This article is a portion of a book titled "Challenges, Risks and Opportunities in Today's Spine World" edited by Stephen Hochschuler, MD, Frank Phillips, MD, and Richard Fessler, MD. You can find links to the previous chapters at the end of this article.
As an emergency physician, I see multiple patients complaining of severe back pain. Many of these patients have had prior back surgery(s) for their pain, and many of those who haven’t yet had surgery are hoping I can help them get a spine surgeon to operate on them. Having had two very successful lumbar spine surgeries myself, I understand why patients are eager to have a procedure that can potentially improve or eliminate their pain. But the reality is, depending on the procedure, 20-30 percent of spine surgery patients do not have positive outcomes, and these numbers get worse with each subsequent operation. Add to this the relatively high rate of intra-operative and post-operative complications for spine surgeries, and the significant impact on cost of care for health systems spinal surgeries create, and it quickly becomes obvious why surgeons, health systems and patients are all looking for strategies to better identify appropriate surgical candidates, stratify risks, and improve outcomes. The use of Predictive Analytics to improve pre-operative planning, increase patient understanding of the risks and benefits of specific procedures, and reduce the risk of complications has rapidly moved to the forefront as a means of addressing these important issues.
Wouldn’t it be great if we could accurately predict, preoperatively, what the outcomes would be post-operatively for spine surgeries? We could use this information to decide who to operate on, what procedures to use, what the optimal post-operative care pathway should be, and set realistic expectations for our patients. Surgeons could use this information to select the procedure with the best risk/benefit ratio for individual patients, or to recommend non-operative care, and patients could participate in the decision making process in a meaningful way. Unfortunately, this level of prediction turns out to be really difficult to achieve. Humans are very complex creatures, and increasingly we are realizing that mental health and social factors may have as much or more impact on outcomes and patient satisfaction as physical health. Chronic health conditions such as diabetes, heart disease, osteoporosis, arthritis, cancer, and COPD significantly impact the risk of intra-operative and post-operative complications, as well as the ability to effectively participate in rehabilitation. Drug use, including prescription drugs such as opioid pain medications, compounds the problem, with recent evidence showing patients who are using opioid pain medications pre-operatively do worse post-operatively in terms of patient satisfaction with the surgery. And we are just scratching the surface on the impact of genetics data in surgical risks and outcomes. There is so much to take into account, and so many different data points that need to be considered in the development of predictive analytics algorithms, coming from so many different sources, and in different formats, that building and automating predictive algorithms that can provide intelligence in real-time and be presented in a format that is easily understood by surgeons and patients significantly exceeds the capabilities of traditional analytics software.
Predictive analytics is not a new idea for healthcare. Risk scoring systems, based on population derived data, have been used for many years to help surgeons and patients make more informed choices. However, tools that are based on population statistics, and lack patient specific data, are inadequate to define risk for individual patients. As access to larger databases grows, the opportunity to leverage this data to develop more accurate predictions of risk and to choose the best care pathway for an individual patient also increases. In order to have a significant impact on outcomes, a predictive analytics solution must have the following characteristics: be able to use all available patient data, be transparent in how risk is calculated, perform in real-time, allow providers to document in their usual format, be easily used by physicians and understood by patients, and be patient specific.
Big data holds a lot of promise, but analytics tools such as machine learning are only as good as the data and have many limitations that must be understood by the users of predictions generated by this technology. Machine learning is an application of artificial intelligence allowing systems to automatically learn and improve from experience without explicit human programming. Machine learning focuses on the development of new computer programs that can access data and use it to learn for themselves, looking for patterns in data and making better decisions in the future. The primary aim is to enable the computers to learn automatically without human intervention.
Like predictive analytics, machine learning is by no means new, having been around since the 1960s. However, recent advances in machine learning technologies have created much more powerful and faster processes, and the widespread adoption of electronic health records has provided access to huge amounts of data for the machine learning technologies to utilize in making predictions.
The use of artificial intelligence models to develop predictive algorithms in healthcare has grown rapidly in recent years. Rapid advancements in machine-learning techniques such “neural networks” and, more recently, “deep learning”, combined with access to very large data sets, are believed by many to be the solution to the challenge of needing real-time, patient specific intelligence for clinical decision making available at the point of care. The synthesis of big data and powerful artificial intelligence models can reveal data trends and identify variables that statisticians might never observe.
Deep learning is a class of machine learning methods based on learning data representations, as opposed to task-specific algorithms. Deep learning involves:
• Using a cascade of multiple layers of nonlinear processing units, with each successive layer using the output from the previous layer as input
• Learning in supervised (classification) and/or unsupervised (pattern analysis) modes
• Learning multiple levels of representations that correspond to different levels of abstraction; the levels form a hierarchy of concepts.
Surgeons (and patients) don’t need to fully understand all the technical aspects of machine learning. They should, however, understand that while machine learning has great promise, it also currently has some significant limitations. In order to develop highly accurate, patient specific predictive intelligence, data elements found in textual format must be combined with structured data in order to have a complete understanding of the patient. Most machine learning solutions lack the ability to combine structured and unstructured data elements, and use only one or the other. Natural language processing technology helps interpret textual data content, but generally lacks the ability to understand context, much less more complex concepts such as temporal relationships of data that can be critical in medicine. Machine learning is most commonly used in the setting of an enterprise data warehouse, where data is gathered and prepared for analysis, but not in a real-time fashion. There is also the risk that machine learning algorithms, because they look at such massive data sets, may include potential risk factors that are very rare, and not relevant to the general population. The algorithms may also be developed from a data set that has patient demographics that differ significantly from the population where it is being applied, missing key factors unique to that population. Machine learning algorithms are also “black box” solutions, meaning you can see the output, but you don’t know how that output was actually determined. I have encountered a machine learning algorithm that identified having asthma as a protective factor for outcomes in pediatric pneumonia, when it was actually the fact that in this particular study population, all pediatric asthma patients who developed pneumonia were admitted to the ICU and treated with different, more aggressive care than non-asthmatics. This accounted for the improved outcomes. Because of these limitations, many physicians are hesitant to trust machine learning predictive algorithms at this point.
The good news is that new, innovative predictive analytics solutions are currently being developed and introduced that address most, if not all, of the above limitations of machine learning. Advanced text analytics solutions are now becoming available that understand data content and context, temporal relationships of data across episodes of care, and finding of vs. no finding of. This allows more of the data to be included, and significantly improves the accuracy of predictions. Social determinates, such as: is the patient an active drug user, do they truly live alone or do they have family nearby that are involved in their care, and do they have transportation, can now be included in the risk equation. This textual data can now also be combined with structured data and used by deep learning programs to improve existing algorithms, to rapidly create new predictive algorithms from small, local data sets that are population and patient specific. As these new technologies are refined and introduced into the market, we should expect to finally have the types of predictive analytics tools that can truly be useful to both surgeons and patients.
Book: Challenges, Risks and Opportunities in Today’s Spine World
Chapter: Predictive Analytics and Machine Learning
Writer: Gregory D. Hobbs MD MPH, OpisoftCare, Inc.
Previous chapters:
Challenges, risks and opportunities in today's spine world
Spine care - Balancing cost with innovation
What are big data and predictive analytics