Artificial Intelligence in the Identification, Management, and Follow-Up of CKD

Challenges in Screening and Identification

Most CKD remains underrecognized as a problem until eGFR is <30–45 ml/min per 1.73 m², and the utility of broad population screening continues to be debated. In the absence of large, randomized trials for CKD screening, the best available evidence comes from observational studies and Markov models, and this literature suggests that screening is cost-effective for patients with diabetes or hypertension or for those with a family history of CKD. The most important factor in determining the cost-effectiveness of screening a cohort of patients is the expected prevalence of CKD in the sample (1). As such, a machine learning approach to classify a patient’s risk of developing CKD may be helpful.

In patients with diabetes, a model using seven routinely available features was developed by Roche in collaboration with IBM to determine incident CKD. The model used a medically supported feature selection strategy, and included the variables age, body mass index, eGFR, creatinine, albumin, glucose, and hemoglobin A1C (2). The final logistic regression model had decent discrimination, with an area under the receiver operating characteristic curve (AUC) of 0.79 in the combined data from the IBM Explorys database and the Indiana Network for Patient Care (INPC) database. Further, a random forest machine learning approach was applied in the INPC dataset, with an AUC of 0.83 for the seven-feature model and a modest improvement to 0.85 for a 35-feature model (2). External validation of the INPC random forest model would be recommended to establish broader utility.

Using a population health approach to identify undiagnosed CKD using data from electronic health records (EHRs) or laboratory information systems can facilitate rapid and cost-effective classification of many patients. However, when laboratory data are not available, insurers and health systems often have access to data from claims and medications. To date, multiple private companies in the CKD space (e.g., Strive Health, Cricket Health, and Monogram Health) all report that they apply machine learning methods to claims data to identify patients with or at risk for CKD. Unfortunately, none of their findings are published, and therefore they are impossible to evaluate for measures of diagnostic accuracy. Although it may be important for these organizations not to present their proprietary models, it would be valuable to present at least summary findings in distinct populations with appropriate peer review.

Changes in CKD Management

Although current guidelines recommend referral of patients with CKD stage 4 (eGFR <30 ml/min per 1.73 m²) to nephrology, there is still a high rate of low-risk referrals in patients with earlier stages of CKD, and similarly a significant number of high-risk patients are still referred late, when the window to prevent kidney failure no longer exists (3).

Since 2015, at least four drugs across two drug classes (sodium-glucose cotransporter 2 inhibitors and mineralocorticoid receptor antagonists) have been approved to slow CKD progression and to reduce cardiovascular events in patients at nearly all stages of CKD (4–7). Although these drugs, when added to renin-angiotensin-aldosterone inhibitors, have clinically meaningful benefits on major adverse kidney and cardiovascular events, the clinical benefit and cost-effectiveness is likely highest in patients at high or intermediate risk of kidney disease progression. In these patients, the number needed to treat to prevent adverse outcomes is likely to be low. When used early in the course of disease, these therapies can potentially help patients avoid dialysis for a lifetime rather than simply delay it by 1–2 years (Figure 1).

Figure 1.

Delay of progression afforded by novel therapies in patients with CKD.

Artificial Intelligence as a Solution

New organizations focused on artificial intelligence (AI)-augmented CKD care have begun to enter the market. One example, pulseData, has received a patent in 2021 for machine learning systems with respect to the management of kidney disease, which apply AI techniques to determine risk scores incorporating data related to demographics, vitals, diagnoses, procedures, diagnostic tests, biomarkers, genetic tests, and patient behaviors or symptoms. The algorithms require at least one of TNF receptor-1 or TNF receptor-2, and at least one laboratory result associated with kidney injury molecule-1. Additional biomarkers evaluated include eGFR, urine albumin-to-creatinine ratio, serum creatinine, tests from the comprehensive metabolic panel, lipid profile, coagulation panel, magnesium, phosphorous, brain natriuretic peptide, hemoglobin A1C, uric acid, and endostatin. Models for the prediction of kidney failure demonstrated excellent discrimination (C statistics >0.90 for 1-year prediction), whereas those for incident CKD had C statistics of 0.84 for 1-year prediction, 0.81 at 2 years, and 0.79 at 5 years (8).

An additional machine learning model has been developed by Renalytix AI, an in vitro diagnostics company. The model, referred to as KidneyIntelX, was developed for use as a clinical decision aid in the management of diabetic kidney disease using information from EHRs and biomarkers. This algorithm also uses TNF receptor-1, TNF receptor-2, and kidney injury molecule-1, and has demonstrated modest predictive accuracy, with a C statistic of 0.77 for the prediction of progression in diabetic kidney disease patients, outperforming the comparator clinical model (AUC of 0.61) (9). The model requires a total of more than 100 features, including the three plasma biomarkers, 27 other laboratory values, 20 ICD diagnostic codes, 30 medications, three measures of vital signs (systolic and diastolic blood pressure and body mass index), and it was developed on a population will fewer than 200 events (9).

Additional data presented from Renalytix AI show that individual biomarkers from the test are associated with CKD progression; however, their use in an algorithm to predict CKD progression has several limitations. First, the models are not externally validated, and given the tendency of machine models to overfit the development dataset, external validation is necessary before implementation. Second, for prediction models, calibration is equally as if not more important than discrimination when applying a model to a clinical risk threshold. The KidneyIntelX model consistently underpredicts across all quantiles of risk in internal validation. Third, an economic analysis of these models may be overly optimistic because it suggests delay or prevention of 5000 dialysis starts in a hypothetical cohort of 100,000 patients. Given the internal validation cohort has a kidney failure rate of 5% over 5 years, this would suggest that simply providing a risk score from a modestly accurate model would affect every dialysis start (9,10).

Next Steps and Future Directions

We believe the ideal solution would be an externally validated model that is broadly applicable in all stages of CKD, using routinely collected laboratory values that can be rapidly accessed from any conventional laboratory system. Involvement of only routinely collected laboratory tests can be helpful to avoid issues with specific assays, many of which can be relatively expensive, especially if they need to be applied to scale on a population with the magnitude of individuals with CKD. Integration of the risk-based information to EHR systems to communicate knowledge efficiently to care providers is also an integral quality of translating AI algorithms into practice, applying knowledge translation techniques to explain findings effectively to both providers and patients. In addition, it is crucial that as models become clinically used to establish a framework: (1) they allow updating of calibration or risk relationships as available treatments change; (2) they allow evaluated models to ensure that they do not disadvantage individuals on the basis of race or socioeconomic factors; and (3) they continue to be externally validated in diverse populations.

In conclusion, in an era of an aging population, rising medical costs, and novel therapies, a focus on personalizing medicine through highly accurate risk stratification can provide substantial benefits to the health care system. AI algorithms can be a useful tool to help guide these clinical decisions and to help align strained resources more efficiently.

Disclosures

T.W. Ferguson reports personal fees from Baxter, Inc., ClinPredict, Quanta Dialysis Technologies Ltd., and Strategic Health Resources, and personal fees/other from Klinrisk. N. Tangri reports grants, personal fees, and other from Tricida, Inc.; grants and personal fees from Astra Zeneca, Inc., Bayer, Boehringer Ingelheim/Eli Lilly, and Janssen; personal fees from Otsuka, Inc., and Roche; other from Mesentech and PulseData; and personal fees and other from ClinPredict and Klinrisk. ClinPredict and Klinrisk are engaged in efforts to develop and implement models for CKD progression in health systems.

Funding

None.

Author Contributions

Both authors conceptualized the study, wrote the original draft of the manuscript, and reviewed and editing the manuscript.