Machine Learning for Risk Stratification in Kidney Disease

Abstract

Purpose of Review.

Risk stratification for chronic kidney is becoming increasingly important as a clinical tool for both treatment and prevention measures. The goal of this review is to identify how machine learning tools contribute and facilitate risk stratification in the clinical setting.

Recent Findings.

The two key machine learning paradigms to predictively stratify kidney disease risk are genomics-based and electronic health record-based approaches. These methods can provide both quantitative information such as relative risk and qualitative information like characterizing risk by subphenotype.

Summary.

The four key methods to stratify chronic kidney disease risk are genomics, multi-omics, supervised, and unsupervised machine learning methods. Polygenic risk scores utilize whole genome sequencing data to generate an individual’s relative risk compared to the population. Multi-omic methods integrate information from multiple biomarkers to generate trajectories and prognostic different outcomes. Supervised machine learning methods can directly utilize the growing compendia of electronic health records such as lab results and notes to generate direct risk predictions, while unsupervised machine learning methods can cluster individuals with chronic kidney disease into subphenotypes with differing approaches to care.

Introduction

Risk stratification for kidney disease has become globally imperative to improve health equity (1). Early identification of kidney disease can prove a useful tool for the treatment of perioperative and postoperative acute kidney injury (2,3), as well as chronic kidney disease (4).However, current measures of kidney function and structure are inadequate especially in the early stages of the disease. Fortunately, increasing multi-omics and health record data has permitted clinicians to stratify individuals prior to the deterioration of kidney function. However, modern machine learning methods are needed to appropriately leverage this wealth of multi-modal data.

There are two classes of machine learning methods that are currently being used to a priori stratify risk of kidney disease. The first involves genomics and multi-omics - utilizing genome-wide association studies (GWAS’s) as well as transcriptome and proteomic studies to generate an individual’s relative risk for developing a kidney disease. The second involves training machine learning methods on electronic health record data to both directly calculate risk, and as importantly, identify sub-phenotypes of chronic kidney disease. This review will cover the basics of generating a polygenic risk score, as well as the machine learning methods for direct stratification and sub-phenotyping (Figure 1). Finally, we will discuss the limitations of current studies as well as the future directions for applying machine learning into risk stratification for kidney disease.

Figure 1.

Four distinct approaches to machine learning for chronic kidney disease.

Biological Approaches for Risk Stratification

The creation of the large biobanks of genomic, transcriptomic, and multi-omic data has prompted a new wave of innovative approaches towards interpretable risk stratification. Here, we show two approaches using genomic data - data generated from whole genome sequencing of large populations (500,000+), and multi-omic data - data generated via whole genome sequencing in conjunction with RNA sequencing and protein analysis from smaller populations (< 100,000).

Polygenic Risk Scores for Risk Stratification

As whole genome sequencing (WGS) data becomes increasingly available for individuals with kidney disease, the ability to generate accurate and precise prognosticators of kidney disease has also improved. Some rare kidney diseases, like polycystic kidney disease, exhibit mendelian inheritance patterns, and therefore can be diagnosed via the identification of disease-causing variants (5). However, the genetic architecture of common phenotypes like chronic kidney disease is attributed to a non-linear mixture of rare variants such as the APOL1 G1 and G2 variants and common variants identified in genome-wide association studies (6).

The estimated contribution of common genetic variants to the expression of chronic kidney disease or the heritability is 39% (7). Genome-wide association studies work by running a linear regression on a given phenotype, with the independent variable representing each sequenced single-nucleotide polymorphism. These studies calculate a coefficient for each sequenced single-nucleotide polymorphism, and these coefficients can be aggregated for a single individual to generate a relative risk for an individual of a given phenotype (8).

This methodology, however, faces some issues that machine learning is helping resolve. Because genome-wide association studies are primarily conducted in European populations, modeling risk in individuals with European ancestry tend to be more accurate because of better accounting for linkage-disequilibrium patterns (9). Novel methods in multi-ethnic polygenic risk scoring has significantly improved polygenic risk scoring in individuals of admixed ancestry by linearly combining coefficients obtained from large samples of European populations and from smaller samples of individuals of diverse ancestry (10). Multi-ethnic polygenic risk for chronic kidney disease is powerful enough to stratify individuals by incident chronic kidney disease, end stage kidney disease, kidney failure and acute kidney injury (11). We anticipate that computational improvements in multi-ethnic polygenic risk as well as the increasing sample size of individuals with admixed ancestry through improving diversity in research will only further improve risk stratification via polygenic risk scoring (12).

Multi-omic Approaches to Risk Stratification

The heterogeneity of kidney disease presentation, treatment and etiology has strongly demonstrated the need for multi-omic approaches to the stratification of kidney disease risk (13). Trends towards growing data quantity and availability has lent itself to improvements in understanding the biological underpinnings of kidney disease. For example, the multi-omic database - a repository consisting of genes, proteins, microRNA, and metabolites - in the context of chronic kidney disease has been expanding since 2017 (14). Preliminary qualitative analysis of the tissue proteome have revealed insights into the biological underpinnings of kidney function during diabetic kidney disease (15).

Multi-omic machine learning approaches have shown the potential to prognosticate outcomes in chronic kidney disease (16). For example, methods like multi-PLIER utilizes differential network enrichment to generate a set of ‘latent’ variables used in downstream prognostication tasks. Integrative analysis of prognostic biomarkers from multi-omics panels help discriminate different trajectories of kidney function (17). Nevertheless, we anticipate that with growing databases, more state-of-the-art methods can be utilized to maximize risk prognostication. For example, graph convolutional neural networks have been used to create predictive models from protein-protein interaction graphs in the field of oncology (18). Future studies may hope to translate this into the complex protein interactions of kidney disease.

Machine Learning

The promise of machine learning in health care is to develop rich, hierarchical models that can accurately and interpretably provide novel clinical insights into the diagnosis and prognosis of disease pathology. The breath of readily available electronic health record data, especially for chronic kidney disease, has yielded important results for patients and clinicians alike.

Direct Risk Prediction

Direct risk prediction involves a class of algorithms called supervised machine learning algorithms, which utilize training data to predict an outcome. In the case for most of the following algorithms, biomarkers, electronic health records, and metabolites are all used as input, and trained to generate a direct prediction of chronic kidney disease or acute kidney failure.

Convolutional Neural Networks, a type of network that accounts for spatial relations within data, were used to develop a model to predict chronic kidney disease with a cohort of 2 million prospective individuals (19). Feedforward neural networks predicted the risk of end stage renal disease in patients with diabetic kidney disease from clinical trial data (20). Publicly available data from the UCI repository permits an empirical approach to model comparison for chronic kidney disease (21).

More nuanced approaches to machine learning for risk stratification - targeting subclinical markers of kidney disease have improved models and yielded insights into the etiology of kidney disease, as well as the associated comorbidities. Using structural data, text data and longitudinal data from electronic medical records from 64,059 individuals, natural language processing models (NLP) could detect the progression of diabetic kidney disease (22). The inclusion of metabolite biomarkers in a machine learning algorithm revealed that 13 metabolites out of 120 were identified to be significantly associated with the development of chronic kidney disease over the span of five years. (23). Machine learning also shows promise in kidney transplantation, with a systematic analysis of currently deployed models demonstrating high sensitivity and specificity for predicting graft failure. (24). An ensemble method applied to 671 cardiac surgery patients predicted the risk of acute kidney injury via preoperative biochemical data, preoperative medical and intraoperative time-series like hemodynamic changes. A study with 1,146 patients demonstrated AI-enabled prognosis can provide insight into the rapid decline of kidney function in diabetic patients (25). Computer assisted diagnosis of chronic kidney disease demonstrates potential for utilization in developing countries to improve screening in low-income and hard-to-reach settings (26).

However, most risk prediction algorithms are trained on patients from a single institution, which limits their generalizability. (27). For example, an external validation for 11 distinct machine learning algorithms to predict kidney failure demonstrated that most of the algorithms perform well in the short term, but perform poorly with horizons longer than 5 years on external populations (28). Institutions are primarily constrained by data-sharing issues, and therefore train algorithms within their own patient database, which limits external validity as well as generalizability. Federated learning presents a potential solution to this problem by allowing institutions to share the weights of a model rather than the patient data to train the model (29). Application of federated learning to acute kidney failure is relatively novel, but has demonstrated improved outcomes in the context of COVID-19 patients (30).

Risk Stratification by Clustering

One challenge that remains in diagnosing and prognosing kidney diseases is that they exhibit a heterogeneity in both expression and etiology (31). This makes risk stratification especially difficult, as individual risk is contingent on the subset of etiologies contributing to an individual’s phenotype. Unsupervised machine learning methods, machine learning methods that find similarity within the data - clustering, or outlier like anomaly detection, can and are be used to sub-phenotype kidney disease. A deep autoencoder is one method that can compress high-dimensional spaces into low-dimensional spaces, and these low-dimensional spaces are very tractable with methods like K-means clustering. One study demonstrated that this method can be used to compress electronic health record data from 4001 sepsis-associated kidney injury patients to generate four distinct sub-phenotypes with varying prognosis (32). Hierarchical clustering, a method that generates a set of hierarchical groups using similarity metrics, demonstrated four distinct subphenotypes of acute kidney injury with varying ages, ICU admission rates and comorbidities amongst clusters when trained upon distinct molecular and clinical patient signatures (33). Similarly, memory networks, a variant of the long-short term memory network, were trained on longitudinal and unstructured electronic health record data to produce patient representations that clustered into three distinct sub-phenotypes (34). This work is beginning to be applied translationally. For example, two different subphenotypes with different biomarker signatures in an AKI patient subset with 794 discovery and 425 validation patients were identified via latent factor analysis. Moreover, these patients respond differently to vasopressin therapy, which demonstrates the translational potential for sub-phenotyping (35).

Nevertheless, external validation remains a significant challenge for kidney machine learning models. First, as noted above, different datasets demonstrated different numbers of clusters in acute kidney failure as well as different characteristics in each cluster. External validation in unsupervised learning depends on metrics that are not currently tested in machine learning models. Several metrics exist for validating unsupervised machine learning models like cohesion, separation and mutual information (36). More rigorous validation of sub-phenotyping algorithms using these metrics on external datasets are necessary before integration of unsupervised cluster-based machine learning methods into the standard of care.

Conclusions

Machine learning for risk stratification in the context of kidney disease has shown great promise to translate complex presentations into direct risk predictions, as well as to disentangle the heterogeneity of etiologies of kidney disease into distinct sub-phenotypes or clusters. With growing amounts of data, the field of nephrology is positioned to optimally utilize machine learning algorithms to provide novel insights into diagnosis, prognosis and treatment of kidney disease. Funding for cross-site collaborations and federated learning approaches should be increased to further improve the generalizability and interoperability of machine learning models. To be successful, machine learning interventions in kidney disease require a team of engaged stakeholders and a systematic approach from ideation to implementation (37). While the ideation stage shows significant promise, we hope more nephrologists will take charge in implementing these ideas into clinical practice.

Key Points.

• Risk stratification for kidney disease is becoming an increasingly important part of prevention and treatment.

• Common genetic variants contribute significantly to the risk of kidney disease, leading to an interest in polygenic risk scores.

• There is an under-representation of individuals of non-European ancestry in published GWAS studies; more studies in diverse populations are indicated.

• EHR data has been demonstrated to be useful in predicting the risk of progression in diabetic kidney disease, graft failure in renal transplantation and identification of subphenotypes of acute kidney injury that differ in response to treatments.

• Rigorous external validation of discoveries based on EHR data has been lacking; thus increased funding for inter-institutional collaborations is necessary to help bridge this gap.