By Robert Pearl, MD | | October 27, 2016

In my last article, I highlighted the shortcomings of some recent, highly touted medical devices and technology “breakthroughs.” A recent JAMA article underscored that point, documenting the failure of exercise monitors to deliver the lifestyle changes and health results promised.

In contrast to technologies that fail to make much of a difference, I offer three that are already delivering the healthcare value anticipated. Last week I highlighted video, a solution relatively inexpensive to acquire that enables physicians to provide care in real time wherever the patient is, and at a lower cost than an in-person office visit. This week, I examine the opportunities available through predictive analytics.

Predictive analytics uses data to establish the best approaches to current medical problems and identify opportunities to intervene today to avoid problems in the future. Rather than being a specific tool, it represents a wide-ranging catalog of possibilities and applications. Today we have only scratched the surface of what is attainable, both in terms of the number of problems to which it can be applied and the specificity for each.

These computer-derived analytics tools can support crucial medical decision-making at no extra cost, with applications that run on currently available computers and mobile devices. The Permanente Medical Group (TPMG), which I lead, makes these clinical tools available to our 9,000 physicians, and they currently provide major improvements in healthcare outcomes. None of them are sold as commercial products.

More than 50 years ago, Dr. Sidney Garfield, the founder of TPMG and its Division of Research (DOR), recognized the potential to combine the power of information technology, even in its earliest iterations, with clinical care and sophisticated data analytics. Today, as one of the nation’s largest research facilities outside of a government or university setting, the DOR derives its power from being a clinical research enterprise embedded in a large, technologically enabled, integrated delivery system. As a result, researchers and physicians caring for patients are directly linked. The problems studied are those most vexing for doctors and patients. And as research discovers the answers, the solutions are systematically and rapidly disseminated and implemented in clinical practice, far faster than the time typically required in community practice.

For the past decade, TPMG researchers have used our robust and comprehensive databases, with information on more than 12 million patients, to understand what is best for patient care, always protecting patient privacy. Next month the Journal of Hospital Medicine will publish the first in a series of articles on this topic, detailing how particular data analytics tools have improved clinical outcomes for thousands of patients. In this column, I offer a preview of the types of opportunities identified and the superior results achieved.

Which Patients Will Need the ICU?

TPMG physician Dr. Gabriel Escobar and his team have developed a tool that serves as an early warning system. It analyzes clinical information about patients sick enough to need hospital care but not appearing sick enough to require treatment in an intensive care unit (ICU). Using information from more than 1 million patients, the tool accurately predicts which patients are most likely to deteriorate later in the day and require transfer to the ICU the next day. When patients deteriorate, their chances of dying increase at least threefold. Intervene today, and the mortality rate drops dramatically.

The current application runs every six hours and alerts the treating physicians which patients have more than an 8% probability of needing transfer to the ICU over the following 12 hours. In addition, the electronically stored patient healthcare preference documents are provided, allowing the physician to intervene immediately while respecting the individual’s personal choices. This approach has lowered both the mortality rate for these individuals and their average length of stay in the hospital.

Medicine is not an exact science. It requires decisions to be made based on best estimates of probabilities. Rarely can physicians be certain what to do, particularly for the sickest patients. Put every patient in the ICU who might deteriorate, and there would not be enough beds in any hospital in the country today. Wait 12 hours until you can be sure which patients will deteriorate, and you will be 12 hours late, with negative consequences. Soon the application described here will be updated on an hourly basis, which should further increase its value while lowering the morbidity and mortality for patients.

Combining the Best of Predictive Analytics and the Human Brain

What we know about the human brain is that it is exceptional, able to reach the right conclusion when the world’s best computers can’t. For example, a recent article compared the ability of physicians versus computers to diagnose medical problems. The clinicians were more accurate by a factor of two to one. But the human brain has blind spots. As an example, people can be biased as a consequence of their most recent experiences, whether good or bad. If a doctor’s last patient with a particular problem responded favorably to a particular medication, the doctor is likely to use it again, even though a single experience is statistically meaningless. Computers avoid these psychological errors, which makes the combination of human and machine so powerful.

Better Decision-Making in the ER

Dr. David Vinson’s and our Emergency Physician Research Network developed a computerized tool that calculates whether a patient in the emergency room should be admitted into the hospital or can go home safely on medication. It analyzes patient demographics, the full medical history including associated diseases and co-morbidities, and ED vital signs. This tool focuses on specific medical problems, such as small pulmonary emboli affecting only a relatively minor part of the lungs. Using this tool, physicians have been able to safely send home 50% more patients than before, with no increase in either morbidity or mortality.

Caring for the Smallest of Patients

The same type of analytics supports the neonatologists caring for the tiniest patients. An application created by Dr. Michael Kuzniewicz estimates the probability that a baby in one of our neonatal ICUs, who doesn’t seem very sick, would benefit from a full evaluation for systemic infection and treatment with antibiotics. The tool has succeeded both in identifying the newborns who would benefit from this extensive work-up and in warning about those for whom the testing would likely cause more harm than good. Once again, the ultimate decision resides with the clinical expert, but the information produced reduces error and improves clinical outcomes.

Pinpoint Cancer Care

Data analytics also enables oncologists to deliver more precise medical treatment to patients with cancer. An application labeled the Cancer Wizard matches an individual who needs chemotherapy with other patients in the database who have not only the same cancer but also the same personal demographics (age, gender and ethnicity). The computer program then updates the physician on which set of drugs were most effective, and the clinical outcomes patients achieved. As opposed to most published clinical trials, where the recommendations are one-size-fits-all, this application targets individuals to improve clinical outcomes. The tool is currently being used for patients with diffuse B-cell lymphoma and will next be applied to cancer of the pancreas and, eventually, multiple other malignancies.

How Harnessing Information Unleashes its Power

The advantages of these analytics tools are obvious. The extremely large size of the databases makes them highly reliable. Learn that your patient has even a relatively small risk of deteriorating—say, 12% rather than 8%—and you can be confident that by transferring the person to the ICU today, rather than tomorrow, you will decrease the chances of his dying. Identify patients about to go home from the ED as having a high probability of coming back, and you are more likely to admit them to the hospital, reducing the likelihood and frequency of complications. And tailor your treatment of patients with cancer based on their personal demographics, and you will select a better combination of medications.

Unlike many hyped technologies, each of these analytics tools will get better on its own as more patient outcomes are entered into databases and incorporated into the algorithms themselves. As computers continuously capture information from millions of electronic health records (EHRs), and researchers adjust the algorithms accordingly, the medical care recommendations will become ever more precise. And over time, researchers will apply these mathematical models to even more areas of clinical practice.

In Thinking, Fast and Slow, Nobel Prize winner Daniel Kahneman describes two ways the human brain works. The first is a form of pattern recognition that can combine disparate facts and reach a conclusion beyond what computers are likely to be able to do for many decades to come. The second is an analytics approach that requires what is, in essence, computation—mathematical calculations that translate data on patient outcomes into clinical algorithms. When it comes to computation and application of algorithms, computers are much faster and unlikely to make a mistake. And as the amount of available data grows through increased use of EHRs, computers will become even more accurate at determining the best treatment for specific patients—always with physicians making the decisions.

The combination of human and machine will be unbeatable in the future, with higher quality, improved performance and lower cost. Moving forward as quickly as possible, without being distracted by shiny new products, will prevent unnecessary complications and save lives.

This article originally appeared on Dr. Pearl’s column on