(P 463) MACHINE LEARNING CORROBORATES TRADITIONAL REGRESSION TECHNIQUES IN PREDICTING MAJOR WOUND COMPLICATIONS AFTER SOFT TISSUE SARCOMA RESECTION WITHOUT IMPROVEMENTS IN PREDICTIVE PERFORMANCE

Objective: Postoperative major wound complications (MWCs) are common complications that occur after wide surgical resection of soft tissue sarcoma (STS). While infections have not been shown to affect oncologic outcomes, they are associated with substantial morbidity. Significant work has been directed at identifying risk factors associated with the development of MWC. Risk factors identified in previous work include neoadjuvant external beam radiation therapy (EBRT), preoperative chemotherapy, diabetes, patient age, resection size, tumor size, lower extremity tumor location, and wound closure methods.

There is considerable complexity in the development of a MWC following STS resection, and the etiology may be multifaceted. Quantification of the relative importance of identified risk factors has not been achieved. Developing predictive models to better understand these risk factors is essential in setting the foundation of future work aimed at reducing the incidence of MWC

The purpose of this study was to utilize logistic regression and machine learning modeling to quantify the risk factors associated with the development of MWC following resection of STS.

Methods: Institutional Review Board approval was obtained for this study. We retrospectively reviewed all patients at a single sarcoma center who were treated for STS of the extremity or trunk between 2008 and 2022. Data collected included the following: sex, age, patient body mass index (BMI), tobacco use, diagnosis of diabetes, presence of metastases on presentation, type of surgical closure, and tumor characteristics (site, size, volume, depth, grade, histological subtype). Laboratory values at time of resection were collected. MWCs were identified and were defined as any of the following within four months of initial resection: formal wound debridement in the operating room, procedural interventions such as percutaneous drainage, treatment with intravenous antibiotics, or deep wound packing for more than 120 days.

Univariate, logistic regression, and machine learning (XGBoost) analyses were performed to identify risk factors for MWCs. Model performance was assessed by receiver operating characteristic area under the curve (AUC).

Results: Six hundred and seventy-seven patients were treated with wide surgical resection of STS between 2008 and 2022 at our institution. Of these patients, 177 had MWCs (26%). In the univariate analysis, the following factors were significantly associated with MWCs: BMI (p < 0.0001), diabetes (p=0.03), neoadjuvant radiation (p < 0.0001), lower extremity tumor location (p < 0.0001), presence of metastases on presentation (p=0.005), chemotherapy (p=0.0017), resection size (p < 0.0001), method of closure (p=0.017), drains (p=0.0015), and low platelet count (p-0.025).

The sensitivity and specificity of the logistic regression model were determined to be 72.4% and 67.9% with an AUC of 0.72. The strongest risk factors for development of a MWC were preoperative radiation and presence of metastases. Figure 1

The sensitivity and specificity of the XGBoost model were 80.6% and 61.4% respectively with an AUC of 0.71. The most important factors identified to be predictive for development of a MWC were preoperative radiation and resection size. Figure 2

Conclusion: Logistic regression modeling and machine learning-based regression models perform similarly. Both models identify preoperative radiation as the most important risk factor for MWCs following STS resection. Accurate modeling can guide efforts to minimize MWCs, and future efforts should be directed at improving and validating models.