• Feature

    The Power of Predictive Analytics

    How patient data can answer questions to help PTs provide better care, improve outcomes, and lower costs.

    The Power of Predictive Analytics

    In 2017, Mike Nilsen, PT, DPT, determined that 40% of his patients with neck pain failed to progress or show clinically meaningful improvement.

    "I wasn't very happy with this," he says, "so I started performing chart reviews. I looked at the patients who didn't improve to see if I'd made any glaring mistakes. Did I not assess something? Did I not progress something?"

    Nilsen, a staff physical therapist (PT) with Intermountain Healthcare in West Valley City, Utah, reviewed APTA's clinical practice guidelines and Intermountain's care process models, looking for discrepancies between patients who did and did not improve. Although he confirmed that his treatment followed protocols and best practices, he also saw that he had better results when he progressed treatment a bit further. He identified a few small tweaks that could make a big difference.

    In 2018, Nilsen progressed treatments further. The failure-to-progress rate of his patients with neck issues dropped to 25%.

    The Big Picture

    The numbers Nilsen used are just a tiny fraction of the patient information and reported outcomes that Intermountain Healthcare has collected over the last 20 years, but they demonstrate in a big way the real-world power of data.

    Some people call it predictive analytics. Others call it probabilistic thinking. Either way, it's about using yesterday's patient data to help make informed care decisions today.

    As HealthITAnalytics.com explains in a September 4, 2018, article by Jennifer Bresnick, "As healthcare organizations develop more sophisticated big data analytics capabilities, they are beginning to move from basic descriptive analytics towards the realm of predictive insights…Instead of simply presenting information about past events to a user, predictive analytics estimate the likelihood of a future outcome based on patterns in the historical data."

    The report "Predictive Analytics in Health Care Trend Forecast," based on an annual survey conducted by the Society of Actuaries in 2019, concludes, "Predictive analytics is poised to reshape the health care industry by achieving the Triple Aim of improved patient outcomes, quality of care, and lower costs." Its survey of health payer executives and health provider executives found that 60% already are using predictive analytics within their organizations — a 13-point year-over-year increase from 2018. Sixty percent also expected to dedicate 15% or more of spending to predictive analytics during 2019.

    The top two desired outcomes cited by the executives for using predictive analytics are "reduced cost" (54%) and "patient satisfaction" (45%). That aligns with the top two actual results the companies are experiencing.

    Among health care uses for predictive analytics, as identified by the HealthITAnalytics.com article, are:

    • Avoiding 30-day hospital readmissions
    • Forestalling appointment no-shows
    • Predicting patient-utilization patterns
    • Developing precision medicine and new therapies
    • Bolstering patient engagement and satisfaction

    Better Care Means Better Outcomes

    At ATI Physical Therapy, Chris Stout, PsyD, PhD, uses data to optimize care provided by the company's 3,000 clinicians at its 900 locations. ATI has been collecting data since 1996 on a variety of metrics, including range of motion, pain, and activities of daily living. In 2015, the company incorporated standardized patient-reported outcome measures (PROMs) across all of its clinics to better measure patient outcomes and clinician performance. That led to development of a national patient registry approved by the National Institutes of Health's Library of Medicine.

    These days, ATI managers come to Stout — the company's vice president of clinical research and data analytics — with such questions as: How are our clinicians doing with a certain type of diagnosis? Where is the smallest level of change? Can we sort this information by state? Can we find the best-performing clinicians for a certain diagnosis?

    "Then we tap those top performers to be our mentors and experts in those areas," Stout says. "They help the folks who aren't quite up to that same level of performance. They help bring up their outcomes maybe a standard deviation or two."

    ATI's clinical excellence department also uses data from these reports to determine where to focus its continuing education efforts — creating custom online courses to meet particular needs.

    Three to six months after the training, Stout's team follows up with another report. "A manager will come to me and say, 'These 12 clinicians completed the shoulder course three months ago. Could you run the data on all the patients they had with shoulder diagnoses so we can see what their outcomes look like? And can you show that in comparison to the last quarter or last four quarters?'"

    Then Stout's team compares patient outcome measures pre- and post-course to determine if outcomes improved.

    How do the therapists feel about this oversight?

    "No one rolls out of bed wanting to do a poor job," Stout says. "Or even a mediocre job. People want to do better. They're excited to have this information to help them."

    Furthermore, he adds, ATI is in the process of introducing a new system that will provide each clinical director with his or her own dashboard. With just a few keystrokes, these directors will be able to sort the data themselves and compare regions, states, clinics, and clinicians against ATI benchmarks.

    The Data

    These reports may be useful, and many PTs would like to access this sort of information, but establishing benchmarks requires having enough data to compare — a clinically significant sample. Where and how is all this data input and stored?

    When Intermountain Healthcare first started collecting patient outcomes 20 years ago, Stephen Hunter, PT, DPT, explains, it used paper surveys with manual data entry. Now, says Hunter — the system's director of internal process control and a board-certified clinical specialist in orthopaedic physical therapy — the information is collected electronically and is automatically entered into an outcomes database.

    Stout describes ATI's custom-built electronic health records (EHR) database as "homegrown" to allow the collection of the information clinicians need and want. "It's the oxygen that allows me to breathe in the work that I do," he says. "It's elemental to my job."

    But what if you're at a small clinic or you work independently? Maybe you don't have a substantial store of records on hand. Can you still learn from patient data and EHRs?

    James Irrgang, PT, PhD, FAPTA, emphasizes to his students the importance of collecting outcomes even on a smaller scale. He is a professor in and chair of the Department of Physical Therapy at the University of Pittsburgh.

    During their year-long clinical internship, Irrgang's students collect data on treatment and outcomes for every patient. When the data is aggregated, the class has information for between 1,500 and 2,000 patients. The students then analyze and compare the patients' classifications, improvement, and outcomes.

    "We ask students to compare patients who had meaningful and important outcomes with those who did not," Irrgang says. "Then the students are asked if there were differences in adherence to clinical practice guidelines or other factors that may explain why one patient got better and another did not."

    Smaller clinics also can optimize the power of the data in their patient EHRs by working with a registry. APTA's Physical Therapy Outcomes Registry [Registry], notes Heather Smith, PT, MPH, collects and reports data for all participating practitioners. She is the association's director of quality. (The Registry will be the subject of a feature article in this magazine later this year. Find more information now at www.apta.org/Registry.)

    "Clients who report data to us get a snapshot of their performance," Smith says. "They get a dashboard where they can see how their performance measures up. They can look at the data by clinician, clinic, or site level, and then we provide internal benchmarks generated from all the providers who report that measure. So, those who participate can see how they perform against those benchmarks."

    According to Irrgang, who is past director of the Registry's scientific advisory panel, as of August 2019 it contained information on 35,000 patients with more than 200,000 tickets.

    The data usually is collected directly from de-identified EHRs. The information allows participants to learn as they treat, and to see what works and what doesn't in order to deliver better outcomes, Smith says.

    APTA is developing new measures to increase the power of the portfolio, she adds. Some of these measures will look at key metrics for certain populations, such as patients with Parkinson disease or those who have undergone total joint replacements, and tie them to clinical practice guidelines to provide information therapists can use when working with patients.

    "It's pretty exciting, because Registry participants are getting fairly real-time data on their performance and can see beyond how their treatment affected a single patient," Smith says. "They can see how they're doing compared with everyone who treats low back pain, or how well they're performing preventive activities such as screening for falls in their patient population. It gives us the ability to learn as we treat."

    Even organizations with their own large data sources are excited by the promise of what PTs can learn from the pooled data of the Registry.

    "It could be helpful to get an even bigger pool to draw from to fine-tune our standard of care," Nilsen says. "Maybe it's normal, for example, for 25% of patients with low back pain to not improve. But if all clinicians who are doing similar work pool their data, then we can see the average. We can see if there is a way that other clinicians or medical groups have used this data to modify treatments that helped these patients."

    Keeping clinicians up-to-date on the latest research and best practices can be especially challenging. Julie Fritz, PT, PhD, FAPTA, publishes her work but has no assurances that practitioners are providing care that's consistently aligned with the latest evidence. She hopes predictive analytics will help change that.

    "When care is more adherent to the principles and the science, you achieve better patient outcomes," Fritz says. "This should surprise no one. But often care doesn't align ideally with what we know we ought to be doing. That's been another area of work I've been interested in: trying to make sure that providers — physical therapists specifically — provide care that's consistently related to and aligned with the best evidence we have." Fritz is a distinguished professor at the University of Utah.

    Clinic Management

    Not only does Kurt Gengenbacher, PT, DPT, see these metrics as powerful tools for determining the course of care, he also says they can be helpful for managing clinic tasks such as scheduling appointments.

    "Say there's a therapist who has 20 patients on the caseload, and 10 of them are not scheduled out," says Gengenbacher, senior director of clinical excellence at ATI. "We can look at the outcomes of those 10 patients. That often opens a conversation with the therapist. It often shows us that those patients are not really seeing the value of therapy. We then can educate clinicians on why some things need to be done, and how they can be done better, to try to get buy-in from the patient."

    With this renewed commitment, he says, patients often start seeing more value in their care. Clinicians see more improvement, and this leads to better outcomes.

    Beyond appointment scheduling and care management, researchers at Duke University used their medical center's EHRs and a predictive analytics model to accurately identify 4,800 patient no-shows in one year. The discovery, written up in a study published in the August 2018 issue of Journal of the American Medical Informatics Association, allowed them to save costs and improve clinician workflow.

    A similar study at the University of Texas Southwestern found that researchers could use predictive analytics to identify patients who, because they had certain problems or adverse events while in the hospital — such as a Clostridioides difficile infection or vital sign instability — were more likely to be readmitted in the following 30 days. The study appeared in the February 29, 2016, issue of the Journal of Hospital Medicine.

    Stout points out that having a standard tool of measurement and data also can be helpful to practitioners who want to work with other providers, payers, and insurance companies.

    "Let's say that a patient has knee replacement surgery," Stout says. "She might be given the knee patient outcome survey for joint replacement preoperatively and before discharge. Then, if she was referred to us, we could use that same instrument for her postsurgical outpatient physical therapy. The nice thing is that the surgeon, the payer, the PT, and the policymaker can all look at that patient throughout that course of care and treatment, and literally have the same instrument given to them."

    And, using that instrument, the whole care team can determine the best course of care for each patient from beginning to end.

    "We certainly can go to payers with this data," Stephen Hunter says, "because we have outcomes every visit. We can go to a payer and say, 'Look, you're only approving us for six visits for low back pain. But if you look at our data from thousands of patients, we can make a bigger difference if we get eight visits.' And most payers know that if patients don't improve in physical therapy, they're going to require more expensive care."

    The Road Ahead

    Stout's team plans to roll out risk adjustment features this year to help account for patient characteristics such as age, fitness level, body mass index, and illnesses when determining patient outcomes. This should help standardize the level of expectation for each PT even if one happens always to see a particular type of patient.

    Irrgang agrees that these risk adjustments are an important next step to maximizing the promise of probabilistic thinking. "With risk adjustments, we can create an algorithm into which the clinician could plug certain variables — such as sex, age, and baseline scores. We can learn what to expect the course of care to be — maybe three weeks and six visits — and what outcome realistically can be achieved."

    Hunter sees the future of predictive analytics as an EHR system that prompts PTs. It will use artificial intelligence to ask questions and make suggestions — helping the PT optimize the course of treatment.

    "Suppose a therapist does an evaluation and the patient starts treatment," Hunter says. "We know from our data that patient recovery should have a certain trajectory. The patient should get better at a certain rate based on hundreds of similar patients. But after three visits the patient hasn't achieved that improvement rate. On visit four, the EHR might suggest a change based on the other records it's analyzed."

    "In the future we'll be able to look at the arc of care for each patient," Stout says. "We'll be able to answer such questions as, 'Do we need to speed up the number of visits during the course of a week?' 'Do we need to tweak what goes on during those visits?' 'How can we know before the patient is discharged if we need to tweak something or do something a little different?'"

    Hunter imagines an artificial intelligence system similar to the one Intermountain has been using for years to help practitioners identify harmful drug interactions. The patient's medications are entered into the EHR and, when a new medication is prescribed, the system alerts physicians to any potentially harmful interactions.

    "I see us doing that with physical therapy," Hunter says. "You do an evaluation and the EHR tells you whether it's complete. It's like a guideline that you can to give the therapist. If you're seeing 15 patients and they all have different complexities, it can be difficult to know everything about every patient. But artificial intelligence could help you do that if you're putting that data in."

    Katherine Malmo is a freelance writer.

    The Darker Side of Predictive Analytics

    Predictive analytics are just algorithms — mathematical formulas. Yet, as observers both inside and outside health care have pointed out and demonstrated, bias can creep in — changing what should be neutral predictions into tools that can promote bias and inequity.

    Peter Kochenburger, an expert on insurance law, cites the development and use of predictive models that purport to establish a "propensity for fraud," especially when based on information or characteristics unrelated to a particular claim.

    In a November 2018 article for Insurance Business magazine titled "The Dark Side of Big Data," he writes: "This data can include comprehensive arrest — not just conviction — records, credit and bill-paying habits well beyond traditional credit reports, social media use or non-use, and shopping habits. These models may be used to screen applicants for this propensity at the underwriting stage, before the possibility of a claim could have arisen."

    Even if the data is accurate, Kochenburger points out, "all models contain preconceptions, including decisions on what data to use and what to omit, that can reflect improper biases. For example, using criminal records as an underwriting tool could ignore the fact that policing and prosecutorial decisions might not be race-neutral." He adds, "The more this type of non-claim-related information is used for insurance purposes, the more likely it is for implicit biases to creep in."

    Bias is occurring in health care as well, according to a number of studies. One, published October 25, 2019, by Obermeyer and colleagues in Science magazine, found that predictive analytics algorithms are referring healthier white patients to care-management programs at higher rates than they are referring less-healthy black patients to those same programs. The problem is that the algorithm calculates patient risk by assessing the amount of health care dollars spent on the patient. However, health disparities are favoring white patients who in many cases are healthier.

    That study explains, "Bias occurs because the algorithm uses health costs as a proxy for health needs. Less money is spent on black patients who have the same level of need, and the algorithm thus falsely concludes that black patients are healthier than equally sick white patients…The bias arises because the algorithm predicts health care costs rather than illness, but unequal access to care means that we spend less money caring for black patients than for white patients." The researchers concluded that eliminating costs as a proxy for needs in determining admission to the care management program would eliminate the bias, and would increase the percentage of black patients receiving additional help from 17.78% to 46.5%.

    There also is risk in sifting through big data looking for relationships that may exist mathematically but are predictive of nothing. Austin Frakt and Steven Pizer, writing in the February 16, 2016, issue of The American Journal of Managed Care, give this example:

    For every 5 million packages of x-ray contrast media distributed to healthcare facilities, about 6 individuals die from adverse effects. With big data, we learn that such deaths are highly correlated with electrical engineering doctorates awarded, precipitation in Nebraska, and per capita mozzarella cheese consumption (correlations 0.75, 0.85, and 0.74, respectively).

    However, because we cannot conceive of a causal mechanism, it is obvious that these variables play no causal role in x-ray contrast media deaths. That such high correlations can be easily mined from big data is concerning nonetheless because it is not always trivial to assess whether they are telling us something useful. For example, observational data reveal that proton pump inhibitor (PPI) use is associated with pneumonia incidence. This could be causal because a mechanism is plausible — gastric acid reduction could increase bacterial colonization — but perhaps the association arises because other factors drive both PPI use and pneumonia incidence.

    Ziad Obermeyer, acting associate professor of health policy and management at the University of California-Berkeley and lead author of the paper, writes, "Algorithms can do terrible things, or algorithms can do wonderful things. Which one of those things they do is basically up to us. We make so many choices when we train an algorithm that feel technical and small. But these choices make the difference between an algorithm that's good or bad, biases or unbiased. So it's often very understandable when we end up with algorithms that don't do what we want them to do, because those choices are hard."

    (You can find more unlikely correlations from Spurious Correlations at tylervigen.com/spurious-correlations.)


    Katherine, fantastic article. Documenting patient outcomes is crucial to our PT profession to promoting how we can impact lives and get people back to function. You make so many good points regarding outcomes data, such as having a sample size so you can make predictions and conclusions, national comparisons, and risk adjusted factors for more accurate predicted scores. Although minimally important difference(MCD) and minimally clinically important difference (MCID) are valuable metrics, a risk adjusted predicted change score and number of visits adds another layer of value to establish specific targets for outcomes. I have been using such a database since 2009, it is called Focus On Therapeutic Outcomes (FOTO), and it has really changed my clinical practice.
    Posted by Kristen Brinks on 2/3/2020 6:06:21 PM

    Leave a comment: