Pre and postoperative machine learning models and point-based scores to predict risk of early recurrence in upfront resected large Hepatocellular carcinoma

Background: Large Hepatocellular Carcinoma (LHCC) are aggressive tumours characterized by a high risk of early recurrence (ER). Although several models predicting this risk exist for HCC, no one is specific for tumours ≥5 cm. The aim of this study is to develop classic and machine learning (ML) models able to identify patients with this pattern of recurrence. Method: A retrospective, multicentric analysis of 12 hepato-biliary centres. Only upfront resected LHCC were included. ER was defined as recurrence within 8 months after resection. Logistic Regression (LR), Elastic Net, Decision Tree, k-nearest neighbors, Random Forest (RF) and Extreme Gradient Boosting were trained and compared though the resulting c-statistic. Results: Between 2016 and 2022, 724 patients met the inclusion criteria. ER was reported in in 225 (31.1 %) patients. Among the five ML models, RF showed the best performance to predict ER (pre- and postoperative c-statistic: 0.685-0.719). LR showed similar accuracy compared to RF, both preoperatively (c-statistic: 0.678) and postoperatively (c-statistic: 0.720). This model was therefore used for two point-based scores, which were split into three groups according to the risk of ER: low, intermediate and high risk (ER for preoperative score: 15 %, 31 % and 45 %; postoperative score 17 %, 40 % and 63 %, respectively). Both scores correctly stratify patients' overall survival and risk of death (p < 0.001). Conclusion: Two easy-to-use point-based scores were created, able to predict the risk of ER. These can be easily implemented in clinical practice and define best candidates for perioperative therapies (https://thibaut-goetsch.shinyapps.io/lhcc_score_preop and https://thibaut-goetsch.shinyapps.io/lhcc_score_postop).