How can we help millions of chronically ill patients lead healthier lives with fewer hospitalizations? Parsing through individual patient remote monitoring data doesn’t seem like a scalable solution. But what if artificial intelligence could rapidly analyze this information and predict an individual’s health problem, such as worsening heart failure, in time for caregivers to intervene and prevent an emergency department visit or hospital admission?

That may sound like a pipe dream, yet the technology to make it happen already exists. This new “prescription” for chronically ill patients would be an algorithm — a personalized disease model unique to each patient, based on streams of data collected from biosensors worn by patients. Aided by advances in clinician-directed machine learning, over time these algorithms could be tuned to predict short-term risk of deteriorating conditions before an acute event requires hospitalization. 

Needless to say, the implications of such a technological advance would be tremendous.

Early remote monitoring

While personalized medicine is a fairly recent development, monitoring patients in their own homes is not. Different home monitoring programs have been launched over the past 20 years or so, and while some have shown some value in reducing avoidable hospitalizations, most of them have not sustained sufficient clinical and economic value. By and large, these efforts have been limited because of overly simplified assumptions and other errors. 

For example, in one program, patients with congestive heart failure would step on a scale every day, and there would be an alert if there was a one-day weight increase of a fixed amount (say 2 or 3 pounds). All patients with CHF would be monitored the same way. However, because abrupt weight gain occurs at a late stage of deterioration, the alert sometimes came too late to change the outcome; it also delivered too many false alarms. 

By the same token, as pathologic progression leading to severe symptoms may start days before the patient becomes aware of it, there were missed opportunities for earlier, more successful intervention. In addition, other chronic diseases weren’t monitored, which meant that exacerbations of comorbidities went undetected. This resulted in more missed opportunities to help patients. 

First-generation monitoring systems also typically chose patients at high risk for readmission. Yet, readmissions are just the tail end of the single most significant cause of excessive health care spending: unnecessary and preventable admissions. Broadening our objective to reduce this larger category can vastly improve the quality of life for more patients while managing chronic illness at a far lower cost. 

Improvements in machine learning

Recent market advancements in machine learning (artificial intelligence) and the devices used for remote patient monitoring have now made it possible to transcend the limitations of first-generation monitoring. The first reason for this is the remarkable decline in the cost of wearable devices, which has plummeted to as little as $100 for a stick-on biosensor. Second, many of the new devices are multiparameter tools and can collect data passively, adding to their usefulness and ease of use. 

Third, advances in computing now make it possible for clinicians, unaided by information technology experts, to easily create initial sets of monitoring rules (algorithms). These rules are automatically personalized for each patient and can detect health deterioration early, in effect predicting hospitalization days in advance with high accuracy. 

Fourth, advances in machine learning provide continual improvements, leading to increasingly accurate models that detect deterioration earlier. This last point is accomplished thanks to a novel concept, adaptive physiological modeling, that lets the system learn from patient data and case manager feedback. For instance, if the system triggers a false-positive alert, machine learning techniques recommend adaptations for clinicians to correct the problem. 

Remote intelligence 

These new advances have enabled a second generation of remote patient monitoring: remote patient intelligence. This approach has recently been deployed across the country by some of the nation’s best-known health plans and health care organizations, including Anthem and Cedars-Sinai in Los Angeles. 

Early results indicate the technology has enormous potential for mainstream adoption. In one such study, 6,314 congestive heart failure patients were enrolled in a managed care program and provided a wireless blood pressure monitor for home measurement and subsequent data analysis. Commonly used rule-of-thumb algorithms that compare a patient’s data with population norms were evaluated against a new set of personalized rules. At study’s end, for a five-day prediction window, the rule-of-thumb algorithm was able to detect deterioration leading to hospitalization with only 38.7 percent accuracy compared with 66.2 percent for the personalized rules. 

Such results indicate that caregivers will soon have an effective tool to reliably provide early warning of impending hospitalization for a specific patient. Enabling clinicians to write the rules for these tools and allowing clinician-directed machine learning to continually improve the rules is a seismic shift. No longer simply programs that perform basic home monitoring, these advances are a new-generation platform that delivers remote — and reliable — patient intelligence. 

Dean Sawyer is the CEO of Sentrian, a remote patient intelligence company in Aliso Viejo, Calif.