Comparison of Patients With and Without Predicted Surgical Indication Between Clusters of Adult Spine Deformity (ASD) Patients.

STUDY DESIGN

Multicentric, retrospective analysis of prospectively collected data.

OBJECTIVE

To utilize machine learning (ML) for clustering and management prediction (conservative vs . operative) in surgically treated adult spine deformity (ASD) patients, and to compare the attainment of the minimum clinically important difference (MCID) between predicted surgical and conservative patients.

SUMMARY OF BACKGROUND DATA

Management choice in ASD is complex. ML can identify patient clusters and predict treatment, but it is unclear whether patients treated according to the prediction also show better clinical outcomes.

MATERIALS AND METHODS

ASD patients (2-yr follow-up) were divided into groups using k-means clustering. Management choice was predicted among operated patients in each cluster. The MCID for the Oswestry Disability Index (ODI) and the Scoliosis Research Society-22 (SRS-22) were calculated and compared between patients with and without surgical prediction.

RESULTS

In cluster 1 (idiopathic scoliosis, n=675, 150 surgeries), 57% of patients had a conservative prediction. Of these, 52% and 49% achieved MCID for ODI and SRS-22, respectively, compared with 68% and 75% for those with surgical predictions [odds ratio (OR)=2 and 3.1, respectively].In cluster 2 (moderate sagittal imbalance, n=561, 200 surgeries), 12% had a conservative prediction. Of these, 29% and 46% achieved MCID for ODI and SRS-22, respectively, compared with 47% and 56% for those with surgical predictions.In cluster 3 (significant sagittal imbalance, n=537, 197 surgeries), 17% had a conservative prediction. Of these, 12% and 15% achieved MCID for ODI and SRS-22, respectively, compared with 37% and 45% for those with surgical predictions (OR=4.2 and 4.5, respectively).

CONCLUSION

Patients with concordant surgical prediction and management had higher odds of achieving the MCID, indicating a good correlation between prediction and clinical outcomes. In cluster 3, the low percentage of patients with conservative prediction achieving the MCID suggests that ML could well identify patients with poor clinical outcomes.