Real-world predictors of adverse clinical outcomes in pancreatic cancer using a machine-learning framework.

Abstract

4186 Background: Completion of chemotherapy is crucial for clinical benefit and has been linked to improved outcomes, but is often challenging given the toxicities of chemotherapy regimens. The main treatment regimen for pancreatic adenocarcinoma (PDA), FOLFIRINOX (FFX), is highly toxic, requires frequent dose-modifications, and most patients are too sick to tolerate multiple lines of therapy. Given the aggressiveness of the disease, selecting the appropriate first-line dosing is crucial as it can be associated with better patient outcomes. The current study objective is to evaluate the association between real-world dosing patterns and clinical toxicity using a machine-learning framework. Methods: Patients attending GI oncology clinics at University of California, San Francisco between November 2011 – December 2023 with a documented administration of FFX were included. Predictors of clinical outcomes included baseline demographic and clinical features, cycle-specific laboratory data, and dosing information for FFX sub-components. Group-based 5-fold cross-validation for logistic regression, random forest, and XGBoost models were used to identify features associated with clinical outcomes (anemia, dehydration, nausea/vomiting, neutropenia, and polyneuropathy). Model performance was evaluated using AUC. Results: Data for 505 patients with PDA receiving FFX across 5,041 cycles were included.The random forest models yielded the best fit for the prediction across all outcomes (Table). Key features consistently associated with outcomes included cycle number, cumulative dose of drug received, laboratory data (PT, INR, albumin), patient demographics (male sex, race, and smoking status), and clinical features, such as hypertension. Conclusions: Our study identifies clinical features in combination chemotherapy leading to specific toxicities, highlighting these as ideal strategies for intervention. Early prognostic indicators of adverse outcomes can guide early management for high-risk individuals through supportive measures and dose modification, before high-grade toxicities and discontinuations wherein patients can no longer benefit from therapy. Summary of FOLFIRINOX model outputs across clinical outcomes. Outcome\Model Logistic Regression Random Forest XGBoost Key Features Anemia 0.61 ± 0.03 0.67 ± 0.03 0.62 ± 0.04 [1] INR[2] Irinotecan dose[3] Any polyneuropathy Dehydration 0.64 ± 0.03 0.72 ± 0.02 0.68 ± 0.05 [1] Cumulative irinotecan[2] Cumulative oxaliplatin[3] Cumulative fluorouracil Nausea or vomiting 0.62 ± 0.04 0.68 ± 0.03 0.66 ± 0.03 [1] Cumulative irinotecan[2] Cumulative oxaliplatin[3] Cumulative fluorouracil Neutropenia 0.57 ± 0.04 0.61 ± 0.06 0.61 ± 0.01 [1] Male sex[2] Cumulative fluorouracil[3] Smoking status – never Polyneuropathy 0.53 ± 0.06 0.53± 0.06 0.52 ± 0.04 [1] Male sex[2] Oxaliplatin[3] Constipation