Broad but Shallow - A Scalable Approach to Target Trial Emulation Using Real-World Data

<p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" class="first" dir="auto" id="d26952052e143">Every Cure is pioneering an AI-driven platform to systematically identify, evaluate, and advance drug repurposing opportunities across all known diseases and FDA approved drugs. This platform generates ~66 million drug-disease pairs, which need to be subsequently prioritised and validated to form a tractable set for human review. Target trial emulation studies using real-world data (RWD) such as health claims data provide a fast and low-cost approach to validating the outputs of model predictions. However such studies are resource-intensive and require considerable domain knowledge, limiting the possibility of scaling across many drug-disease pairs. We propose a ‘broad but shallow’ (BBS) methodology that prioritises increased generalisability over study specificity and signal, through the use of a broadly applicable study design framework with standardised input parameters and outcome measures. We focus on three classes of outcome: hospitalisation events, medication prescription patterns and healthcare resource utilisation metrics. In order to increase specificity to a disease, we integrate large language models to cohort building, generate disease-specific clinical measures based on measures that we might expect to find within RWD, and parameters such as whether proportion of days covered by a drug prescription should be required for a given drug. Here we present the initial results of a proof of concept study across ten disease areas, and ~1 million patients, using US health claims data. Initial assessments of the latter show high concordance between LLM generated outputs and clinical expectation, supporting the use of LLMs for scalable cohort and endpoint generation. Our framework provides initial evidence that supports scaling this approach to thousands of AI-prioritised drug-disease pairs, for validation of prediction strength and further triage, enabling a more systematic triage, clinical review, and eventual advancement towards the patient.