Patient risk can improve value-based cancer care. Find out how.
Risk stratification and predictive modeling can translate clinical oncology data into better decision making to protect chemotherapy patients and curb avoidable cancer care costs.
That’s according to speakers at the AMCP Managed Care & Specialty Pharmacy Annual Meeting, in Denver, during the March 28 session, “Using Predictive Analytics to Improve Value-Based Cancer Care Delivery.”
Cancer care quality and cost management “drive health plan actuaries crazy,” said Andrew Hertler, MD, FACP, chief medical officer of New Century Health. “It has the ‘perfect storm’ of challenges for predictability: high variability, high cost, and low volume.”
Chemotherapy drugs represent 20% of cancer care costs but potentially-avoidable chemotherapy-related hospitalizations represent nearly as much, 18%, making it a major cost driver in clinical oncology, Hertler said.
Hospitalizations among patients undergoing cancer treatment are frequent and “commonly related to chemotherapy toxicity,” he said. “Studies have shown that many chemotherapy-related hospitalization risks are predictable using routinely collected clinical data.”
Data can reduce hospitalizations
Clinical data can be employed in two ways to reduce chemotherapy-associated hospitalizations, Hertler said: risk stratification modeling and predictive analytics.
Risk stratification uses prospective clinical data collected during treatment (prior to authorization), allowing timely use of that data.
“While it is nowhere near as robust as claims data, claims data has a lag of up to 30 days-and that’s not going to do us any good,” he explained.
Analyzing clinical data, New Century Health developed a proposed risk-stratification model to identify which patients are most likely to wind up in the emergency room-a major predictor of hospitalization and inpatient care. ER admissions risk factors in their risk stratification model include emetogenic or neutropenic chemotherapies, patient age, cancer stage and functional performance status, and line of cancer therapy.
The model also includes patient body-mass index (BMI) or weight “and particularly changes in weight,” Hertler said. “When a patient begins losing weight, that’s a pretty global indicator that they are not doing well.”
Each risk factor is weighted in the model, depending on how important a contributor it is to hospitalization risk. The result is a hospitalization risk-stratification score (high, moderate or low) that can be used to guide clinical decisionmaking, patient monitoring, and anticipation of interventions.
The risk stratification model is easy to use, Hertler said. But it has limited data and lacks comorbidity data that can also influence hospitalization risk, he noted.
Predictive modeling capabilities
Whereas risk stratification models use data “up front,” early in patient care, to identify which patient subpopulations are most at risk, predictive models can be more dynamic, helping to assess the probability of a future event based on “input” variables, said Mir T. Mirnazari, PhD, assistant vice president at New Century Health.
In other words, predictive modeling can predict specific behaviors or outcomes based on particular clinical decisions.
Predictive modeling in clinical oncology can be challenging, Mirnazari cautioned: Patient volumes are small, particularly for any given type and stage of cancer-and most cases are “very high-risk.”
“To build a predictive model, we rely on a large and complete historic dataset that contains the risk factors and outcome information for every patient, using authorization data for information on predictions and claims data to identify outcomes, like chemotherapy costs,” he explained.
Building predictive models
Building predictive models involves splitting data into three sets:
The estimation and test sets are used to generate a model; the validation set is used to make sure the model works-that it accurately predicts patient outcomes.
It is crucial that the clinical and cost-analytic objectives (preventing inappropriate chemotherapy utilization, for example) be clearly articulated before modeling is attempted, Mirnazari emphasized.
Also key is a sufficiently large set of patient data. About 1,000 patients is probably the minimum size for reliable model development and validation, he said.
Model testing and validation assess sensitivity (true-positive) and specificity (true-negative) rates: the percentage of patients correctly identified as having, or not having, respectively, a predicted condition or outcome like high chemotherapy costs. Sensitivity and specificity often behave independently; they do not always improve or worsen together as models are adjusted.
Predictive modeling is iterative; if testing or validation reveals insufficient sensitivity or specificity, then it’s back to the drawing board to “start from scratch,” Mirnazari said.
Once internally validated, predictive models can be validated with new datasets. Once a model is deemed robust and predictive, it can be combined with clinical pathways to create “a continuous learning loop that leads to better cancer care delivery,” Mirnazari said.
“It’s not magic,” he concluded. “The completed model is essentially a simple equation. On the left side is … your predicted outcome. On the right side are the risk factors, each of them associated with a predictive weight or value. That becomes a mathematical equation and you take it to your IT department to implement.”
“At the end of the day, it’s a simple equation,” Mirnazari said.