(P 100) ADVANCING PROGNOSTICS IN ONCOLOGY: MACHINE LEARNING MODELS FOR PREDICTING 2-YEAR AND 5-YEAR SURVIVAL RATES IN PATIENTS WITH UNDIFFERENTIATED PLEOMORPHIC SARCOMA

Objective: Undifferentiated pleomorphic sarcoma (UPS) is a common soft tissue sarcoma subtype associated with high metastatic potential and poor prognosis. Current UPS survival prediction tools rely on arbitrary thresholds for certain parameters and fail to provide individualized estimates. To address this limitation, we sought to develop and validate a newly developed machine learning (ML) algorithm to predict two- and five-year overall survival in UPS patients.

Methods: A total of 3,596 patients with histologically-confirmed UPS from the Surveillance, Epidemiology, and End Results (SEER) database were included for model development. An institutional database of 127 patients was used for external validation. For model development, we transformed tabular patient data into high-dimensional embeddings using a pre-trained language model. To classify patient outcomes from these embeddings, we developed a neural network tailored for high-dimensional input. Our developed model (BERT + attention model) was compared to traditional ML algorithms. Model performance was assessed using area under the curve (AUC), F1 score, and Brier scores. Primary outcomes of interest were two- and five-year overall survival.

Results: Median age for the SEER and institutional datasets was 66 and 62, respectively (p=0.005) (Table 1). Median tumor size was higher in the SEER cohort than in the institutional cohort (8 vs. 6.5 cm, respectively). There were no differences in the proportion of patients presenting regional LN metastasis and distant metastasis between groups (Table 1). Two- and five-year OS were 70.4% and 60.6% for the SEER cohort, and 89% and 76.4% for the institutional cohort (p< 0.001 in both). Tumor size, age, metastases, and lymph node involvement were factors associated with overall survival. On internal validation, our model showed a better performance than standard ML algorithms (AUC=0.81, precision=0.78, F1 score=0.79) (Table 2). For 5-year survival, the model had an AUC=0.82, precision=0.74, and an F1 score=0.84 (Table 3). On external validation, the 2-year and 5-year survival showed excellent discriminative performance with an AUC=0.94 and an AUC=0.96.

Conclusion: In this study, we successfully developed and validated a ML algorithm that accurately predicts 2-year and 5-year survival in patients with UPS. Based on our findings, it is evident that the AI model we have developed holds significant clinical relevance and potential in aiding orthopedic surgeons in predicting survival percentages for UPS patients, utilizing a deeply personalized approach. To confirm generalizability, further external validation of this algorithm is encouraged.