Hello. My name is Joseph Zabinski. I'm Director of Data Solutions at OM1 in Boston, and today I'm going to be speaking to you about some work we've been doing in improving risk management in spine surgery patients through cost burden prediction. So by way of background or motivation for this work, in recent years value-based care programs have been becoming increasingly important as hospitals and health systems try to manage populations of patients that are undergoing similar procedures. And one of the key characteristics of these programs that is different from more traditional or more familiar fee-for-service programs is that while fee-for-service programs reimburse the health system in a sort of one-to-one matching way for each service provided, value-based care programs provide reimbursement for episodes of care as a whole. And for this reason, risk management for the patient population under undergoing surgery or receiving care is especially important both to optimize clinical outcomes and also to keep track of resource utilization for the entire program. One particular sensitivity that some of these value-based care programs have is to a few patients who end up with unusual and highly disproportionate resource utilization. Because there's a sort of leveled amount of reimbursement for each case within these programs, those few patients who have very disproportionate resource utilization can change the overall financial performance of the program quite drastically. And so for this reason, it's important to consider preoperative risk stratification of patients in value-based care settings, especially for patients who are undergoing elective spine surgeries, both because these spine surgeries are fairly expensive in an absolute sense and also because they are characterized by substantial variation in total cost. In other words, if you have two patients who are undergoing the same spine surgery, they can end up having very different experiences and very different levels of post-surgical resource utilization in their episodes of care. And from a value-based care setting, that's very important with respect to the overall performance of patients in that program for the health system where the care is occurring. So a number of programs have been designed around this concept of value-based care. Some of these programs are called bundled payment programs. One such program is CMS's Bundled Payments for Care Improvement, or BPCI, initiative. And what these programs try to do is to incentivize value-based care by providing set target payments for surgical care episodes. And so in other words, instead of reimbursing for each specific procedure or each specific element of the care episode, these bundled payment programs provide a single payment that covers the surgery itself as well as a pre-specified post-surgical recovery period. One of the most relevant concepts in these bundled payment programs is the idea of bundle breaking. And that occurs when a care episode ends up costing the hospital or health system more than they receive as reimbursement for that episode. And so predicting which patients are at greatest risk of breaking the bundle is crucial for health systems that are trying to succeed in these programs. In particular, predicting that risk using only preoperative information on the patient that would be available to their surgeon and their care team could help the surgeon and care team to target risk mitigation efforts before the surgery actually occurs and hopefully reduce the risk of negative post-surgical clinical complications as well as the risk of breaking the bundle. And so we sought to develop a predictive algorithm that would do exactly this. It would use preoperative information that would be available to the surgeon and the care team for a spine surgery patient in order to try to predict their risk of being a bundle breaker or their more general level of resource utilization post-surgically. So in order to do this, the foundation of course of any predictive model is the data. And we used our proprietary data set, the OM1 Real-World Data Cloud, which contains de-identified records on hundreds of millions of patients. Within this data set, we defined a cohort of de-identified patients who underwent common spine procedures like spinal fusions during the years between 2013 and 2019. And after applying the qualifiers that we designed to generate this cohort, we ended up with a little shy of about 600,000 patients who had experienced one of these qualifying spine procedures during that period of time. And in addition to knowing that these patients had undergone a spine procedure, we also had available to us a year of those patients' medical history data prior to their spine surgery for each patient. So in other words, for each patient, we indexed the date of their spine surgery and then looked one year into their prior medical history in order to give the algorithm some data to be used to try to predict post-surgical outcomes. So as is standard practice when developing any kind of AI predictive algorithm, we partitioned this data set cohort into testing and training subsets. And we use that training subset to train an artificial intelligence algorithm to predict patients' post-surgical resource utilization using that single year of their pre-surgical medical history data. And the output of the model in generating this prediction took the form of a preoperative risk score, which we call the OMB Spine Score. One of these was generated for each patient, ranging from one, which indicated the lowest level of risk, up to a score of 1,000, which indicated the highest level of risk. And so the question was really when we were using the testing data subset here to evaluate the predictive performance of this algorithm, what we were looking at was the association between these scores, preoperative scores generated for patients that ranged from 1 to 1,000, and then their post-operative resource utilization. So as we were hoping and expecting, the preoperative risk score, these OMB Spine Scores that we generated were, in fact, highly predictive of future resource utilization. And in particular, the preoperative risk score was useful for identifying patients who ended up being the highest utilizers of resources after their surgery. One way that we looked at quantifying this predictive capacity was to see how well the model did in identifying patients who ended up in the top 1% of utilizers. So in other words, if you imagine patients kind of distributed in terms of increasing post-surgical resource utilization, these are the patients all the way at the right-hand side of that distribution. And when we targeted the model and used these OMB Spine Scores to try to evaluate that outcome, we ended up with an area under the receiver operating characteristic curve of 0.82, which if you're familiar with that metric for evaluating predictive models, you'll know that this is quite a good level of performance, especially in a healthcare setting. Even more excitingly for us, we found that this preoperative risk score that we generated was also highly predictive of post-operative bundle breaking more generally. And I've tried to illustrate that in the graph that you see here. So each of these bars represents an episode of care from the data set analyzed. And each of the bars representing an episode of care is also labeled with the preoperative OMB Spine Risk Score that was generated for that patient using their pre-surgical information. And as you can see, the bars are organized from lowest score on the left-hand side of the graph all the way up to highest score on the right-hand side of the graph. The colors indicate bundle breaking in the case of the red and non-bundle breaking in the case of the blue. And as you can see, there's quite a clear association between lower risk patients as assessed by this score, those over on the left-hand side of the graph, and much lower risk of bundle breaking. And then also a much higher risk, much higher prevalence of bundle breaking at the right-hand side of the graph where patients had the highest OMB Spine Scores. You also see a clustering of sort of the most severe cases of bundle breaking. Those are the largest red bars over at the right-hand side of the graph, indicating that these scores are useful for identifying not just bundle breakers, but also the riskiest cases of bundle breaking that ended up with the greatest bundle breaking magnitude. So to conclude, as hoped, the score that we developed did have strong predictive performance and potential application for presurgical risk stratification of spine surgery patients. Since examining those components of performance, we've also developed the ability for the system to surface key risk factors in patients' medical histories in order to allow for better targeting of preoperative resources. And this whole system is currently being studied in clinical spine surgery practice at a major Midwestern academic medical center.

risk management

cost burden prediction

spine surgery patients

value-based care programs

preoperative risk stratification