeutic Approaches Front. Nephrol. 2024 4 1436896 10.3389/fneph.2024.1436896 PMC11341478 39185276
7. Hanna T.P. King W.D. Thibodeau S. Jalink M. Paulin G.A. Harvey-Jones E. O’Sullivan D.E. Booth C.M. Sullivan R. Aggarwal A. Mortality Due to Cancer Treatment Delay: Systematic Review and Meta-Analysis BMJ 2020 371 m4087 10.1136/bmj.m4087 PMC7610021 33148535
8. Floris M. Trevisani F. Angioi A. Lepori N. Simeoni M. Cabiddu G. Pani A. Rosner M.H. Acute Kidney Disease in Oncology: A New Concept to Enhance the Understanding of the Impact of Kidney Injury in Patients with Cancer Kidney Blood Press. Res. 2024 49 745 752 10.1159/000540908 39159615
9. Gupta S. Gudsoorkar P. Jhaveri K.D. Acute Kidney Injury in Critically Ill Patients with Cancer Clin. J. Am. Soc. Nephrol. 2022 17 1385 1398 10.2215/CJN.15681221 PMC9625110 35338071
10. Noble R.A. Lucas B.J. Selby N.M. Long-Term Outcomes in Patients with Acute Kidney Injury Clin. J. Am. Soc. Nephrol. 2020 15 423 429 10.2215/CJN.10410919 PMC7057296 32075806
11. Christiansen C.F. Johansen M.B. Langeberg W.J. Fryzek J.P. Sørensen H.T. Incidence of Acute Kidney Injury in Cancer Patients: A Danish Population-Based Cohort Study Eur. J. Intern. Med. 2011 22 399 406 10.1016/j.ejim.2011.05.005 21767759
12. Kitchlu A. McArthur E. Amir E. Booth C.M. Sutradhar R. Majeed H. Nash D.M. Silver S.A. Garg A.X. Chan C.T. Acute Kidney Injury in Patients Receiving Systemic Treatment for Cancer: A Population-Based Cohort Study JNCI J. Natl. Cancer Inst. 2019 111 727 736 10.1093/jnci/djy167 PMC6624169 30423160
13. Miyamoto Y. Andes L.J. Koyama A.K. Xu F. Pavkov M.E. Association Between Cancer and Incident Acute Kidney Injury Among Medicare Fee-for-Service Population Kidney Med. 2025 7 101069 10.1016/j.xkme.2025.101069 PMC12447095 40978193
14. Munch P.V. Nørgaard M. Jensen S.K. Birn H. Schmidt H. Christiansen C.F. Risk of and Mortality After Acute Kidney Injury Following Cancer Treatment: A Cohort Study Cancer Med. 2025 14 e70646 10.1002/cam4.70646 PMC11808929 39927749
15. Shi Y. Chen G. Lu Z. Wang H. Xu J. Li Y. Teng J. Acute Kidney Injury in Hospitalised Cancer Patients: Incidence, Risk Factors and Outcomes Nephrology 2025 30 e70025 10.1111/nep.70025 40189824
16. Launay-Vacher V. Oudard S. Janus N. Gligorov J. Pourrat X. Rixe O. Morere J. Beuzeboc P. Deray G. On behalf of the Renal Insufficiency and Cancer Medications (IRMA) Study Group Prevalence of Renal Insufficiency in Cancer Patients and Implications for Anticancer Drug Management: The Renal Insufficiency and Anticancer Medications (IRMA) Study Cancer 2007 110 1376 1384 10.1002/cncr.22904 17634949
17. Ciorcan M. Chisavu L. Mihaescu A. Gadalean F. Bob F.R. Negru S. Schiller O.M. Grosu I.D. Marc L. Chisavu F. Chronic Kidney Disease in Cancer Patients, the Analysis of a Large Oncology Database from Eastern Europe PLoS ONE 2022 17 e0265930 10.1371/journal.pone.0265930 PMC9183451 35679539
18. Campus A. Vischini G. Vetrano D. Di Nunzio M. Chiocchini A.L.C. Stalteri L. Grandinetti V. La Manna G. Renal Complications in Solid Tumor Patients Referred to an Onconephrology Clinic: A Three-Year Italian Experience G. Ital. Nefrol. 2025 42 2025-vol5 10.69097/42-05-2025-03 41184186
19. Kidney Disease: Improving Global Outcomes (KDIGO) Acute Kidney Injury Work Group. KDIGO Clinical Practice Guideline for Acute Kidney Injury Kidney Int. Suppl. 2012 2 1 138
20. Ostermann M. Bellomo R. Burdmann E.A. Doi K. Endre Z.H. Goldstein S.L. Kane-Gill S.L. Liu K.D. Prowle J.R. Shaw A.D. Controversies in Acute Kidney Injury: Conclusions from a Kidney Disease: Improving Global Outcomes (KDIGO) Conference Kidney Int. 2020 98 294 309 10.1016/j.kint.2020.04.020 PMC8481001 32709292
21. Churpek M.M. Carey K.A. Edelson D.P. Singh T. Astor B.C. Gilbert E.R. Winslow C. Shah N. Afshar M. Koyner J.L. Internal and External Validation of a Machine Learning Risk Score for Acute Kidney Injury JAMA Netw. Open 2020 3 e2012892 10.1001/jamanetworkopen.2020.12892 PMC7420241 32780123
22. Tomašev N. Glorot X. Rae J.W. Zielinski M. Askham H. Saraiva A. Mottram A. Meyer C. Ravuri S. Protsyuk I. A Clinically Applicable Approach to Continuous Prediction of Future Acute Kidney Injury Nature 2019 572 116 119 10.1038/s41586-019-1390-1 PMC6722431 31367026
23. Kidney Disease: Improving Global Outcomes (KDIGO) KDIGO 2026 Clinical Practice Guideline for Acute Kidney Injury (AKI) and Acute Kidney Disease (AKD): Public Review Draft Kidney Disease: Improving Global Outcomes Brussels, Belgium 2026 Available online: https://kdigo.org/wp-content/uploads/2026/03/KDIGO-2026-AKI-AKD-Guideline-Public-Review-Draft-March-2026.pdf (accessed on 3 September 2026)
24. Renaghan A.D. Ostermann M. Ronco C. Ballen K. Cosmai L. Fenoglio R. Floris M. Forni L.G. Gladstone D.E. Glezerman I.G. The Nephrotoxic Effects of Anti-Cancer Therapies: Consensus Report of the 34th Acute Disease Quality Initiative Workgroup Nat. Rev. Nephrol. 2026 22 283 300 10.1038/s41581-025-01031-3 41361704
25. Cerda J. Kashani K. Ostermann M. Basu R.K. Bell S. Cantaluppi V. Chakaravarthi R. Costa J.M. Claure-Del Granado R. Macedo E. The Global Epidemiology of Acute Kidney Injury: Challenges and Opportunities Nat. Rev. Nephrol. 2026 22 179 198 10.1038/s41581-025-01030-4 41350436
26. Ostermann M. Lumlertgul N. Jeong R. See E. Joannidis M. James M. Acute Kidney Injury Lancet 2025 405 241 256 10.1016/S0140-6736(24)02385-7 39826969
27. Perazella M.A. Onco-Nephrology: Renal Toxicities of Chemotherapeutic Agents Clin. J. Am. Soc. Nephrol. 2012 7 1713 1721 10.2215/CJN.02780312 22879440
28. Quickfall D. La A.M. Koyner J.L. 10 Tips on How to Use Dynamic Risk Assessment and Alerts for AKI Clin. Kidney J. 2024 17 sfae325 10.1093/ckj/sfae325 PMC11586629 39588357
29. Scanlon L.A. O’Hara C. Garbett A. Barker-Hewitt M. Barriuso J. Developing an Agnostic Risk Prediction Model for Early AKI Detection in Cancer Patients Cancers 2021 13 4182 10.3390/cancers13164182 PMC8393922 34439336
30. Tangri N. Cheungpasitporn W. Crittenden S.D. Fornoni A. Peralta C.A. Singh K. Usvyat L.A. Waterman A.D. on behalf of the American Society of Nephrology (ASN) Artificial Intelligence (AI) Workgroup. Responsible Use of Artificial Intelligence to Improve Kidney Care: A Statement from the American Society of Nephrology J. Am. Soc. Nephrol. 2026 37 881 890 10.1681/ASN.0000000929 PMC13065138 41201255
31. Abdelhafez M. Nayfeh T. Atieh A. AbuShamma O. Babaa B. Baniowda M. Hrizat A. Hasan B. Hassett L. Hamadah A. Diagnostic Performance of Fractional Excretion of Sodium for the Differential Diagnosis of Acute Kidney Injury: A Systematic Review and Meta-Analysis Clin. J. Am. Soc. Nephrol. 2022 17 785 797 10.2215/CJN.14561121 PMC9269645 35545442
32. Seethapathy H. Fenves A.Z. Fractional Excretion of Sodium (FENa): An Imperfect Tool for a Flawed Question Clin. J. Am. Soc. Nephrol. 2022 17 777 778 10.2215/cjn.06410522 PMC9269665 35613884
33. Gudsoorkar P. Langote A. Vaidya P. Meraz-Muñoz A.Y. Acute Kidney Injury in Patients with Cancer: A Review of Onconephrology Adv. Chronic Kidney Dis. 2021 28 394 401.e1 10.1053/j.ackd.2021.09.008 35190106
34. Darmon M. Vincent F. Canet E. Mokart D. Pène F. Kouatchet A. Mayaux J. Nyunga M. Bruneel F. Rabbat A. Acute Kidney Injury in Critically Ill Patients with Haematological Malignancies: Results of a Multicentre Cohort Study from the Groupe de Recherche En Réanimation Respiratoire En Onco-Hématologie Nephrol. Dial. Transplant. 2015 30 2006 2013 10.1093/ndt/gfv372 PMC4832999 26597921
35. Canet E. Zafrani L. Lambert J. Thieblemont C. Galicier L. Schnell D. Raffoux E. Lengline E. Chevret S. Darmon M. Acute Kidney Injury in Patients with Newly Diagnosed High-Grade Hematological Malignancies: Impact on Remission and Survival PLoS ONE 2013 8 e55870 10.1371/journal.pone.0055870 PMC3573047 23457485
36. Stevens P.E. Ahmed S.B. Carrero J.J. Foster B. Francis A. Hall R.K. Herrington W.G. Hill G. Inker L.A. Kazancıoğlu R. KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease Kidney Int. 2024 105 S117 S314 10.1016/j.kint.2023.10.018 38490803
37. Hodgson L.E. Sarnowski A. Roderick P.J. Dimitrov B.D. Venn R.M. Forni L.G. Systematic Review of Prognostic Prediction Models for Acute Kidney Injury (AKI) in General Hospital Populations BMJ Open 2017 7 e016591 10.1136/bmjopen-2017-016591 PMC5623486 28963291
38. Wang Y. Wei Y. Yang H. Li J. Zhou Y. Wu Q. Utilizing Imbalanced Electronic Health Records to Predict Acute Kidney Injury by Ensemble Learning and Time Series Model BMC Med. Inform. Decis. Mak. 2020 20 238 10.1186/s12911-020-01245-4 PMC7507620 32957977
39. Feng Y. Wang A.Y. Jun M. Pu L. Weisbord S.D. Bellomo R. Hong D. Gallagher M. Characterization of Risk Prediction Models for Acute Kidney Injury: A Systematic Review and Meta-Analysis JAMA Netw. Open 2023 6 e2313359 10.1001/jamanetworkopen.2023.13359 PMC12011341 37184837
40. Vagliano I. Chesnaye N.C. Leopold J.H. Jager K.J. Abu-Hanna A. Schut M.C. Machine Learning Models for Predicting Acute Kidney Injury: A Systematic Review and Critical Appraisal Clin. Kidney J. 2022 15 2266 2280 10.1093/ckj/sfac181 PMC9664575 36381375
41. Rehman A.U. Neyra J.A. Chen J. Ghazi L. Machine Learning Models for Acute Kidney Injury Prediction and Management: A Scoping Review of Externally Validated Studies Crit. Rev. Clin. Lab. Sci. 2025 62 454 476 10.1080/10408363.2025.2497843 40322979
42. Park N. Kang E. Park M. Lee H. Kang H.-G. Yoon H.-J. Kang U. Predicting Acute Kidney Injury in Cancer Patients Using Heterogeneous and Irregular Data PLoS ONE 2018 13 e0199839 10.1371/journal.pone.0199839 PMC6053162 30024918
43. Motwani S.S. McMahon G.M. Humphreys B.D. Partridge A.H. Waikar S.S. Curhan G.C. Development and Validation of a Risk Prediction Model for Acute Kidney Injury After the First Course of Cisplatin J. Clin. Oncol. 2018 36 682 688 10.1200/JCO.2017.75.7161 PMC5946720 29320311
44. Gupta S. Glezerman I.G. Hirsch J.S. Chen K.L. Devaraj N. Wells S.L. Seitter R.H. Kaunfer S.A. Jose A.M. Rao S.P. Derivation and External Validation of a Simple Risk Score for Predicting Severe Acute Kidney Injury after Intravenous Cisplatin: Cohort Study BMJ 2024 384 e077169 10.1136/bmj-2023-077169 PMC10964715 38538012
45. Sakuragi M. Uchino E. Sato N. Matsubara T. Ueda A. Mineharu Y. Kojima R. Yanagita M. Okuno Y. Interpretable Machine Learning-Based Individual Analysis of Acute Kidney Injury in Immune Checkpoint Inhibitor Therapy PLoS ONE 2024 19 e0298673 10.1371/journal.pone.0298673 PMC10950216 38502665
46. Thakar C.V. Arrigain S. Worley S. Yared J.-P. Paganini E.P. A Clinical Score to Predict Acute Renal Failure after Cardiac Surgery J. Am. Soc. Nephrol. 2005 16 162 168 10.1681/ASN.2004040331 15563569
47. Kheterpal S. Tremper K.K. Englesbe M.J. O’Reilly M. Shanks A.M. Fetterman D.M. Rosenberg A.L. Swartz R.D. Predictors of Postoperative Acute Renal Failure after Noncardiac Surgery in Patients with Previously Normal Renal Function Anesthesiology 2007 107 892 902 10.1097/01.anes.0000290588.29668.38 18043057
48. Li Y. Chen X. Wang Y. Hu J. Shen Z. Ding X. Application of Group LASSO Regression Based Bayesian Networks in Risk Factors Exploration and Disease Prediction for Acute Kidney Injury in Hospitalized Patients with Hematologic Malignancies BMC Nephrol. 2020 21 162 10.1186/s12882-020-01786-w PMC7201633 32370757
49. Li Y. Chen X. Shen Z. Wang Y. Hu J. Zhang Y. Xu J. Ding X. Prediction Models for Acute Kidney Injury in Patients with Gastrointestinal Cancers: A Real-World Study Based on Bayesian Networks Ren. Fail. 2020 42 869 876 10.1080/0886022X.2020.1810068 PMC7472473 32838613
50. Guo D. Bai J. Zhang J. Xi X. Chen Y. Luo Z. Feng K. Zeng J. Zhang M. Dong W. Multicenter Validation of an Explainable Machine Learning Model for Early Prediction of Acute Kidney Injury in Critically Ill Patients with Digestive System Tumors Digit. Health 2026 12 20552076261461368 10.1177/20552076261461368 PMC13269983 42311845
51. Okawa T. Mizuno T. Hanabusa S. Ikeda T. Mizokami F. Koseki T. Takahashi K. Yuzawa Y. Tsuboi N. Yamada S. Prediction Model of Acute Kidney Injury Induced by Cisplatin in Older Adults Using a Machine Learning Algorithm PLoS ONE 2022 17 e0262021 10.1371/journal.pone.0262021 PMC8765666 35041690
52. Ambe K. Aoki Y. Murashima M. Wachino C. Deki Y. Ieda M. Kondo M. Furukawa-Hibi Y. Kimura K. Hamano T. Prediction of Cisplatin-Induced Acute Kidney Injury Using an Interpretable Machine Learning Model and Electronic Medical Record Information Clin. Transl. Sci. 2025 18 e70115 10.1111/cts.70115 PMC11702448 39760531
53. Gupta S. Motwani S.S. Seitter R.H. Wang W. Mu Y. Chute D.F. Sise M.E. Glazer D.I. Rosner B.A. Curhan G.C. Development and Validation of a Risk Model for Predicting Contrast-Associated Acute Kidney Injury in Patients with Cancer: Evaluation in Over 46,000 CT Examinations Am. J. Roentgenol. 2023 221 486 501 10.2214/AJR.23.29139 37195792
54. Rodrigues N. Fragão-Marques M. Costa C. Branco C. Marques F. Vasconcelos P. Martins C. Leite-Moreira A. Lopes J.A. Predictive Risk Score for Acute Kidney Injury in Hematopoietic Stem Cell Transplant Cancers 2023 15 3720 10.3390/cancers15143720 PMC10377961 37509381
55. Yu X. Wu R. Ji Y. Huang M. Feng Z. Identifying Patients at Risk of Acute Kidney Injury among Patients Receiving Immune Checkpoint Inhibitors: A Machine Learning Approach Diagnostics 2022 12 3157 10.3390/diagnostics12123157 PMC9776989 36553164
56. Liu W. Ji K. Tang Q. Xia W. Zhang W. Shao L. Shi J. Li Y. Huang P. Ye X. Development and Validation of Interpretable Machine Learning Models for Predicting AKI Risk in Patients Treated with PD-1/PD-L1: A Retrospective Study BMC Med. Inform. Decis. Mak. 2025 25 295 10.1186/s12911-025-03142-0 PMC12333227 40781625
57. Lou Q. Luo L. Ye X. Zheng A. Ren Y. Zhang W. Bu S. Li Y. Zhu B. Shao L. Predictive Nomogram for Severe Acute Kidney Injury in Patients with Cancer Receiving Anti-PD-1/PD-L1 Antibodies: A Multicenter Retrospective Study Sci. Rep. 2025 15 33279 10.1038/s41598-025-16940-3 PMC12474994 41006410
58. Lin J. Dong X. Tang W. Liu S. The Incidence, Severity and Risk Factors of Renal Injury in Lung Cancer Patients Receiving Osimertinib Therapy: A Real-World Study Cancer Med. 2025 14 e71382 10.1002/cam4.71382 PMC12638203 41271435
59. Lei L. Wang Y. Xue Q. Tong J. Zhou C.-M. Yang J.-J. A Comparative Study of Machine Learning Algorithms for Predicting Acute Kidney Injury after Liver Cancer Resection PeerJ 2020 8 e8583 10.7717/peerj.8583 PMC7047869 32140301
60. Lee Y. Ryu J. Kang M.W. Seo K.H. Kim J. Suh J. Kim Y.C. Kim D.K. Oh K.-H. Joo K.W. Machine Learning-Based Prediction of Acute Kidney Injury after Nephrectomy in Patients with Renal Cell Carcinoma Sci. Rep. 2021 11 15704 10.1038/s41598-021-95019-1 PMC8333365 34344909
61. He S. Wu Y. Liu Y. Multicenter Machine Learning Study for Long-Term Prediction of Acute Kidney Injury after Complete Mesocolic Excision: Integrating Inflammatory Biomarkers and Transfusion-Related Risk Factors Front. Oncol. 2026 16 1782613 10.3389/fonc.2026.1782613 PMC13062227 41971434
62. Scanlon L.A. O’Hara C. Barker-Hewitt M. Barriuso J. Validation of a cancer population derived AKI machine learning algorithm in a general critical care scenario Clin. Transl. Oncol. 2025 27 3803 3810 10.1007/s12094-025-03906-0 PMC12399679 40140139
63. Motwani S.S. Choueiri T.K. Partridge A.H. Hu J. Kaymakcalan M.D. Waikar S.S. Curhan G.C. Comparison of Equations to Estimate Glomerular Filtration Rate and Their Impact on Frequency of Cisplatin-Associated Acute Kidney Injury Kidney360 2021 2 205 214 10.34067/KID.0000572020 PMC8741004 35373014
64. Mehran R. Aymong E.D. Nikolsky E. Lasic Z. Iakovou I. Fahy M. Mintz G.S. Lansky A.J. Moses J.W. Stone G.W. A Simple Risk Score for Prediction of Contrast-Induced Nephropathy after Percutaneous Coronary Intervention J. Am. Coll. Cardiol. 2004 44 1393 1399 10.1016/j.jacc.2004.06.068 15464318
65. Abramson M.H. Gutgarts V. Zheng J. Maloy M.A. Ruiz J.D. Scordo M. Jaimes E.A. Sathick I.J. Acute Kidney Injury in the Modern Era of Allogeneic Hematopoietic Stem Cell Transplantation Clin. J. Am. Soc. Nephrol. 2021 16 1318 1327 10.2215/CJN.19801220 PMC8729581 34135023
66. Renaghan A.D. Jaimes E.A. Malyszko J. Perazella M.A. Sprangers B. Rosner M.H. Acute Kidney Injury and CKD Associated with Hematopoietic Stem Cell Transplantation Clin. J. Am. Soc. Nephrol. 2020 15 289 297 10.2215/CJN.08580719 PMC7015091 31836598
67. Kanbay M. Mizrak B. Alper E.N. Copur S. Ortiz A. Acute Kidney Injury Following CAR-T Cell Therapy: A Nephrologist’s Perspective Clin. Kidney J. 2025 18 sfae359 10.1093/ckj/sfae359 PMC11704793 39781479
68. Vincendeau M. Joseph A. Thieblemont C. Rabian F. Harel S. Valade S. Zafrani L. Acute Kidney Injury after CAR-T Cell Therapy: Exploring Clinical Patterns, Management, and Outcomes Clin. Kidney J. 2024 17 sfae123 10.1093/ckj/sfae123 PMC11195623 38915438
69. Cortazar F.B. Kibbelaar Z.A. Glezerman I.G. Abudayyeh A. Mamlouk O. Motwani S.S. Murakami N. Herrmann S.M. Manohar S. Shirali A.C. Clinical Features and Outcomes of Immune Checkpoint Inhibitor–Associated AKI: A Multicenter Study J. Am. Soc. Nephrol. 2020 31 435 446 10.1681/ASN.2019070676 PMC7003302 31896554
70. Seethapathy H. Zhao S. Chute D.F. Zubiri L. Oppong Y. Strohbehn I. Cortazar F.B. Leaf D.E. Mooradian M.J. Villani A.-C. The Incidence, Causes, and Risk Factors of Acute Kidney Injury in Patients Receiving Immune Checkpoint Inhibitors Clin. J. Am. Soc. Nephrol. 2019 14 1692 1700 10.2215/CJN.00990119 PMC6895474 31672794
71. Meraz-Muñoz A. Amir E. Ng P. Avila-Casado C. Ragobar C. Chan C. Kim J. Wald R. Kitchlu A. Acute Kidney Injury Associated with Immune Checkpoint Inhibitor Therapy: Incidence, Risk Factors and Outcomes J. Immunother. Cancer 2020 8 e000467 10.1136/jitc-2019-000467 PMC7326260 32601079
72. Vanhoutte T. Sprangers B. Pseudo-AKI Associated with Targeted Anti-Cancer Agents—The Truth Is in the Eye of the Filtration Marker Clin. Kidney J. 2023 16 603 610 10.1093/ckj/sfad011 PMC10061433 37007700
73. Solanki S.L. Salunke B. Gangakhedkar G. Ambulkar R. Kuberkar D.V. Bhatt A. Acute Kidney Injury after Cytoreductive Surgery and Hyperthermic Intraperitoneal Chemotherapy—A Systematic Review Eur. J. Surg. Oncol. 2024 50 108622 10.1016/j.ejso.2024.108622 39216462
74. Li Y. Sperrin M. Ashcroft D.M. Van Staa T.P. Consistency of Variety of Machine Learning and Statistical Models in Predicting Clinical Risks of Individual Patients: Longitudinal Cohort Study Using Cardiovascular Disease as Exemplar BMJ 2020 371 m3919 10.1136/bmj.m3919 PMC7610202 33148619
75. Watson M. Chambers P. Steventon L. Harmsworth King J. Ercia A. Shaw H. Al Moubayed N. From Prediction to Practice: Mitigating Bias and Data Shift in Machine-Learning Models for Chemotherapy-Induced Organ Dysfunction across Unseen Cancers BMJ Oncol. 2024 3 e000430 10.1136/bmjonc-2024-000430 PMC11557724 39886186
76. Körner A. Sailer B. Sari-Yavuz S. Haeberle H.A. Mirakaj V. Bernard A. Rosenberger P. Koeppen M. Explainable Boosting Machine Approach Identifies Risk Factors for Acute Renal Failure Intensive Care Med. Exp. 2024 12 55 10.1186/s40635-024-00639-2 PMC11178719 38874694
77. Nagendran M. Chen Y. Lovejoy C.A. Gordon A.C. Komorowski M. Harvey H. Topol E.J. Ioannidis J.P.A. Collins G.S. Maruthappu M. Artificial Intelligence versus Clinicians: Systematic Review of Design, Reporting Standards, and Claims of Deep Learning Studies BMJ 2020 368 m689 10.1136/bmj.m689 PMC7190037 32213531
78. Vollmer S. Mateen B.A. Bohner G. Király F.J. Ghani R. Jonsson P. Cumbers S. Jonas A. McAllister K.S.L. Myles P. Machine Learning and Artificial Intelligence Research for Patient Benefit: 20 Critical Questions on Transparency, Replicability, Ethics, and Effectiveness BMJ 2020 368 l6927 10.1136/bmj.l6927 PMC11515850 32198138
79. Graversen H.V. Jensen S.K. Vestergaard S.V. Heide-Jørgensen U. Christiansen C.F. Defining Baseline Creatinine for Identification of AKI in Population-Based Laboratory Databases: A Danish Nationwide Cohort Study Kidney360 2022 3 232 241 10.34067/KID.0006082021 PMC8967652 35373126
80. Che Z. Purushotham S. Cho K. Sontag D. Liu Y. Recurrent Neural Networks for Multivariate Time Series with Missing Values Sci. Rep. 2018 8 6085 10.1038/s41598-018-24271-9 PMC5904216 29666385
81. Shukla S.N. Marlin B.M. Multi-Time Attention Networks for Irregularly Sampled Time Series Proceedings of the International Conference on Learning Representations Virtual Event 3–7 May 2021 PMC12927590 41737905
82. Heo S. Kang E.-A. Yu J.Y. Kim H.R. Lee S. Kim K. Hwangbo Y. Park R.W. Shin H. Ryu K. Time Series AI Model for Acute Kidney Injury Detection Based on a Multicenter Distributed Research Network: Development and Verification Study JMIR Med. Inform. 2024 12 e47693 10.2196/47693 PMC11263760 39039992
83. On Behalf of Topic Group ‘Evaluating Diagnostic Tests and Prediction Models’ of the STRATOS Initiative Van Calster B. McLernon D.J. Van Smeden M. Wynants L. Steyerberg E.W. Calibration: The Achilles Heel of Predictive Analytics BMC Med. 2019 17 230 10.1186/s12916-019-1466-7 PMC6912996 31842878
84. Vickers A.J. Elkin E.B. Decision Curve Analysis: A Novel Method for Evaluating Prediction Models Med. Decis. Mak. 2006 26 565 574 10.1177/0272989X06295361 PMC2577036 17099194