Scalable High-Recall Constraint-Satisfaction-Based Information Retrieval for Clinical Trials Matching

A Stanford University team led by Professor Monica S. Lam has published a groundbreaking AI research paper introducing SatIR (Satisfiability-based Information Retrieval), a new system designed to solve the critical bottleneck of patient recruitment for clinical trials. The method fundamentally shifts away from traditional keyword or embedding-similarity matching, which often yields low recall and poor interpretability. Instead, SatIR employs formal methods from computer science—specifically Satisfiability Modulo Theories (SMT) and relational algebra—to represent the complex, multi-faceted eligibility criteria of trials as a set of precise, logical constraints. A key innovation is its use of Large Language Models (LLMs) to interpret and formalize the often ambiguous, implicit, or incomplete reasoning found in both trial protocols and patient medical records, converting them into a structured format the system can process.

In a rigorous evaluation, SatIR was tested on a dataset of 59 patient profiles against a pool of 3,621 clinical trials from ClinicalTrials.gov. The results were decisive: SatIR outperformed the previous state-of-the-art system, TrialGPT, across all measured objectives. Most notably, it retrieved between 32% and 72% more relevant-and-eligible trials for each patient. It also improved overall recall—the ability to find all potentially useful trials—by 22 to 38 percentage points, and succeeded in finding at least one suitable trial for more patients. Crucially, this high-recall, high-precision matching is achieved at scale, with the system processing a patient's profile against the entire trial database in an average of just 2.95 seconds.

The success of SatIR demonstrates a powerful hybrid AI architecture that combines the reasoning strength of symbolic AI (formal logic) with the linguistic understanding of modern LLMs. This approach not only yields superior performance but also provides the interpretability that is essential in healthcare. Each match can be traced back to the specific constraints that were satisfied, offering clinicians a clear audit trail. This addresses a major trust barrier for AI in medical decision-support systems and paves the way for more reliable automation in connecting patients with life-saving research opportunities.