Systematic Review of Mortality Risk Prediction Models for Patients with Kidney Failure

Academic Background

Kidney failure is a severe chronic condition, with a significant decline in both quality of life and survival rates compared to patients with other chronic diseases, including some forms of non-metastatic cancer. Patients with kidney failure often face difficult treatment decisions, such as whether to start or continue dialysis (peritoneal or hemodialysis) or choose conservative care (non-dialysis). To aid patients and clinicians in making informed decisions, mortality risk prediction models have been proposed as tools to support treatment choices by providing individualized risk estimates. However, challenges exist in applying these models in clinical practice, including their quality, relevance, and usability.

This study aimed to systematically review existing mortality risk prediction models for patients with kidney failure, assessing the quality and clinical applicability of these models. Through this research, the authors intended to provide guidance for developing future models and promoting their use in clinical decision-making.

Source of the Study

The paper was authored by Faisal Jarrar, Meghann Pasternak, Tyrone G. Harrison, Matthew T. James, Robert R. Quinn, Ngan N. Lam, Maoliosa Donald, Meghan Elliott, Diane L. Lorenzetti, Giovanni Strippoli, Ping Liu, Simon Sawhney, Thomas Alexander Gerds, and Pietro Ravani. The researchers are affiliated with institutions such as the University of Calgary in Canada, the University of Bari in Italy, and the University of Sydney in Australia. The paper was published on January 3, 2025, in JAMA Network Open under DOI e2453190.

Research Methods

Data Sources and Search Strategy

The research team conducted searches in databases including Ovid MEDLINE, Ovid Embase, and the Cochrane Library for studies published between January 1, 2004, and September 30, 2024. The search strategy combined three key concepts: chronic kidney failure, mortality, and prediction modeling. Predictive modeling-specific filters were employed, and all references in eligible studies were manually reviewed to ensure completeness.

Study Selection

After screening titles and abstracts of 7,184 articles, 77 studies were selected for full-text review, and 50 met the inclusion criteria. These studies encompassed 2,963,157 participants, with a median age of 64 years (range: 52–81 years) and a median proportion of female participants at 42% (range: 2%–54%). All models aimed to predict all-cause mortality, with prediction horizons ranging from 3 months to 10 years.

Data Extraction and Quality Assessment

Two reviewers independently extracted data and evaluated each study for risk of bias and applicability using CHARMS (Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies) and PROBAST (Prediction Model Risk of Bias Assessment Tool). They also considered adherence to TRIPOD+AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) reporting standards.

Prediction Framework and Model Evaluation

The team evaluated the prediction framework for each study, including the target population, prediction start time, prediction horizons, predictors, outcomes, and competing risks. Model validation methods were also assessed, including cross-validation, bootstrapping, and single random data splits for internal validation.

Research Findings

Study Characteristics

Among the 50 included studies, all models were developed using data from kidney failure patients already undergoing dialysis treatment. None of the studies included patients who had not yet decided on treatment. Most studies (78%) enrolled patients newly initiating dialysis, while a minority (4%) included those receiving dialysis for some time. All studies showed flaws in their analytical strategies, leading to high risks of bias. Additionally, many studies performed poorly in interpretability and usability for clinical practice.

Model Performance

Most studies used metrics such as the C-index and time-dependent AUC (area under the receiver operating characteristic curve) to evaluate model performance. However, these metrics have limited direct application to clinical decision-making. Only a few studies (4%) reported Brier scores, which measure both calibration and discrimination.

Clinical Usability

Among the 50 studies, only 15 (30%) provided user-friendly tools like risk calculators or nomograms to facilitate clinical application. Just one study (2%) conducted decision curve analysis to assess the net benefit of using the model in clinical practice.

Conclusions and Implications

This systematic review of 50 studies on mortality prediction models for patients with kidney failure revealed significant issues with bias risk and clinical applicability. Most models failed to meet the demands of clinical practice and lacked adequate validation and usability tools. The research team advocated for the development of new prediction models to better assist in treatment decisions for kidney failure patients.

Study Highlights

1. Comprehensive Literature Review: The research team conducted an exhaustive search across multiple databases to ensure coverage of all relevant studies.
2. Rigorous Evaluation Criteria: Tools like CHARMS and PROBAST were used to thoroughly assess bias and applicability.
3. Lack of Clinical Usability: Existing models showed considerable weaknesses in their application for real-world clinical decision-making.

Future Research Directions

Future research should focus on the following: 1. Inclusion of Undecided Patients: New prediction models should include patients who have not yet decided on dialysis to provide better decision-making support. 2. Improved Validation Methods: Future studies should adopt robust validation methods such as cross-validation or external validation to ensure model reliability and applicability. 3. Enhanced Usability: Risk calculators and other user-friendly tools should be developed, incorporating input from patients, caregivers, and clinicians to increase model acceptability and clinical uptake.

This systematic review offers critical guidance for the development and application of mortality risk prediction models for kidney failure patients, advancing research in this important field.