Predictive modeling improves care, cuts costs for kidney patients
Analytics and a team-based intervention method optimize care for at-risk patients, an approach Duke hopes to expand to those with diabetes, heart failure and surgical complications.
The predictive model that Duke Health uses to identify patients at risk of kidney failure is like “a watch tower in the sky,” says Will ElLaissi, program manager for the Innovation Living Lab at the Duke Institute for Health Innovation (DIHI). “We use a statistical model that can predict with 80 percent accuracy how long it will take a patient to hit end-stage renal disease,” he says.
Just as watch towers have guards who rapidly react to threats, DIHI brings a small team of clinical experts—including a nephrologist, care manager, and pharmacist—around a table to review each at-risk patient. With the help of a web-based, data analytics tool, team members can quickly access the most relevant information on each patient (e.g., kidney function trends, current medications), allowing them to make appropriate care decisions within a few minutes. DIHI calls this team-based method "population rounding."
“What we are doing is creating a completely new model of care to intervene on patients and shuttle them to the most appropriate care pathway, as well as the lowest possible cost of care setting,” ElLaissi says.
Treating kidney disease patients quickly and in the most effective care settings is important in preventing declines in patient conditions and delaying the need to move them to expense treatments, such as end-stage renal dialysis.
In partnership with Dev Sangvai, MD, Eugenie Komives, MD, Blake Cameron, MD and Mark Sendak, DIHI began using this two-pronged approach—a predictive model combined with analytics-driven population rounding—about one year ago with a Medicare population covered through Duke Health’s accountable care organization. Of the 46,000 patients in that population, the predictive model identified 1,100 as having a greater than 15 percent risk of developing end-stage kidney disease within five years. More than half of these patients were weeded out due to various statistical constraints (e.g., they had moved away), which left 413 patients for the team to round on and determine the most appropriate care pathway.
Most commonly, the patients are referred to a nephrologist, ElLaissi says. For patients whose kidney disease is stable enough to be managed in a primary care setting, the team consults with the patient’s primary care physician about medications, patient education, and care management strategies. Some patients are connected to disease management or social work services, as needed. When necessary, patients are fast-tracked for a kidney transplant.
“We have identified at-risk patients who would not have been identified if we did not have population rounding,” says DIHI’s Suresh Balu, program director. “Prior to this, we might have only caught them when they needed dialysis. By getting to them early on and providing a nephrology consult, we can level their kidney function and push out their dialysis by years.”
Chronic kidney disease is known as a silent killer because it often causes severe organ damage before symptoms (e.g., blood in the urine) develop. To identify those patients who slip through the cracks, DIHI’s predictive model assesses numerous variables associated with kidney failure, including age, race, blood pressure, and key lab values such as creatinine levels and estimated glomerular filtration rate (eGFR).
DIHI teamed up with Larry Carin, Vice Provost for Research at the Department of Statistics at Duke University, to employ machine learning to create the predictive model. “It is about applying Bayesian methods to predict what kind of trajectory people are going to take,” Balu says.
The data for the model is pulled from Duke Health’s electronic health record (EHR) and claims data. Patients typically get their creatinine and eGFR measured whenever they get a common blood test, like a basic metabolic panel. But, until now, those values were buried in the patient’s record, making it difficult to detect trends. “Over a period of five or six years, we can see whether a patient’s eGFR value is dropping, but nobody notices because no one sees the data in a form that is actionable,” Balu says.
Now, the Duke Health team members who gather for the population rounding meetings can see each patient’s lab trends displayed graphically via the data analytics tool, alongside other key information like medical history and current medications. Plus, each patient is assigned a risk score by the predictive model.
“The clinicians can more clearly see the best care opportunity for this patient,” ElLaissi says. “Then they make notes and order those notes to different people. Let’s say we decide to change the medications for a patient. That note is made to the pharmacist, and the pharmacist can pull up a discrete view and see his or her action items for all of the patients rounded on.”
Following the success of its kidney disease initiative, DIHI is moving onto congestive heart failure and diabetes. “We have started the modeling and are working with subspecialists to identify the care pathways for patients,” ElLaissi says. “This is something in which we see a lot of real-world value.”
In addition, Duke Health has created a machine learning model that can predict a patient’s risk of developing surgical complication after surgery, by working in close collaboration with Erich Huang, MD and Katherine Heller, MD. Crucial to developing this model was the availability of national data from the American College of Surgeon’s National Surgical Quality Improvement Project (NSQIP).
“The database from NSQIP includes every single patient in the nation who has undergone surgery and what complications have occurred in the first three days post-surgery,” Balu says. “Based on that data, we were able to predict surgical complications based on a machine learning model that our team developed. The model can be extended and customized to a large hospital or a community hospital, and it can predict absolute and relative risk.”
Duke Institute for Health Innovation is currently running an internal trial of the model to test its effectiveness and determining the appropriate care interventions to take when a patient is identified as being at risk for a surgical complication.
Balu commends DIHI’s multidisciplinary approach to population health for the institution’s success to date. “This requires strong data scientists, machine learning experts, and clinicians,” he says. “We’re bringing all these people together to look at the specific problems. By doing so, we are improving the ‘system-ness’ of the health system."
Just as watch towers have guards who rapidly react to threats, DIHI brings a small team of clinical experts—including a nephrologist, care manager, and pharmacist—around a table to review each at-risk patient. With the help of a web-based, data analytics tool, team members can quickly access the most relevant information on each patient (e.g., kidney function trends, current medications), allowing them to make appropriate care decisions within a few minutes. DIHI calls this team-based method "population rounding."
“What we are doing is creating a completely new model of care to intervene on patients and shuttle them to the most appropriate care pathway, as well as the lowest possible cost of care setting,” ElLaissi says.
Treating kidney disease patients quickly and in the most effective care settings is important in preventing declines in patient conditions and delaying the need to move them to expense treatments, such as end-stage renal dialysis.
In partnership with Dev Sangvai, MD, Eugenie Komives, MD, Blake Cameron, MD and Mark Sendak, DIHI began using this two-pronged approach—a predictive model combined with analytics-driven population rounding—about one year ago with a Medicare population covered through Duke Health’s accountable care organization. Of the 46,000 patients in that population, the predictive model identified 1,100 as having a greater than 15 percent risk of developing end-stage kidney disease within five years. More than half of these patients were weeded out due to various statistical constraints (e.g., they had moved away), which left 413 patients for the team to round on and determine the most appropriate care pathway. Most commonly, the patients are referred to a nephrologist, ElLaissi says. For patients whose kidney disease is stable enough to be managed in a primary care setting, the team consults with the patient’s primary care physician about medications, patient education, and care management strategies. Some patients are connected to disease management or social work services, as needed. When necessary, patients are fast-tracked for a kidney transplant.
“We have identified at-risk patients who would not have been identified if we did not have population rounding,” says DIHI’s Suresh Balu, program director. “Prior to this, we might have only caught them when they needed dialysis. By getting to them early on and providing a nephrology consult, we can level their kidney function and push out their dialysis by years.”
Chronic kidney disease is known as a silent killer because it often causes severe organ damage before symptoms (e.g., blood in the urine) develop. To identify those patients who slip through the cracks, DIHI’s predictive model assesses numerous variables associated with kidney failure, including age, race, blood pressure, and key lab values such as creatinine levels and estimated glomerular filtration rate (eGFR).
DIHI teamed up with Larry Carin, Vice Provost for Research at the Department of Statistics at Duke University, to employ machine learning to create the predictive model. “It is about applying Bayesian methods to predict what kind of trajectory people are going to take,” Balu says.
The data for the model is pulled from Duke Health’s electronic health record (EHR) and claims data. Patients typically get their creatinine and eGFR measured whenever they get a common blood test, like a basic metabolic panel. But, until now, those values were buried in the patient’s record, making it difficult to detect trends. “Over a period of five or six years, we can see whether a patient’s eGFR value is dropping, but nobody notices because no one sees the data in a form that is actionable,” Balu says.
Now, the Duke Health team members who gather for the population rounding meetings can see each patient’s lab trends displayed graphically via the data analytics tool, alongside other key information like medical history and current medications. Plus, each patient is assigned a risk score by the predictive model.
“The clinicians can more clearly see the best care opportunity for this patient,” ElLaissi says. “Then they make notes and order those notes to different people. Let’s say we decide to change the medications for a patient. That note is made to the pharmacist, and the pharmacist can pull up a discrete view and see his or her action items for all of the patients rounded on.”
Following the success of its kidney disease initiative, DIHI is moving onto congestive heart failure and diabetes. “We have started the modeling and are working with subspecialists to identify the care pathways for patients,” ElLaissi says. “This is something in which we see a lot of real-world value.”
In addition, Duke Health has created a machine learning model that can predict a patient’s risk of developing surgical complication after surgery, by working in close collaboration with Erich Huang, MD and Katherine Heller, MD. Crucial to developing this model was the availability of national data from the American College of Surgeon’s National Surgical Quality Improvement Project (NSQIP).
“The database from NSQIP includes every single patient in the nation who has undergone surgery and what complications have occurred in the first three days post-surgery,” Balu says. “Based on that data, we were able to predict surgical complications based on a machine learning model that our team developed. The model can be extended and customized to a large hospital or a community hospital, and it can predict absolute and relative risk.”
Duke Institute for Health Innovation is currently running an internal trial of the model to test its effectiveness and determining the appropriate care interventions to take when a patient is identified as being at risk for a surgical complication.
Balu commends DIHI’s multidisciplinary approach to population health for the institution’s success to date. “This requires strong data scientists, machine learning experts, and clinicians,” he says. “We’re bringing all these people together to look at the specific problems. By doing so, we are improving the ‘system-ness’ of the health system."
More for you
Loading data for hdm_tax_topic #better-outcomes...