Clinically enhanced predictive modeling for end-of-life care in colon cancer: Identifying opportunities to support clinical decision-making.

553 Background: Timely palliative care (PC) integration improves outcomes for colon cancer patients with serious illness, but optimal identification remains a challenge. This study assessed PC's impact on hospital outcomes for deceased colon cancer patients and developed clinically enhanced machine learning (ML) models to predict PC utilization, aiming to identify patient subgroups that might warrant earlier PC consideration and thereby support clinical decision-making. Methods: A retrospective cohort study utilized the HCUP National Inpatient Sample (NIS; 2018-2021 training; 2022 validation), an approximate 20% sample of U.S. discharges. Deceased colon cancer patients (PC N=3,454; No-PC N=2,647 in 2018-2021 sample) were compared on median total hospital charges (TOTCHG), length of stay (LOS), and mean clinical characteristics. ML models (Logistic Regression, Random Forest, Gradient Boosting) using demographic, administrative, and extensive engineered clinical features predicted PC use. Models were internally and externally validated. The best model by PR-AUC (Precision-Recall Area Under Curve) was calibrated and used to generate risk scores for PC utilization among deceased non-PC patients, highlighting those with probabilities >0.5 as potential "missed opportunities" for clinical review. National estimates were derived. Results: For deceased colon cancer patients (2018-2021 NIS sample), PC was associated with significantly lower median TOTCHG ($49,773 vs. $85,790, p<0.0001) and shorter median LOS (5.0 vs. 6.0 days, p<0.0001). Gradient Boosting demonstrated the best PC prediction (Internal Test: PR-AUC 0.392; External Validation: PR-AUC 0.412), with ROC-AUCs ~0.85. Though its sensitivity for identifying PC cases was modest (~0.12-0.13), the calibrated model identified 246 deceased non-PC patients in the 2018-2021 sample (extrapolating to ~1,230 nationally) with a predicted PC probability >0.5. These patients represent a cohort that, based on their complex clinical profiles, might have benefited from earlier PC discussion; if PC had been provided, a projected national charge reduction of approximately $44.3 million and 1,230 fewer hospital days could have been realized. On 2022 data, a similar cohort was identified with comparable projected impacts. Conclusions: Palliative care is associated with substantial reductions in hospital charges and LOS for deceased colon cancer patients. Clinically enhanced machine learning models can robustly predict PC utilization with high discrimination. While not replacing clinical judgment, these models can serve as a decision support tool by identifying high-risk patient profiles who might benefit from earlier palliative care consultation, potentially leading to significant improvements in care quality and resource optimization on a national scale.