Interpretable artificial intelligence-driven selection of doublet chemotherapy regimens as second-line treatment in patients with advanced pancreatic cancer.

Abstract

4171 Background: Guidelines recommend second-line (2L) doublet chemotherapy, NALIRI (liposomal irinotecan + 5-fluoruracil and leucovorin), FOLFIRI or FOLFOX, for patients (pts) with metastatic pancreatic ductal adenocarcinoma (mPDAC) after failure of gemcitabine+Nab-paclitaxel (GemNabP). A head-to-head comparison between doublets has not been performed. We aimed to apply interpretable artificial intelligence (IAI, i.e., AI systems in which the prescription logic can be understood) methods on real-world data to establish which pts should receive NALIRI vs other doublets to maximize the benefit in this setting. Methods: In this observational cohort study, we compared progression-free survival (PFS) of consecutive pts with mPDAC who received 2L doublets after GemNabP failure at 42 Italian centers between 2013 and 2023. The dataset was randomly split into a training set (70%) and a test set (30%). In the former, a counterfactual Cox proportional hazard model, including baseline characteristics to infer the 12-month PFS probability for a given patient under each regimen, was trained. An Optimal Policy Tree (OPT), a state-of-the-art IAI-based method, was used to read the complete reward matrix by training a decision tree with the counterfactual predictions, and OPT recommendations were validated in the test set. The potential gain of the new policy was evaluated by 12-month PFS net-benefit curves. Results: Among 571 eligible pts, 209 (36.6%), 209 (36.6%) and 153 (26.8%) received NALIRI, FOLFOX and FOLFIRI, respectively. Median PFS was similar among the three groups (3.3 months for NALIRI, 3.5 months for FOLFOX and 3.6 months for FOLFIRI), with a long-term benefit observed only in the NALIRI group (12-month PFS 12.0% vs 2.6% for FOLFIRI and 5.2% for FOLFOX). The OPT recommended NALIRI as the preferred regimen for pts with pancreatic head/body cancers, with ECOG PS 0 or with Ca19.9 < 109 U/ml if ECOG PS > 0. The net-benefit curves revealed that the OPT consistently outperformed the uniform strategies of administering either NALIRI or FOLFOX/FOLFIRI to all pts, attaining a 2.5 percentage-point net-benefit at a threshold probability of roughly 9%. Conclusions: Our findings show that 2L NALIRI can offer long-term PFS advantage in a subgroup of mPDAC pts compared with other doublets. The AI-derived policy provides a higher net benefit than treating all pts with NALIRI, avoiding unnecessary clinical and financial toxicity.