(P 173) DEVELOPMENT AND EXTERNAL VALIDATION OF A MACHINE LEARNING ALGORITHM TO PREDICT OVERALL SURVIVAL IN PATIENTS WITH PELVIC OR APPENDICULAR LIPOSARCOMA

Objective: Liposarcoma accounts for approximately 20% of all soft tissue sarcomas, making it one of the most common subtypes. Despite this prevalence, predicting survival outcomes remains challenging due to significant heterogeneity across histologic subtypes, anatomical locations, and patient characteristics. Current prognostic approaches rely on limited clinical factors and lack the ability to integrate complex personalized data. Machine learning (ML) offers a promising solution to enhance predictive accuracy through pattern recognition in high-dimensional datasets. Thus, this study aimed to develop and validate an ML algorithm capable of predicting one-, two-, five-, and ten-year overall survival (OS) in liposarcoma patients, with adaptability to treatment type.

Methods: We included 6,643 patients with histologically confirmed liposarcoma from the SEER database for model training and internal validation, plus 200 patients from an institutional database for external validation. Patient data were converted into high-dimensional embeddings using a pre-trained language model. Multiple ML algorithms were developed and compared, with performance assessed using AUC, precision, recall, F1-scores, and Brier scores.

Results: Key predictors included histologic subtype, tumor size, anatomical location, age, metastatic status, and surgical margins. Our neural network demonstrated superior performance with internal validation AUCs of 0.82, 0.77, 0.68, and 0.6 and external validation AUCs of 0.93, 0.96, 0.95, and 0.95 for one-, two-, five-, and ten-year OS, respectively.

Conclusion: We successfully developed an ML algorithm that accurately predicts survival in liposarcoma patients, adaptable to treatment type. Additional multi-institutional external validation is recommended for broader clinical applicability.

Comparison of performance metrics for the internal validation and external validation of the neural network