A brighter future for kidney disease?

The Lancet Journal

PERSPECTIVES|DIGITAL MEDICINE| VOLUME 395, ISSUE 10219, P179, JANUARY 18, 2020

A brighter future for kidney disease?
Evan D Muse
Eric J Topol
Published:January 18, 2020DOI:https://doi.org/10.1016/S0140-6736(20)30061-1

It wasn't until the 19th century that diseases of the kidney began to be recognised under the eponymous diagnosis of Bright's disease, named after the father of nephrology Richard Bright (1789–1858). While our understanding of various types of kidney injury and pathologies has broadened over time, the complexities and clinical overlap of kidney diseases have led to less than adequate prevention and treatment strategies. Chronic kidney disease is present in about 10% of the world's population—more than diabetes and cancer combined—ranks as the ninth leading cause of death in the USA, and accounts for billions of dollars in medical costs, suffering, and lost quality life-years. Despite this high burden, the public and most medical providers rarely give chronic kidney disease the priority it deserves. However, machine learning predictive algorithms are poised to make a real difference here.
One of the main barriers to early detection and subsequent prevention of kidney disease is access to and feasibility of testing, especially when it comes to urine studies. Most patients find it commonplace to have blood collected for tests, but providing a urine sample is not always so convenient. Notwithstanding guideline recommendations for annual urine screening for individuals with hypertension and diabetes, these tests are sometimes skipped. Smartphone app-based platforms for urine testing could improve adherence to albumin creatinine ratio (ACR) testing. One study showed screening of at-risk patients almost doubled with a home urine test kit that uses a smartphone camera to easily and accurately quantify ACR from a user-performed urine dipstick. If independently validated in a large, diverse population, this low-cost strategy could change the often dim trajectory for individuals with declining kidney function. A similar approach has been taken for urinary tract infection diagnosis through pharmacy or home-based testing, with the goal not only to minimise barriers to appropriate testing but also to enhance antibiotic stewardship. This low-cost, app-based approach could facilitate broader home screening for at-risk individuals and more frequent serial testing for people with abnormal kidney function. Ultimately, this technology has the potential to help prevent and slow the progression of kidney disease in both low-income and high-income settings.

The identification of meaningful changes in kidney function on the basis of blood and urine biomarkers is inadequate. That it takes nearly a 50% decline in glomerular filtration rate before a noticeable change is seen in serum creatinine—one of the most common metrics used to track kidney function—is astounding. As we await the identification of new biomarkers and validation of developing biomarkers for assessing overall kidney function, novel analytics and machine learning algorithms are surfacing that might provide deeper predictive power to existing large electronic health record datasets. Although automated diagnostic and treatment pathway algorithms for kidney disease are not new, the use of machine learning tools could improve their predictive power and thereby promote prevention.

An in-depth retrospective analysis of 703 782 patients in the US Department of Veterans Affairs database showed the potential of a deep learning approach to anticipate the high risk of acute kidney injury in hospital inpatients up to 48 h before it occurred. Additionally, this model independently predicted the need for haemodialysis within that time for more than nine of every ten occurrences of kidney injury that progressed to that stage. These findings are noteworthy given the complex and multifactorial causes of acute kidney injury in hospital patients, such as sepsis, shock, heart failure, hypovolaemia, and drug or contrast dye toxicity. In the outpatient setting, a Japanese team used machine learning and natural language processing to predict disease progression and need for dialysis over 6 months in patients with diabetic nephropathy. And while the increased risk of contrast-induced acute kidney injury has been long appreciated, a machine learning algorithm trained and tested on 3 million adults effectively quantified the degree of kidney injury on the basis of the volume of contrast used and individual patient-level characteristics.
Although the performance of each of these models is impressive, the challenge now is that prospective clinical trials have yet to be done to validate these predictive algorithms. At some point, if and when it is shown that kidney damage can be pre-empted by use of algorithms in clinical practice, we'll be able to declare that the scene has brightened.