Chief of spinal deformity surgery Columbia University NYP/Och Spine Hospital New York, NY, US
Disclosure(s):
Michael Fields, MD: No financial relationships to disclose
Introduction: Quantifying a patient’s risk for neuro deficit during ASD surgery remains challenging given the myriad of perioperative factors. Therefore, a ML approach was used to generate an accurate scoring system to predict root-level IONM data loss.
Methods: 735 ASD patients from a single-center (2015-2023) were studied. 193 distinct perioperative variables were included (demographics, diagnosis, medical history, physical exam, operative factors, labs, preop/intraop x-rays, preop MRI/CT). Root-level IONM data loss was defined as either complete signal loss or when motor-evoked potentials met institutional warning criteria and was reviewed with the senior member of the IONM team. Patients were randomly allocated into training/testing(75%/25%) cohorts to train a random forest and logistic regression ML classifier. Threshold values for features were calculated from trained random forest model, and scores were derived by rounding up weights from the logistic regression model. Variables in the final scoring calculator were selected to optimize predictive performance (accuracy/sensitivity/specificity/AUC).
Results: The rate of root-level IONM loss was 5%(39/735) and most involved the tibialis anterior(72%), extensor hallucis longus(56%), and vastus(33%). Of those with root-level loss, 26% had severe motor deficit(≤grade 3 anti-gravity strength) immediately postop. Through the ML approach, 5 features were included. Calculated scores for Lumbar Foraminal/Lateral Recess Stenosis, Preop to Intraop Delta in Hematocrit≥12, Lower Instrumented Level at Ilium, Dural Tear, and Three Column Osteotomy were 3, 1, 1, 1, and 1 point, respectively. An aggregate score of ≤3 had an associated risk of nerve-root IONM loss of 1.7%, while those with total scores ≥6 had a rate of 60%. On the test cohort, the scoring system achieved an accuracy, sensitivity, specificity, and AUC of 86.4%, 80%, 86.6%, and 0.83, respectively.
Conclusion : This study introduces the first ML-derived scoring system that can be used preoperatively to predict root-level IONM motor data loss for ASD patients undergoing reconstructive surgery with excellent model performance.
.jpg)
