Presentation (CTI)

OBJECTIVES: In healthcare analytics, deriving actionable insights from complex patient data remains a critical challenge. While AI/ML tools enable cohort profiling, they often require extensive manual exploration and yield an overwhelming volume of findings—many of which may be trivial or redundant. To address these limitations, we present an enhanced version of the Cohort Comparison Synthesis (CCS) framework, which integrates Agentic AI to augment generative summaries with literature-based findings.
METHODS: The CCS methodology uses a multi-stage pipeline designed to generate high-quality, literature-informed qualitative summaries from prescription data: -Cohort Definition: Patient cohorts are constructed using claims and prescription data, enabling robust characterization of treatment patterns and clinical attributes. -Statistical Analysis: Contingency tables are generated to compare cohort characteristics. Statistical significance is assessed using hypothesis testing and effect size metrics, with adjustments for multiple comparisons to ensure result robustness. -Generative AI Synthesis: A GPT-4-based large language model (LLM) is orchestrated through a chain-of-thought prompting strategy to produce coherent, human-readable summaries that synthesize the key differentiators between cohorts. -Agentic AI Integration: To enhance the credibility and depth of the generated insights, we incorporate an agent which autonomously queries external literature sources (e.g. PubMed) to validate, contextualize, and expand upon the LLM-generated summaries. The agent’s behavior is dynamically tailored to the research question, ensuring relevance and interpretability.
RESULTS: When applied to a real-world problem using patient cohorts extracted from prescription data, CCS successfully highlighted statistically significant and clinically relevant differences across subgroups. The summaries facilitated strategic evaluations of market opportunities and patient targeting. The Agentic AI component successfully retrieved relevant literature to validate and expand upon key findings.
CONCLUSIONS: The CCS framework demonstrates the potential of combining statistical rigor, generative synthesis, and Agentic AI to streamline insight generation in healthcare research. Its modular design supports broad applicability across therapeutic areas, enabling more informed, evidence-backed decision-making.