Over the last five years, electronic health records (EHRs) have been widely implemented in the United States, and health care systems now have access to vast amounts of data. While they are beginning to apply “big data” techniques to predict individual outcomes like post-operative complications and diabetes risk, big data remains largely a buzzword, not a reality, in the routine delivery of health care. Health systems are still learning how to broadly apply such analytics, outside of case examples, to improve patient outcomes while reducing spending. From a review of the literature on health systems that have successfully integrated predictive analytics in clinical practice, we have identified steps to make predictive algorithms an integrated part of routine patient care.
Determine the clinical decision. There is now a plethora of data available for nearly every potential clinical outcome. And where you have data, there is a potential
algorithm. But while it may be easy to develop clinical algorithms, it is equally necessary to be specific about which specific clinical decision(s) that algorithm will inform.
For example, there are many algorithms predicting a patient’s risk of hospital readmission (although the vast majority perform poorly). But simply knowing the percentage risk of readmission does not answer the questions that physicians and nurses typically ask before a patient is discharged: Should I discharge this patient now? Should I assign this patient to a readmission prevention intervention? Should this patient go to a short-term rehabilitation facility? Does she need a home care visit in the next two days?
Parkland Health and Hospital System in Dallas, Texas, has developed a validated EHR-based algorithm to predict readmission risk in patients with heart failure. Patients deemed at high risk for readmission receive evidence-based interventions, including education by a multidisciplinary team, follow-up telephone support within two days of discharge to ensure medication adherence, an outpatient follow-up appointment within seven days, and a non-urgent primary-care appointment. In a prospective study, the algorithm-based intervention reduced readmissions by 26%. Parkland’s success stems from focusing its algorithm on a specific population and tying it to discrete clinical interventions.
Leverage the data from EHRs. Algorithms are only as reliable as the data they are based on. While algorithms for acute clinical issues (e.g., heart attack, septic shock) may not require large amounts of data to predict risk, algorithms that utilize greater amounts of clinical data have greater accuracy and potential clinical applications.
The Veteran’s Health Administration (VHA), the largest health system in the United States, has collected electronic data from its patients for over three decades. Beginning in 2006, the VHA built a corporate data warehouse as a repository for patient-level data across its national sites. The sheer amount of inpatient and outpatient data has allowed the VHA to create comprehensive algorithms that reliably predict meaningful outcomes such as risk of death and hospitalization. Nurse care managers use these scores to guide intensity of outpatient services, including end-of-life and palliative care, delivered by multidisciplinary teams. The VHA’s investment in an integrated EHR and data repository — 5% of its total health spending — is substantial. However, the ability to reliably predict outcomes to improve quality of care may explain why the VHA’s net return on EHR investment is over $3 billion.
Focus on low-value decision points. Uncertainty over a clinical decision often leads physicians to overtreat or undertreat patients. Predictive analytics can allow clinicians to steer high-cost interventions to those high-risk patients who actually need them.
Consider the use of antibiotics to treat newborns. While less than 0.05% of all newborns have infection confirmed by blood culture, 11% of them receive antibiotics. Kaiser Permanente of Northern California has used predictive analytics to reduce this overuse. Its researchers have developed an algorithm to accurately predict the risk of severe neonatal infection based on a mother’s clinical data and the baby’s condition immediately after birth. Using this algorithm OB/GYNs can better determine which babies need antibiotics, sparing up to 250,000 American newborns each year from receiving unnecessary antibiotics. This could reduce medication costs and side-effects among vulnerable newborns.
Integrate — don’t force — analytics into the existing workflow. Physicians see hundreds of numbers (vital signs, laboratory values, etc.) each day. So there’s a danger that an algorithm’s output may just be another number that physicians ignore if it does not fit well into a daily workflow.
AWARE, a decision-support tool, is an example of how a predictive output can be iterative enough to be useful for a practicing physician. A critically ill patient’s risk of deteriorating varies constantly in the hospital. Thus, static risk percentages would be of little use to physicians, who need real-time information to make real-time decisions. AWARE’s decision-support algorithm output extracts high-value data from the EHR and uses analytics to reflect a patient’s current clinical status across organ systems. In prospective trials, providers in intensive care units using AWARE’s output were more efficient in gathering clinical information and made fewer cognitive errors — a perfect example of how just-in-time analytics can improve a provider’s existing workflow.
Predictive algorithms are becoming more complex and sophisticated. However, this sophistication means little if health systems cannot apply such algorithms to improve value in everyday clinical care. In this next era of value-based care, health systems must critically think about the clinical situations where enhanced analytics can be useful, help providers use them routinely in patient care, and develop strategies to evaluate the clinical impact of algorithms. By doing so, organizations can reduce spending and improve outcomes by targeting interventions to patients who need them the most.