Addressing patients to neoadjuvant systemic chemotherapy followed by surgery rather than surgical resection upfront is controversial in the case of resectable colorectal -liver metastases (CLM). The aim of this study was to develop a machine-learning model to identify the best potential candidates for upfront surgery (UPS) versus neoadjuvant perioperative chemotherapy followed by surgery (NEOS). Patients at first liver resection for CLM were consecutively enrolled and collected into two groups, regardless of whether they had UPS or NEOS. An inverse -probability weighting (IPW) was performed to weight baseline differences; survival analyses; and risk predictions were estimated. A mortality risk model was built by Random-Forest (RF) to assess the best -potential treatment (BPT) for each patient. The characteristics of BPT-upfront and BPT-neoadjuvant candidates were automatically identified after developing a classification -and -regression tree (CART). A total of 448 patients were enrolled between 2008 and 2020: 95 UPS and 353 NEOS. After IPW, two balanced pseudo-populations were obtained: UPS = 432 and NEOS = 440. Neoadjuvant therapy did not significantly affect the risk of mortality (HR 1.44, 95% CI: 0.95-2.17, p = 0.07). A mortality prediction model was fitted by RF. The BPT was NEOS for 364 patients and UPS for 84. At CART, planning R1vasc surgery was the main factor determining the best candidates for NEOS and UPS, followed by primitive tumor localization, number of metastases, sex, and pre-operative CEA. Based on these results, a decision three was developed. The proposed treatment algorithm allows for better allocation according to the patient's tailored risk of mortality.
Keywords: colorectal liver metastases; liver surgery; machine-learning; neoadjuvant chemotherapy; upfront surgery.