Skip to content

Pioneering Machine Learning Tools to Inform Decision Making in Spine Surgeries

In This Article

  • The Orthopaedic Department of Massachusetts General Hospital leads the innovative use of probabilities and prediction models to improve decision-making to help patients understand the risks and outcomes of spine surgeries
  • People mistakenly fear that the rise of ML (machine learning) and AI (artificial intelligence) will supplant the physician-patient relationship. Research suggests otherwise
  • Decision-making is the basis of optimal clinical and surgical outcomes, and ML, AI-based tools are a potentially valuable adjunct to help guide decisions

In the traditional clinical consult, a physician makes decisions based on an understanding of probability—and in particular, the probability that she or he can improve the patient's outcome.

However, despite the most sophisticated command of probabilities, cognitive science research shows that people are prone to making decisions based on faulty information or assumptions.

"Most human decisions are based on ideas that are frankly wrong," says Joseph H. Schwab, MD, chief of the Orthopaedic Spine Center at Massachusetts General Hospital, and co-director of the Stephan L. Harris Center for Chordoma Care. Factors such as memory, cognitive bias, limited experience and emotional filtering can influence a person's decisions.

"Yet decision-making is the foundation of medicine," Dr. Schwab says. He and colleagues are at the forefront of developing machine learning (ML) algorithms to help patients understand the risks and outcomes of their surgeries, and to mitigate inborn human biases in decision making.

Probability Modeling for Surgical Outcomes

In the last five years, the rise of ML and its advanced mathematics that drive artificial intelligence (AI) have accelerated dramatically. The confluence of these trends with the development of graphic processing unit (GPU) computer chips from the gaming industry means computers can now parallel process large amounts of data at speeds never before possible.

Dr. Schwab and his team have used this technology to develop a probability model of surgery outcomes to inform treatment choices, and are developing another model to assess the probability that patients will experience complications. The models are relatively broad, with some used to predict which factors are associated with long-term opiate usage after surgery, to others that predict which patients are most likely to have a major complication after surgery.

“The models are developed by comparing several different machine learning algorithms,” says Dr. Schwab. “We compare the ability of the algorithms to discern differences between two groups, and then we compare the accuracy of the models over a range. We select the best algorithm for each case on an individual basis.”

While Dr. Schwab and his team have used nationally available data to design algorithms, they prefer to use the extremely large spine database that includes 15,000 patients belonging to Mass General and affiliated hospitals through Partners HealthCare. It includes many more parameters than the national databases, such as those from Medicare and the National Surgery Quality Improvement Program (NSQIP) collection of inpatient data.

The Partners data provides the level of granularity missing in the government resources, and can help predict such practical outcomes as identifying which patients should go to a skilled nursing facility and the probability of the risk for complications.

“The real benefit of using our data is that it provides the ability to follow patients longitudinally. It is not limited by pre-selected data points, such as the ones in national databases,” says Dr. Schwab. He also points out that the Partners data can be linked to genetic data via the biobank. Patient-reported outcomes measures (PROMS) are soon to be available for research efforts.

Advantages of Machine Learning

People often fear the application of machine learning in medicine as a way of supplanting the physician-patient relationship, and the many aspects of intimate knowledge that embodies. Dr. Schwab believes these concerns can be overcome by the clear and demonstrable advantages that machine learning can bring to certain kinds of decisions. He is confident machine learning will not replace or diminish the role of humans in medicine.

"It will augment and extend it," he says, "because machine learning provides better or more accurate tools to make decisions."

For example, consider this common scenario. In developing the care plan for a patient with metastatic cancer, you are trying to decide if surgery is going to improve the patient's quality of life. A key question here, Dr. Schwab says is, "How long is the patient likely to survive?" This is a good application for AI-based tools, because the whole treatment goal is to palliate the condition during the time remaining.

"If they are going to live one month, and it takes them three months to recover from surgery, it doesn't make much sense to undergo surgery," Dr. Schwab says.

Future Directions

Despite the proven clinical utility of applying ML to some aspects of patient care, Dr. Schwab emphasizes that probabilities in any field are still calculated likelihoods and not crystal ball divinations.

"Eighty percent probability of condition X means that if the model were run 100 times, the condition would occur 80 times," he says. "That means it would not occur 20 times out of 100. If condition X does not occur, the model is still valid even though the condition did not occur. This is akin to the weather report stating there is an 80% chance of rain. One still has to decide whether or not to bring an umbrella. But the weather forecast is much more accurate than our human ability to predict outcomes."

Research by Dr. Schwab and others shows that a patient's preoperative depression and anxiety scores are correlated with surgical outcomes for degenerative spinal surgery such as laminectomy for lumbar stenosis.

"Perhaps incorporating preoperative scores into prediction models would hopefully change the way we treat the patient," Dr. Schwab says. By treating anxiety before the procedure, the outcome might improve.

Dr. Schwab explains that knowing when and how to enlist the power of machine learning in patient care is key to its widespread use. He adds that Mass General is uniquely suited to lead the field because of its broad multidisciplinary medical, surgical and research network, coupled with leadership partners in computational sciences. Mass General also has experts who treat many rare conditions, such as bone tumors, which are now captured in the medical records. This provides researchers with fertile ground to study these conditions using new methods including machine learning, which may shed new light on treatment options. It is therefore invaluable in training high-precision machine learning systems and algorithms to aid in the diagnosis.

The goal is to harness the power of machine learning to help create a future of individualized medicine serving empowered patients.

"If patients have this information, they'll be able to make better decisions about their care," he says.


Paulino Pereira NR, Janssen SJ, van Dijk E, Harris MB, Hornicek FJ, Ferrone ML, Schwab JH. Development of a Prognostic Survival Algorithm for Patients with Metastatic Spine Disease. J Bone Joint Surg Am. 2016 Nov 02; 98(21):1767-1776.

van der Vliet QM, Paulino Pereira NR, Janssen SJ, Hornicek FJ, Ferrone ML, Bramer JA, van Dijk CN, Schwab JH. What Factors are Associated With Quality Of Life, Pain Interference, Anxiety, and Depression in Patients With Metastatic Bone Disease? Clin Orthop Relat Res. 2017 Feb; 475(2):498-507

About the Orthopaedic Spine Center

About the Stephan L. Harris Center for Chordoma Care


Less invasive spine procedures can reduce adverse events and improve clinical outcomes, taking advantage of new imaging technologies and protocols.


In an award-winning study, researchers at Mass General show for the first time that giant cell lesions may escape immune surveillance because of the inhibitory effect of immune modulating protein B7-H3.