Phesi has launched a new Protocol Complexity analysis module within its AI-powered Trial Accelerator platform that lets sponsors identify and reduce protocol complexity before it drives up cost, timelines and execution risk.
Client protocols or synopses are uploaded, deconstructed and extracted using agents, then benchmarked against comparable historical trials among more than 600,000 protocols in Trial Accelerator, in minutes. The feature generates a complexity score using a transparent, rule-based engine across five domains: patient burden, investigator site burden, endpoint complexity, eligibility complexity and operational complexity. Protocol Complexity functionality also produces targeted recommendations to address flagged complexity issues. Typical cost and complexity reductions of 10-25% have been achieved in different indications and target patient subpopulations.
Unlike generic benchmarking, Protocol Complexity identifies the drivers behind a complexity score, flagging where it departs from modal protocol design values in comparable, successful trials – for example, too many patient visits, endpoint and outcome measure selection or overly narrow eligibility criteria. It produces recommendations drawing on Phesi's proprietary patient, disease, protocol design, enrolment, country and investigator site data. Recommendations are split between direct complexity-reduction actions, such as relaxing an overly restrictive eligibility criterion, and planning recommendations, such as reviewing the number of countries and investigator sites. Built-in scenario modelling lets users test the impact of a change before committing to a protocol amendment, so sponsors can model the effect before they act.
"What looks like protocol complexity is usually a symptom of something else – poor design, overcollection of data, too many countries, too many investigator sites, too many patients," said Dr. Gen Li, Phesi founder and CEO. "Too often, this complexity is simply treated as an unavoidable cost of doing business. In reality, understanding the causes of complexity is sponsors' clearest route to reducing risk and costs, and decreasing the burden on patients and investigator sites. Left unaddressed, these factors are what turn into ‘mid-flight’ amendments and operational delays, and often lead to trial failures."
Protocol Complexity uses agents to read, extract and verify protocol data in minutes. Once a document is extracted, a large language model (LLM) converts the unstructured protocol into structured data, including study visits, procedures, endpoints, objectives and eligibility criteria. A second LLM then re-reads the protocol and audits the extraction, producing findings with an evidence quote from the source document, a suggested fix and a confidence level. The complexity score is produced via a rule-based engine rather than generative AI, built on a peer-reviewed scoring framework with every score traceable back to a specific factor in the protocol. A human remains in the loop to accept or dismiss each finding – the system proposes, but the user decides.
"Phesi has been building and refining AI-driven methods for over 20 years. This is the evolution of work we've been doing since long before 'AI' became a buzzword," said Dr. Li. "Protocol Complexity is an important step in a more agentic approach, with AI handling the heavy lifting of structuring data, and humans remaining in control of decisions. We combine AI and algorithmic methods to give sponsors actionable insights, without sacrificing rigor or explainability. Analysing protocols against contextualised, outcome-linked data is essential – without it, even the most sophisticated AI is only as good as the data it's built on."
Protocol Complexity supports more than 4,000 indications, from the most studied to ultra-rare conditions, across single-country and global studies in all trial phases. The feature draws on Phesi's contextualised database: tabulated inclusion/exclusion criteria, outcome measures, endpoints and treatment duration data from more than 615,000 protocols, 119,662 schedule-of-activities tables, and almost 100,000 full protocols. Earlier analysis from Phesi found that only around 29% of protocols in these datasets are linked to trials with documented patient outcomes – underlining why protocol volume alone is not a reliable basis for benchmarking or AI-driven recommendations.