Evidence-linked complementary care AI supporting standard oncology: Real-world outcomes in pancreatic, HNSCC, and prostate cancer.

Abstract

e23381 Background: Complementary care may improve oncology treatment tolerance but often lacks rigorous safeguards. We developed an evidence-linked AI system generating individualized supportive protocols (micronutrients aligned with DNA damage response) under strict "No evidence → No claim" constraints. This system integrates with standard-of-care (SoC) to enhance safety and auditability. Methods: We established a prospective registry (N = 23; data cutoff 01/07/2026) with structured clinical/SoC data capture. This retrospective analysis of de-identified secondary data was deemed exempt by Advarra IRB (Protocol BTBC-2026-00; Ref Pro00092368). AI logic selected supportive products by matching micronutrient mechanisms to DNA damage pathways within safety constraints. The AI platform employed a locked prompt sequence (Extract → Validate → Transform → Analyze → Narrate) to construct an evidence ledger comprising verified data (Level 1) and derived outcomes (Level 2). Outcomes were compared between adherence levels (0/1/2) using covariate-adjusted Restricted Mean Survival Time (RMST) truncated at 12 months. RMST regression utilized pseudo-observations (R package pseudo) adjusting for age, gender, stage, and time-to-entry. Immediate non-cancer deaths in non-adherent patients were excluded to prevent bias. Toxicity metrics were assessed via multivariable regression; missing data used complete-case analysis. Results: Among 23 patients (Full Adherence n = 17; Lower Adherence n = 6), higher adherence was associated with superior outcomes: Mean planned dose intensity (PDI) was 18.3% higher (95% CI: 10.5 to 26.1, estimated) 55% reduction in treatment interruptions (IRR 0.45) 9.2 fewer grade ≥2 adverse event days per patient (95% CI: -13.5 to -4.9) 85% faster resolution of adverse events (HR 1.85) 48% fewer unplanned acute-care visits (IRR 0.52) 88% of patients reported stable or improved quality of life In the covariate-adjusted RMST analysis truncated at 12 months, each level increase in adherence (0 to 1 to 2) yielded +1.55 months of survival; full adherence (Level 2) was associated with a +3.1-month survival gain compared to non-adherence (Level 0) after adjusting for age, gender, stage, and time-to-entry. Although the survival trend was positive (p = 0.088), statistical significance was limited by sample size. CBR increased from 47.4% (low) to 95.5% (full). No safety events were attributed to the protocol. Conclusions: An AI-driven complementary care system demonstrated reductions in adverse events and acute-care visits, improved chemotherapy dose delivery, and suggestive survival benefits (+3.1 months in full adherence; p = 0.088). This structured, evidence-based approach ensures auditability. A larger prospective, multi-site validation is recommended to confirm these findings and explore subgroup-specific effects.