OC-032 IMAGE-BASED DEEP LEARNING MODELS (DLMS) TO PREDICT LONG TERM QUALITY OF LIFE (QOL) FOLLOWING ABDOMINAL WALL RECONSTRUCTION (AWR)

Abstract

Listen

Translate article

Abstract Aim To apply image-based DLMs to predict post-operative QOL following AWR. Materials & Methods A prospective, institutional hernia-database was queried for patients with preoperative abdominal CT-imaging, a preoperative and 1-year postoperative Carolinas Comfort Scale(CCS) survey, and no recurrence. “Symptomatic” was defined as CCS-score≥2(2=mild and bothersome). Google Xception existing architecture model was used with ImageNet database pre-initialized weights to classify symptomatic and non-symptomatic patients. Patients were divided into 80:20-training:testing samples for model generation and evaluation. Model training, test accuracies, and loss-functions were evaluated to determine performance and discriminative ability. Results Of 244 patients, mean age was 60.4±11.8 years, mean BMI:33.0±7.1kg/m2, female:57.1%, tobacco use:14.3%, diabetic: 24.5%. Median[IQR] hernia defect size was Exactly 180cm2[90–324]; 66.1% had a failed repair. CDC wound classifications Included: 75.9% class-I, 8.3% class-II, 9.1% class-III, 6.6% class-IV. Preoperatively, hernia-related pain(70.2%) and movement limitations(72.3%) were common. Mesh position was predominantly preperitoneal(91.6%). Median[IQR] mesh size was 900cm2[572–1050]. Anterior component separation was required in 17.9% and posterior in 20.4%. One-year postoperatively, reported symptoms included: mesh sensation-39.5%, discomfort-37.8%, movement limitations-37.0%. DLMs utilized 6,441-CT-images(5,097 training-sample). Proportions of symptomatic patients were 48.9%(85/174) in the training-sample and 50%(35/70) in the test-sample. Highest DLM training accuracy was 85.37%(loss=0.3766) at epoch 15/50 with 79.30%(loss=0.3766) comparative validation accuracy, demonstrating strong discriminative ability in model classification between symptomatic and asymptomatic patients. Lower accuracy due to model overfitting was observed after 50 epochs. Conclusions Image-based DLMs using standard, preoperative CT images very successfully predicted 1-year AWR QOL. The impact of DLMs on preoperative counseling/consent for surgery could be revolutionary.