Predicting which scoliosis patients will suffer from spinal imbalance after surgery has always been an educated guess, but a new machine learning model aims to replace intuition with math.

Surgeons treating adolescent scoliosis face a delicate trade-off. They must fuse enough of the spine to straighten it, but leave enough unfused so the teenager can still bend and move. If they miscalculate, the patient ends up with coronal imbalance, a painful trunk lean that often requires a second, highly invasive surgery to fix.

This risk is not new, but how we calculate it is. Traditionally, surgeons rely on rigid, manual classification systems that treat complex spinal curves as static shapes. By using machine learning to parse these variables, this study challenges the old habit of relying on a surgeon’s “gut feel” for distal fusion limits. It suggests that postoperative alignment is not a mystery of healing, but a predictable equation of skeletal maturity and mechanical forces.

The predictive formula

Researchers analyzed data from 282 patients with Lenke 1/2 adolescent idiopathic scoliosis who underwent selective posterior thoracic fusion. They started with 24 candidate predictors and used a dual-stage dimensionality reduction process to narrow the field. They then trained ten different machine learning architectures to see which could best predict postoperative coronal imbalance.

The standout performer was a LightGBM model. It achieved an area under the curve (AUC) of 0.885 during training and maintained a strong AUC of 0.824 in internal validation. This level of accuracy suggests the model is highly capable of distinguishing which patients are at risk before the first incision is made.

Using SHAP (SHapley Additive exPlanations) to peek inside the algorithm’s black box, the researchers identified three critical variables that drive the risk.

• LIV-LSTV relationship: Lower values between the lowest instrumented vertebra and the last substantially touching vertebra sharply increased risk.
• Lumbar Modifier C: Patients with this specific lumbar curve morphology were much more vulnerable to postoperative imbalance.
• Risser grade: Lower skeletal maturity grades correlated with higher failure rates, meaning younger, growing spines are harder to balance.

Rethinking fusion limits

This matters because it directly targets the “last joint” dilemma in spinal fusion. Surgeons often struggle to decide where to stop fusing the spine to preserve motion without risking collapse. By highlighting the relationship between the lowest instrumented vertebra and the last touching vertebra, this model provides a concrete, mathematical boundary for that decision.

Instead of looking at the spine as a single curve, the AI forces clinicians to look at the interaction between skeletal maturity and the specific vertebra chosen for the anchor. If a patient has a low Risser grade and a Lumbar Modifier C, the surgeon must accept that aggressive motion preservation might guarantee a mechanical failure. The model shifts the conversation from how much motion can be saved to how much fusion is required to keep the patient upright.

However, the tool is not ready for the operating room just yet. The study relied entirely on a single-center cohort for its split-sample internal validation. Without independent external validation on diverse patient populations, we cannot be sure if these algorithmic rules hold true across different surgical techniques and demographic groups.

Read the full study in International Orthopaedics.