I'll admit it ... I had never even heard of "predictive analytics" while practicing medicine or even in the decade after leaving clinical practice.
Having spent years in both academic and community hospital operating rooms, intensive care units, patient wards, and outpatient clinics, I was among the majority of practicing physicians and non-physician inpatient bedside providers who had never seen, let alone utilized, patient-specific "predictions.
Nor did this bother me. After all, isn't prediction the overall job description for every physician? Haven't we spent years in training, learning the signs and symptoms and myriad of other clinical factors that predict "what's going on" with each of our patients, that generate a differential diagnosis list, that guide our diagnostic work-up and, ultimately, that drive us to the correct diagnosis? The answer is, yes. And no.
Yes, as physicians, we are trained and gain increasing experience in incorporating patient signs and symptoms and lab results and imaging studies into our own "predictive analytics" system: our brains. And yes, this often works well in predicting and confirming a patient diagnosis or treatment outcome.
However, we are far less successful in accurately predicting individual risks for specific inpatient health and care events, such as predicting each of our inpatients' risk of sepsis or risk of mortality or risk of transfer to an intensive care unit (ICU) during hospitalization, the risk of an extended length of stay, and similar inpatient risks that have dramatic impact on the patient, the provider, and the payer.
Thus, while we can diagnose sepsis (that is, predict based on initial clinical information that a patient is indeed septic), we are significantly less successful in predicting who is at risk of becoming septic during their hospitalization and, more specifically, at what level of sepsis risk any individual patient truly is and how that sepsis risk is changing at any given time for any individual patient.
As a result, sepsis remains a major driver of morbidity and mortality for thousands of inpatients annually, as physicians correctly predicting that a patient has sepsis are already "too late" to intervene and prevent the patient's journey to sepsis. The result for septic patients is increased hospital stays, additional organ insult and injury, and, for many, in-hospital death. For providers, the inability to routinely stratify sepsis risk and appreciate risk trending represents a major challenge in the delivery of quality healthcare. And for payers, simply waiting for members to become septic and then paying the high cost of care associated exemplifies poor cost efficiency.
"Risk prediction" is where predictive analytics enters the clinical equation. Complex computer algorithms are capable of incorporating and weighing multitudes of potentially meaningful clinical factors, resulting in individual patient risk stratification and trending information.
Such reliable, personal risk stratification and trending predictions are powerful tools in guiding physicians and other inpatient bedside providers, empowering caregivers to focus clinical attention and resources on those specific inpatients who are truly at higher or growing risk in order to intervene and prevent an unwanted clinical event or outcome.
As the medical community has recognized, implementing preventative processes can be enormously beneficial to patients, caregivers, and payers (the use of rapid response teams is a superb example of the benefits of preventative intervention).
Let's take a real-world example of such preventative intervention based on predictive analytics:
In a busy cardiac facility, trying to determine which inpatients are at greatest or growing risk of clinical decompensation, which will require transfer to the ICU, is extremely challenging. Transferring a floor patient to the ICU increases inpatient length of stay (LOS), requires additional hospital staff and resources, drives up overall cost of care, and (most importantly) worsens patient outcome.
On most cardiology wards, physicians (cardiologists and hospitalists) round in the morning and evening, rapidly assessing each of the often numerous inpatients. Based on a handful of lab values, vital signs, and cursory physical examinations, these physicians get a general sense of how each inpatient is doing. In more communicative settings, the physicians may share their concerns about specific inpatients with an individual nurse. "Watch Mr. Smith...he's not looking so good today," is an example of such communication.
Now imagine those cardiologists and hospitalists, and the cardiology unit nurse managers, charge nurses, and case managers were provided on their iPad or other mobile device real-time, regularly updated (each shift, for example) inpatient specific ICU transfer risk stratification and risk trending.
Instead of the physicians or, at best, the physicians and a few nurses, knowing who's "not looking so good today," and instead of this generalized physician prediction, the caregivers were all provided risk specific stratification and trending, all those providing care for the cardiology inpatients would know on which patients to focus their clinical attention and resources.
Every stakeholder would greatly benefit. Staff and resources availability for those inpatients truly at-risk would increase, allowing for targeted evaluation and intervention to address the clinical processes driving the risk of ICU transfer. Providers, too, would realize greater success and satisfaction in knowing on whom they should focus their attention, and in "wasting less time" evaluating patients who are not truly in need of such attention.
Hospitals and payers would benefit from improved quality of care and reduced cost of care
This is just one example of how predictive analytics can truly be personalized (patient specific) and actionable, resulting in increased value, via both increased quality and reduced costs. And while predictive analytics is initially focused on those risks that play a critical role in reimbursement and healthcare reform (such as the risk of extended LOS, 30-day readmission, sepsis, mortality, etc.), predictive models can be created for countless additional clinical events and outcomes for which real-time electronic clinical data is available.
So here's a common question: if predictive analytics can provide clinical providers with inpatient-specific risk stratification and trending predictions, why can't these tools simply "tell" clinicians what clinical processes and factors are driving that risk?
That is, in addition to reporting that "Mr. Johnson has trended into the 'high sepsis risk" category," can't predictive analytics also explain that this trend is the result of Mr. Johnson's rapidly-developing Klebsiella urinary tract infection?
Unfortunately, it's not quite that simple...yet. There are often hundreds of dynamic clinical results and factors (every changing vital sign and other biometrics, the numerous changing laboratory results, number and duration of antibiotic use, additional pharmaceutical use data, updated imaging study results, microbiology results, procedures, and so on and so on), which drive predictive analytics risk stratification and trending outcomes. Thus, the "drivers" of risk predictions are rarely truly "clinically meaningful."
However, predictive analytics will soon penetrate into the world of clinical utility. Initially, predictive analytics tools will likely provide a ranking of the "clinical drivers" that most greatly drive each risk prediction, in much the same way as cost utilization prediction drivers are presented today.
Such predictive analytics would thus not only say that "Mr. Johnson has trended into the 'high sepsis risk" category," they would additionally share that "Mr. Johnson's trending sepsis risk is based 23 percent on cardiovascular (pulse and blood pressure) factors, 17 percent on additional antibiotic utilization, and 11 percent on altered intake/output fluid balance." Thus, predictive analytics will not only identify each inpatient's individual risk level and risk trending, these solutions will also provide the clinical drivers of each risk prediction for each patient, further empowering bedside clinical providers to focus their attention and resources on truly at-risk inpatients, allowing for proactive intervention rather than reactive care once an undesired event or outcome has occurred.
Peter Edelstein, MD, is chief medical officer of Elsevier/MEDai, a provider of analytics solutions. Edelstein was in private practice before serving on the Surgical Faculty at Stanford University. Before joining Elsevier/MEDai, Edelstein most recently served as a physician executive within the Adventist Health System.