James Mooney, MD, spent seven years training in neurosurgery and another year specializing in minimally invasive and complex spine deformity surgery.
The most difficult part of the job, he said, is not the operation. It is deciding who should have one. “Decision-making is definitely the hardest part of the process,” Dr. Mooney, a neurosurgeon at VCU Health in Richmond, Va., told Becker’s. “Who to operate on and when is vastly more important to outcomes than executing a technically perfect surgery.”
That is the problem he believes AI could eventually help spine surgery solve. The opportunity is not to replace surgeon judgment with an algorithm. It is to give surgeons and patients a more complete picture of what is likely to happen before they commit to an operation, drawing from imaging, bone quality, muscle quality, comorbidities, postoperative outcomes and even patients who never undergo surgery.
For a field where similar scans can produce very different recommendations, Dr. Mooney sees the potential to make decisions more consistent without making them uniform. But getting there will require confronting some of AI’s biggest weaknesses: incomplete data, embedded bias, false precision and an unsettled question of who is responsible when an algorithm sees something a physician does not.
The scan does not decide the operation
Two patients can arrive with nearly identical imaging and require entirely different treatment. One may want relief from a highly localized symptom that can be addressed with a smaller decompression. Another may have broader disability that makes treating a larger deformity reasonable. Age, bone health, smoking status, diabetes, weight and expectations can alter the risk-benefit equation further.
The goal is not to treat the X-ray. It is to determine which problem matters enough to the patient to justify the risk of fixing it. Dr. Mooney said that distinction has become more obvious as he has followed patients beyond the operating room. “BMI, osteoporosis, diabetes, smoking, all of those factors play a much bigger role in outcomes compared to how well the surgery was executed,” he said.
Patient goals can matter just as much. A patient with significant scoliosis may not necessarily need the entire deformity corrected. If the dominant problem is focal leg pain, Dr. Mooney said a targeted minimally invasive procedure could sometimes address what the patient actually wants fixed. That requires surgeons to distinguish what they are technically capable of treating from what they should treat.
Sometimes the right answer is, ‘I can’t meet that expectation’
Imaging can give a surgeon a reason to operate. Expectations can give them a reason not to. Dr. Mooney is particularly cautious when patients expect surgery to eliminate every symptom or restore them completely to a previous level of function. He prefers to undersell the likely benefit rather than promise an outcome he cannot guarantee.
“The most unhappy patients are the ones that you overpromise and can’t follow through,” he said. Sometimes that means telling a patient directly: “I don’t have the ability to meet your expectations where they are currently.”
That conversation is part of informed consent, but Dr. Mooney believes spine surgery still has too much variability in how those discussions happen. Different surgeons can evaluate the same patient and recommend substantially different operations based on training, experience and comfort with particular techniques.
AI could help, if the field first builds the right dataset.
Spine has big datasets. Dr. Mooney wants deeper ones.
Existing spine registries have given the field access to outcomes from large numbers of patients. Dr. Mooney sees a limitation in what they capture. Many were designed before surgeons understood the importance of some variables now linked with outcomes. Imaging data may be absent. Bone and muscle quality may not be characterized in sufficient detail. Wearables and real-time postoperative monitoring create streams of information that older registries were never designed to collect.
“The databases we have now don’t capture all of those data points,” he said.
That is one reason Dr. Mooney is helping develop an AI-based platform intended to collect more granular spine data from the ground up. The project is in early development and validation. Its broader goal is to combine patient-reported information with clinical variables and, eventually, imaging and electronic health record data to support surgical decision-making.
Dr. Mooney envisions a system capable of helping structure a patient’s history before the appointment and, as the dataset matures, estimating diagnoses, operative risk and the likelihood that surgery is appropriate.
The ambition is larger than creating another predictive calculator. He wants participating surgeons to contribute to, and ultimately benefit from, a dataset built around the questions spine surgeons actually need answered.
“The future of spine care and spine decision-making is going to be generated from the ground up by surgeons who are on the front lines of this,” he said.
The missing control group
There is a fundamental problem with trying to teach AI which spine patients should not undergo surgery. Much of the outcomes data available to surgeons comes from people who already did. That makes it possible to compare one surgical technique with another or identify risk factors associated with complications.
It is much harder to know what would have happened if the surgeon had never operated. Dr. Mooney believes longitudinally following nonsurgical patients could help close that gap. “If you operate on this patient, there’s a 50% chance of a complication,” he said. “Giving patients a more accurate picture of their risk allows them to make a more informed decision.”
Just as important, a dataset that includes conservative care could begin answering a question surgical registries cannot: What was the benefit of surgery compared with not operating at all? “Right now, all the databases are surgical patients,” Dr. Mooney said. “We don’t have good comparisons to patients that underwent conservative management.” That could make AI especially valuable before the OR rather than inside it.
The end goal is a ‘digital twin’
Dr. Mooney’s longer-term vision is what data scientists often call a digital twin. Collect enough high-quality information about enough patients, and an algorithm could theoretically identify patients who closely resemble the person sitting in front of the surgeon.
Then it could compare different paths. What happens if this patient receives physical therapy? What happens after a limited decompression? What happens after a larger reconstruction?
Instead of relying entirely on population averages, the surgeon could compare likely outcomes among patients with increasingly similar characteristics. “Before we even operate or pick an intervention, we can simulate how that patient’s going to do based on patients that are almost identical to them,” Dr. Mooney said.
The concept is powerful precisely because the decision is rarely binary. A surgeon may have several technically reasonable options. AI could potentially show the consequences of each one. But Dr. Mooney does not believe it can make the final choice. Every patient still brings values, goals and circumstances that are difficult to encode. That is where the digital twin stops being identical.
Better prediction could also reproduce bad history
A model learns from what happened before. That creates a problem when the past reflects unequal access or disparities in outcomes.
If disadvantaged patients historically experienced worse results, an algorithm could interpret that pattern as evidence that similar patients are poor surgical candidates, turning an existing disparity into a recommendation against treatment. Dr. Mooney sees that as one reason human oversight cannot disappear.
“If the data shows these disparities, then an algorithm could potentially reproduce and amplify them,” he said. “We can’t fully rely on these algorithms.”
Where the model is trained matters too. A patient population at a major academic center may differ substantially from patients treated in community hospitals or private practices. Predictions derived from one environment may not generalize cleanly to another. Dr. Mooney’s goal is therefore to include data from multiple practice settings rather than allowing one type of institution to define the model.
AI can calculate the pattern. The surgeon still has to understand the context behind it.
Precision can become falsely reassuring
There is another risk Dr. Mooney thinks will become increasingly important as prediction tools improve. Algorithms can sound certain even when they should not.
Anyone who has used generative AI has seen the phenomenon: a confident answer presented with far more certainty than the underlying information warrants. Dr. Mooney calls it “false precision.” A risk estimate of 51% can appear meaningfully different from one of 48%, even if the data behind both is incomplete or highly uncertain. “We have to temper that a little bit and step back and say, ‘Should I be this confident about that?’” he said.
The same technology also creates an unresolved liability question. If postoperative data are collected continuously and an algorithm detects that a patient is declining, who is responsible for acting on it? Does the physician have an obligation to monitor every data point in real time? Does the technology have to escalate a warning? And what happens if it fails to?
“There’s a whole new level of liability that needs to be ironed out,” Dr. Mooney said. Those questions will become more important as AI shifts from analyzing retrospective datasets to following patients continuously.
The goal is not to eliminate surgeon judgment
For Dr. Mooney, the endpoint is not an algorithm that tells a surgeon whether to operate. It is a better conversation. A patient considering a major reconstruction could understand with greater precision the likelihood of a complication, the expected improvement and how similar patients performed with and without surgery.
The surgeon would still interpret those numbers. The patient would still decide which risks and outcomes matter. “This is ultimately us in combination with these artificial intelligence systems, not us just blindly feeding information in,” Dr. Mooney said.
That distinction matters because spine surgery will never be made entirely consistent. Nor should it be. Two patients with the same MRI may still choose different paths because they want different lives. AI’s opportunity is not to erase those differences. It is to make sure that when the surgeon and patient choose a path, they understand far more clearly where it is likely to lead.
At the Becker’s 32nd Annual Meeting: The Business and Operations of ASCs, taking place October 29-31 in Chicago, ASC leaders, surgeons and healthcare executives will explore strategies to drive growth, enhance operational performance, navigate reimbursement challenges and prepare for the future of ambulatory surgery. Apply for complimentary registration now.
