
Can Machine Learning Help Predict Recovery After Surgery for DCM?
One of the hardest parts of living with Degenerative Cervical Myelopathy (DCM) is the uncertainty. Surgery can stop further spinal cord damage for many people, but recovery afterwards varies widely. Some patients notice meaningful improvements in function, while others improve only a little, despite similar scans or similar operations.
This raises a practical question patients often ask: “What are my chances of getting better?” A recent study by Shuai Zhou and colleagues explored whether machine learning (computer methods that learn patterns from large datasets) can help predict short-term neurological improvement after decompression surgery for DCM, using information already collected before surgery [1].
Why Was This Study Conducted?
Clinicians already know that factors like age, symptom severity, and overall health can influence outcomes after surgery. The problem is that these factors interact in complex ways, and it can be difficult to combine them into a clear, individualised estimate of recovery using traditional approaches.
Machine learning models are designed for exactly this kind of task: they can combine many variables at once to estimate the likelihood of a particular outcome. The authors aimed to build a model that could support pre-operative counselling and shared decision-making, and then simplify it so it could be easier to use across different hospitals.
How Was the Study Designed?
The researchers reviewed records from 1,895 adults who had surgery for DCM at a large hospital in Beijing between 2011 and 2020. After applying eligibility criteria (including having follow-up data), 672 patients were included in the final analysis.
They trained and tested five different machine learning approaches, using information available before surgery, including:
- age and sex
- neurological status measured using the JOA score (Japanese Orthopaedic Association score), one of the most commonly used ways to assess spinal cord function in DCM (it captures things like movement, sensation, and function)
- health and quality-of-life measures from the SF-36 questionnaire (a widely used survey covering areas such as physical function, general health, and pain)
- clinical symptoms and examination findings
- whether there was a high signal within the spinal cord on T2-weighted MRI (a type of MRI finding that can reflect stress or injury within the cord)
The outcome they tried to predict was whether someone achieved a minimal clinically important difference (MCID) in their JOA score at 3–6 months after surgery. MCID means the improvement is big enough to be noticeable and meaningful, not just a small change on paper.
They also tested a simplified version of the model on an external dataset of 58 patients from two other hospitals.
What Were the Findings?
1) One model performed best overall.
Among the tested approaches, a model called LightGBM (a type of machine learning method that is especially good at finding patterns in complex clinical data) performed the best at predicting who would reach a meaningful improvement by 3–6 months after surgery.
The authors reported an AUROC of 0.745 in the testing dataset. AUROC (area under the receiver operating characteristic curve) is a standard way of describing how well a prediction model separates “more likely to improve” from “less likely to improve”. Values closer to 1.0 are better; 0.5 is no better than chance.
2) A simpler model worked almost as well.
To make the approach easier to apply in real-world clinics, the authors created a “feature-reduced” model — meaning it used fewer inputs. This simpler model relied mainly on age, sex, baseline JOA score, and SF-36 scores, and still performed similarly (AUROC 0.734) in internal testing.
3) The simplified model held up in external validation.
When tested on the independent external dataset, the feature-reduced model achieved an AUROC of 0.785, suggesting it may generalise reasonably well beyond the original centre.
4) Some pre-operative factors mattered more than others.
The most influential predictors included:
- Baseline JOA score (how severe the myelopathy was before surgery)
- Age
- SF-36 domains related to physical function, general health, and bodily pain
One important nuance the authors highlight is that people with very high baseline JOA scores may have less “room” to improve on that specific scale (a “ceiling effect”), which can affect whether they reach MCID — even if they still benefit in other ways.
Why Does This Matter?
Studies like this point towards a future where patients and clinicians may have better tools for personalised pre-operative discussions. If a model can estimate the likelihood of meaningful short-term improvement, it could help:
- set expectations more realistically
- identify patients who might benefit from closer follow-up or additional rehabilitation support
- strengthen shared decision-making by integrating multiple factors at once
It’s also encouraging that the feature-reduced model performed well, because tools that depend only on routinely collected measures are more likely to be usable across different healthcare systems.
What Do These Findings Mean (And What Don’t They Mean)?
Prediction models can support conversations, but they cannot guarantee outcomes for an individual person. They estimate probabilities based on patterns seen in previous patients.
The authors also note limitations:
- outcomes were measured at 3 -6 months, which is relatively short-term
- most data came from a single centre
- imaging inputs were limited (for example, they focused on whether T2 signal change was present rather than using more detailed imaging measures)
Despite this, the external validation is a real strength, and it’s encouraging that the simpler model retained performance when tested outside the original hospital.
We commend the authors for tackling a clinically meaningful question and for prioritising a model that could be practical to apply, rather than only reporting the most complex version.
If you found this interesting, consider sharing it to help raise awareness of DCM, grow the Myelopathy.org community, and support future work aimed at improving prediction, rehabilitation, and outcomes for people living with myelopathy.
Want To Learn More?
[1] Zhou S, Liu Z, Huang H, et al. Predicting postoperative neurological outcomes of degenerative cervical myelopathy based on machine learning. Frontiers in Bioengineering and Biotechnology. 2025. doi: 10.3389/fbioe.2025.1529545
