Presentation (CTI)

OBJECTIVES: Survival analysis and long-term extrapolation are critical components of health technology assessment (HTA) submissions and economic modeling in health economics and outcomes research (HEOR). Traditional approaches are time-intensive and require specialized expertise. This study presents a novel multi-agentic generative AI system utilizing retrieval-augmented generation (RAG) with human-in-the-loop validation to automate survival analysis workflows, generating HTA-ready outputs for regulatory submissions and economic models.
METHODS: A multi-agentic AI framework was developed incorporating three specialized agents: clinician, statistician, and medical writer, each trained on domain-specific knowledge bases. The system accepts individual patient-level time-to-event data or Kaplan-Meier curve images, with automatic digitization and pseudo-IPD generation via the Guyot algorithm. Users specify analysis requirements through natural language prompts, or the system automatically detects optimal model types. Available models include standard parametric distributions (Exponential, Weibull, Log-normal, Log-logistic, Gompertz, Gamma, and Generalized Gamma), spline-based models on odds, hazard, or probit scales with flexible knot configurations (up to three knots), cure models, piecewise models, joint-fitted models, and parametric mixture models. An iterative feedback loop enables user refinement based on clinical plausibility assessments and external validation data.
RESULTS: Validation across 10 HTA case studies demonstrated 87% concordance with expert-selected models based on minimum AIC/BIC criteria and visual fit. The system reduced analysis time by 85% (from approx. 80 to 12 hours), and produced HTA-ready deliverables including interactive dashboards, Excel-formatted economic model inputs, and dynamically generated Word reports with interpretations and narratives added. Human-in-the-loop validation ensured clinical appropriateness at each decision point.
CONCLUSIONS: This multi-agentic framework demonstrates feasibility of automating survival analyses while maintaining methodological rigor. Current limitations include inability to perform Bayesian survival modeling and computational constraints with complex mixture models. Future development will incorporate Bayesian approaches, expand external validation across therapeutic areas, and enhance real-time quality assurance mechanisms for broader HTA applicability.