Deep Learning-Based Prediction for Risk of Durotomy in Midline vs. Lateral Incisions during Robotic-Assisted Transforaminal Lumbar Interbody Fusion (TLIF)

Introduction: Durotomy is an undesirable event that can occur during robotic-assisted Transforaminal Lumbar Interbody Fusion (TLIF), potentially causing postoperative complications. While midline and lateral incisions are clinically adopted, a systematic comparison between them is lacking. This study aims to compare patient characteristics and clinical outcomes of the two incision approaches, identify potential risk factors, and establish a machine learning-based calculator for the risk of durotomy.

Methods: We identified patients from Johns Hopkins Hospital who underwent robotic-assisted TLIF and categorized them into lateral and midline incision groups. Univariate analyses were used to compare demographics, operation details, the primary outcome (durotomy), and other outcomes. Multiple machine learning survival models were evaluated through Receiver Operating Characteristic (ROC) analysis, with the Area Under the Curve (AUC) calculated.

Results: Of a total of 223 TLIF patients, 183 underwent lateral and 40 underwent midline incisions. The incidence of durotomy was significantly higher in the midline group (22.5% vs. 2.7%, p< 0.001), along with increased mortality, revision, and long-term reoperation rates. Demographics were similar across groups. The lateral approach was preferred for lumbar procedures (68.9% vs. 52.5%), while the midline approach was favored for lumbosacral, thoracic, and thoracolumbar regions (p < 0.0002). Procedures involving 4 and 6 screws were predominantly used in the lateral approach, while more complicated procedures (>6 screws) were more common in the midline group (p < 0.001). Midline patients often had poorer preoperative status, with a higher proportion scoring 50-70 on the Karnofsky Performance Status (KPS) (p=0.018) and an ASA status of 3 (p=0.016). They also experienced longer surgeries (p < 0.001), greater blood loss (p < 0.001), and lower home discharge rates (p=0.022). Postoperative complications, including delirium, septic shock, and wound infection, were significantly more common in the midline group. The Gradient Boosting Classifier model was the best-performing model (AUC=0.977), with the SHAP method identifying the midline approach and high ASA status significantly associated with elevated risk of durotomy.

Conclusion : Midline incisions in TLIF show higher risks of durotomy and complications, underscoring the need for careful surgical planning. Machine learning models could enhance patient-specific risk assessments, guiding safer incision approach selection.