Overview
Chronic heart failure poses a major public health problem in developed countries, including in the United States. The use of new treatments and medications has vastly improved the outcomes of patients with heart failure in recent decades, but mortality in this population still remains high.
Among patients in the Veterans Health Administration (VHA) in particular, heart failure is the leading cause of admissions among ambulatory care-sensitive conditions, responsible for twice as many hospitalizations as the second-leading cause, chronic obstructive pulmonary disease. In addition, researchers have found that there is an association with post-traumatic stress disorder (PTSD) and heart failure, a devastating correlation for the US Veteran population.
Chronic heart failure is one of the most high-cost cardiovascular conditions in the VHA, resulting in high readmission rates, lowered quality of life, and greater risk for mortality. Patients with heart failure have less adherence to medications and dietary recommendations, which plays a role in patient outcomes. Roughly 22% of Veterans with heart failure are readmitted within 30 days, and 25% of them die within one year of discharge. Finally, gender disparities exist in the cardiovascular disease population of Veterans, further complicating treatment and outcomes.
Challenges
Heart failure can be difficult to manage because the disease is variable. A treatment regimen that one Veteran responds to may not work for another Veteran, and vice versa. When a Veteran is discharged from the hospital, their therapy must be regularly adjusted based on their dynamic physiological state. The heart failure management guidelines provided by the VHA indicate that hospitalizations due to heart failure could be largely prevented if ambulatory care were provided in a timely and effective manner. But siloed data delays the intervention.
The VHA needed a robust synthetic version of a patient cohort that they could use to build a neural network model to better understand their data and predict which heart failure patients were at risk of readmission after an initial event. They could then potentially use this model in a clinical setting to gain insights to prescribe better treatments and ultimately decrease the occurrence of adverse events for heart failure patients.
KEY QUESTIONS
The VHA sought to use a recurrent neural network (RNN) model to capture non-linear temporal trends in patients’ journeys to predict readmission risk of heart failure patients. The VHA wanted to know:
-
What kind of data did it need to generate accurate predictions?
-
What variables did it need to add to its patient cohort model to be able to accurately assess heart failure outcomes?
-
Could it obtain a robust enough synthetic version of its patient cohort to develop a strong predictive model that could then be shared with third parties?
The VHA aimed to develop an RNN that would enable prediction of adverse events in its heart failure patient population. After a successful RNN model is developed, the VHA will begin working to validate the model with real setting or aclinical trial before any decision is made for deployment at the point of care.
Results to Date
The VHA performed an initial cohort analysis and included a wide variety of clinical variables such as labs, conditions, and vital signs to build a prediction model to understand the phenotypes of patients more likely to die after discharge. They generated several synthetic derivatives of real data using MDClone technology in the Arches platform at the VHA IE by iteratively adding clinically relevant attributes for the cohort while preserving the fidelity of the synthetic data. The predictive models for adverse events were trained on these synthetic datasets and then re-run on de-identified, underlying data.
The team was able to demonstrate that the model results obtained from Synthetic data were statistically similar to those using the deidentified data, thereby enabling the VHA to begin developing a strong predictive model and gain early insights within the constraints of their data governance policies for sharing data for external collaborations.
Conclusion
Without MDClone, the VHA IE would have faced significant data governance challenges that would have presented hurdles to overcome and delay the innovation. With MDClone, the VHA is empowered to share synthetic data freely without risking patient privacy. Additionally, using MDClone’s tools, the VHA was able to generate high-quality synthetic data for use in its RNN model, enabling it to create early models, validate its results, and ultimately drive down cycle time for learning.
Finally, the VA Boston Healthcare System and Empallo are in a collaboration applying machine learning and AI techniques to existing synthetic and patient-derived clinical data to predict future clinical events as part of a quality-improvement initiative. This collaboration will inform future prospective evaluations of this AI technology, which the VHA can then use on its own synthetic data.
