5 Ways Artificial Intelligence (AI) Is Impacting Kidney Disease

5 Ways Artificial Intelligence (AI) Is Impacting Kidney Disease

Healthcare practitioners constantly look for ways to improve the well-being of their patients. In nephrology care, promising advancements are happening through the application of artificial intelligence (AI). AI technology enables computer programs to emulate human behaviors, tasks, and outcomes, and it is utilized by virtually every industry, including healthcare.

Our leaders at Fresenius Medical Care North America (FMCNA) believe that AI-based kidney disease solutions are helping to transform the future of kidney care. Through machine learning (ML), we can provide physicians and clinicians more advanced tools to care for people with chronic kidney disease (CKD) — tools that can help detect and lower risks, decrease hospitalizations, improve outcomes, and forecast adverse events.

1. Decreasing Hospitalizations of Dialysis Patients with AI Kidney Disease Models

On average, in-center hemodialysis (HD) patients are hospitalized about twice every year, which is more often than 99 percent of all adults in the U.S., according to the Centers for Disease Control and Prevention.1 Hospital stays may cause disruptions to dialysis therapy schedules and can decrease patients’ quality of life while significantly increasing the cost of care to our healthcare systems. To help counter this trend, FMCNA uses AI to build predictive models and tools that show when patients are at a higher risk for hospitalization.

Our Imminent Hospitalization Predictive Model (IHPM) explores more than 1,000 variables from dialysis center records — including treatment data, lab values, notes from nurses and dietitians, and more. Computer algorithms process this information, tabulating a numerical risk score and a list of the top 10 predictors of hospitalization within seven days.

While the IHPM was designed to predict imminent hospitalizations, another ML model and tool — the Dialysis Hospitalization Reduction Program (DHRP) — can calculate the risk of multiple hospital admissions for a single patient within the coming 12 months.

Using historical data from 150,000 in-center HD patients, the DHRP identifies patterns in the factors leading to hospital admissions. By analyzing variables such as demographics, comorbidities, hospitalization/treatment history, labs, survey results, and even patients’ geographic region, the model can predict the number of hospitalizations a patient is likely to have in a year’s time.2

In a study, FMCNA scientists put the DHRP to the test in 54 dialysis centers over a two-year period. They saw a statistically significant lower “all-cause” patient hospitalization rate compared to other clinics where the DHRP was not implemented.2

The IHPM and DHRP AI models provide nurses a snapshot of patients’ risk factors, which may help improve workflow and enable timely care team interventions. So far, these tools have proven successful for small teams of our nurses and their patients. Our goal is to see these and other AI kidney disease solutions adopted more broadly, helping patients stay heathier and out of the hospital.

2. Reducing Therapy Interruptions for Peritoneal Dialysis Patients

While hospitalization risk is often top of mind for dialysis patients and caregivers, continuity of care is another area of great importance. We recently developed an AI innovation that can help us address the issue of patients leaving peritoneal dialysis (PD) therapy to start (or return to) in-center HD or home hemodialysis (HHD). Our PD Drop Risk ML tool analyzes patient data in the time frame leading up to a lapse in PD therapy, then identifies and reports on key risk factors for these stoppages. Predictors include patients’ lab results, infections, hospitalizations, medications, and others.

By examining the influencers of short-term (one- to three-month) and long-term (three- to six-month) shifts from PD to HD or HHD, we may prevent some lapses in PD therapy and enable smooth transitions between modalities, as required. To accomplish these objectives, FMCNA is piloting the PD Drop Risk model at a small number of our dialysis centers, with plans to expand implementation over time.

3. Identifying Signs of Arteriovenous Access Failure in Advance

For dialysis patients with an arteriovenous fistula (AVF) as vascular access, a primary concern is the risk of access failure. A study in the American Journal of Kidney Diseases showed that 23 percent of primary AVF placements fail for HD patients.3

Exploring the conditions and behaviors that lead to AVF deficiencies can help predict whether a fistula is likely to malfunction and allow for interventional procedures that can correct the problem. With this goal in mind, we developed an ML model to predict the likelihood of an AVF failure within 30 days.

Data gathered from more than 15,000 HD patients with a working AVF (and no history of AVF failure) were used to help determine the failure risk. Key predictors included the time elapsed since fistula placement, days since last treatment, and clinician observations. Variables such as lab values, treatment data, and clinical notes were also included in the model. So far, this model has shown the potential for providing early indications of access failure. 

Scientists at our Renal Research Institute also created a smartphone app that can analyze images of a dialysis patient’s vascular access and identify the presence and stage of an access aneurysm (a weak area in the vein or artery used to create an AVF or arteriovenous graft). Aneurysms can swell and possibly rupture if not detected and treated, causing bleeding and requiring immediate medical attention.4

The app uses a deep learning algorithm — like those used for facial recognition — to examine photos of accesses taken by patients or caregivers. The algorithm looks for patterns in the images and quickly alerts app users to the presence and severity of an aneurysm if one is discovered. Clinicians will then determine the best course of action, such as surgery or other measures. Our app can potentially extend the life of AVFs and arteriovenous grafts and lower the risk of ruptures. We are assessing the best methods of piloting these technologies at our dialysis centers with the eventual goal of widespread clinical usage.

4. Improving Dialysis Outcomes with Better Fluid Monitoring

One of the most critical aspects of dialysis therapy is maintaining optimal fluid levels in patients. Both fluid overload (hypervolemia) and fluid depletion can cause serious health issues, including heart problems, morbidity, or death. Our Renal Therapies Group was granted FDA Breakthrough Device Program designation for its computer-assisted ultrafiltration (UF) control software that can guide ultrafiltration rate to attain favorable relative blood volume targets. The Breakthrough Device Program intends to speed up the FDA’s development and review process and provide faster access to new technologies that treat life-threatening or debilitating diseases.

To design this software, we examined patient data gathered from 17 dialysis centers over a median 31-month span. The data were captured from our HD machines, patient electronic health records, and our CLiC™ device, a noninvasive tool that enables real-time monitoring of relative blood volume (RBV).5

RBV represents changes in blood volume (by percent) throughout a dialysis treatment, using a pre-treatment reading as a baseline. Relative blood volume monitoring (RBV-M) is the method used to track RBV. Studies have shown that HD patients have better outcomes when RBV decreases between −3 and −6.5 percent per hour during dialysis.6

Using CLiC technology and UF control software, we created a dashboard that demonstrates how RBV-M could be used to alert clinicians in real time when RBV numbers present a risk. A longer-term goal is to create an AI-enabled dialysis machine that guides clinicians in attaining target RBV levels.

5. Exploring COVID-19 Trends in Hemodialysis Patients

Our care teams see the same patients multiple times a week and use computer programs to enter health data after each visit. Using this information, we built an AI program that can detect changes in patient health and potentially flag when infected patients acquire COVID-19.

Our predictive model evaluated more than 80 variables (such as body temperature, heart rate, and interdialytic weight gain) from a pool of more than 40,000 HD patients. The model proved successful at identifying patients with an undetected COVID-19 infection three days before the onset of clinically suspicious symptoms. This AI application can help detect COVID-19 infections in HD patients earlier, enabling proactive isolation and testing to protect staff and other patients.7

In a separate but related undertaking, we also built a predictive model using data on COVID-19 deaths among HD patients in the U.S., Argentina, Colombia, and Ecuador. Mortalities were divided into three time-specific categories:

• Short-term — patients who died within 14 days of COVID-19 presentation
• Intermediate — patients who died 15–30 days after COVID-19 presentation
• Long-term — patients who died 30 days or more after COVID-19 presentation

By utilizing this advanced AI innovation, our researchers discovered several universal risk factors for death among the groups studied. These include demographics, comorbidities, vital signs, weight, chronic inflammation, and malnutrition. Our team concluded that predicting mortality in HD patients appears possible with the availability of specific data. With this type of analysis, we may be able to help HD patients with COVID-19 get the treatment they need to survive.

Artificial Intelligence and the Advancement of Kidney Care

In recent years, AI has gone from science fiction to widespread practical use, even in healthcare. We view AI as a major contributor to the future of renal technology and have developed more than a dozen advanced predictive models to improve how we treat and care for people with kidney disease.

With its capacity for high-volume, big-data analysis, AI can offer us insights that might otherwise never be realized. It provides us with the tools to fuel innovation and transform healthcare. Using AI kidney disease solutions, we can learn more about our patients, and offer them the most advanced, precise, and personalized treatments and care. 
 

CITATIONS:

1. “Summary Health Statistics: National Health Interview Survey,” Centers for Disease Control and Prevention, 2018, https://ftp.cdc.gov/pub/Health_Statistics/NCHS/NHIS/SHS/2018_SHS_Table_P-10.pdf.
2. Sheetal Chaudhuri, Hao Han, Len Usvyat, Yue Jiao, David Sweet, Allison Vinson, Stephanie Johnstone Steinberg, et al. “Machine Learning Directed Interventions Associate with Decreased Hospitalization Rates in Hemodialysis Patients,” Int J Med Inform 153 (September 2021): 104541. https://doi:10.1016/j.ijmedinf.2021.104541.
3. Ahmed A. Al-Jaishi, Matthew J. Oliver, Sonia M. Thomas, Charmaine E. Lok, Joyce C. Zhang, Amit X. Garg, Sarah D. Kosa, Robert R. Quinn, Louise M. Moist. “Patency rates of the arteriovenous fistula for hemodialysis: a systematic review and meta-analysis,” Am J Kidney Dis 63, no. 3 (March 2014): 464-78. doi: 10.1053/j.ajkd.2013.08.023. Epub 2013 Oct 30. PMID: 24183112.
4. Deep learning to classify arterio-venous access aneurysms in hemodialysis patients Hanjie Zhang, Dean Preddie, Warren Krackov, Murat Sor, PEter Waguespack, Zuwen Kuang, Xiaoling Ye, Peter Kotanko. Clinical Kidney Journal, sfab278 https://doi.org/10.1093/ckj/sfab278. Published:16 December 2021
5. Priscila Preciado, Hanjie Zhang, Stephan Thijssen, Jeroen P. Kooman, Frank M. van der Sande, Peter Kotanko. “All-cause mortality in relation to changes in relative blood volume during hemodialysis,” Nephrol Dial Transplant 34, no. 8 (August 1, 2019): 1401-8. doi:10.1093/ndt/gfy286.
6. Sheetal Chaudhuri, Hao Han, Caitlin Monaghan, John Larkin, Peter Waguespack, Brian Shulman, Zuwen Kuang, et al. “Real-time prediction of intradialytic relative blood volume: a proof-of-concept for integrated cloud computing infrastructure.” BMC Nephrol 22, no. 1 (August 9, 2021): 274. doi:10.1186/s12882-021-02481-0.
7. Caitlin K. Monaghan, John W. Larkin, Sheetal Chaudhuri, Hao Han, Yue Jiao, Kristine M. Bermudez, Eric D. Weinhandl, et al. “Machine Learning for Prediction of Patients on Hemodialysis with an Undetected SARS-CoV-2 Infection,” Kidney360 2, no. 3 (March 2021): 456-68. hoi: https://doi.org/10.34067/KID.0003802020.