Technical deep dive: how Monsana’s AI agents decode complex clinical trial criteria

At Monsana, one of the biggest challenges we face when matching patients to clinical trials is the ambiguity of clinical trial criteria and the complexity of medical patient data. In this article, we focus on two specific examples we encountered during our research: the ambiguity of medical history and the complexity of genetic mutation.

The challenge of interpreting medical history

Patient records often blend current diagnoses, past medical events, treatments, and incidental findings in a way that’s not always straightforward to interpret. Take, for example, this snippet from a patient file below.

“The patient completed their last chemotherapy cycle for breast cancer 9 months ago and is currently in follow-up. She was diagnosed with hypertension 5 years ago and now presents with new neurological symptoms.”

Here, is the breast cancer still an active diagnosis or part of the medical history? Is the hypertension relevant to the current condition? How should “currently in follow-up” be interpreted? These are tricky questions, even for experienced clinicians, and they’re even harder for a standard AI model.

To address this, we developed a specialized AI agent that doesn’t jump to conclusions but instead reasons through the information step-by-step. This agent asks itself a series of targeted questions, much like a doctor would:

1. What is the current active diagnosis?

2. What other medical conditions are present, and when were they diagnosed?

3. Do these conditions relate to or affect the current diagnosis?

4. Are any of these conditions exceptions noted in the trial’s exclusion criteria?

5. Based on this, is the patient eligible for the trial?

Example: Metastatic breast cancer patient

Consider a patient with metastatic breast cancer diagnosed in 2023 who is being treated with hormonal therapy. Her history includes a Hodgkin's Lymphoma in 2012 and she has had a recent influenza infection. The trial exclusion criterion states:

History of another primary malignancy except for malignancy treated with curative intent with no known active disease ≥ 3 years before the first dose of study intervention with the exception of treated non-melanoma skin cancer or adequately treated carcinoma in situ without evidence of disease (e.g., cervical cancer in situ).

Our AI agent reasons through this by identifying the breast cancer as the active diagnosis and noting that prior malignancies were treated long ago, well outside the 3-year exclusion window. The influenza infection isn’t a malignancy, so it doesn’t affect eligibility. As a result, the patient is deemed eligible based on their medical history. Below is the reasoning made by Monsana's AI agent.

The patient has a current diagnosis of breast cancer, which is a malignant condition. The exclusion criterion states that a history of other malignancy within the past 2 years would lead to exclusion from the trial. The patient had Hodgkin lymphoma treated in 2012, which is outside the 2-year window. The influenza A infection is not a malignancy. Therefore, the patient does not meet the exclusion criteria based on their history of malignancies.

Decoding genetic mutation data: a different kind of challenge

Genetic mutation reports pose a whole new set of difficulties. These pathology documents are often extremely long and dense, listing numerous genes analyzed, with complex jargon peppered throughout. Crucially, the gene of interest might be mentioned only briefly, sometimes just a few letters like “MTAP” or “CDKN2A”, somewhere deep in the text.

Standard large language models (LLMs) struggle to maintain context across such lengthy documents and often “lose” important information buried in the middle.

To overcome this, our genetic mutation AI agent uses a clever two-step approach:

• First, it scans the entire report to detect mentions of the specific genes of interest.

• Then, it extracts focused snippets surrounding each gene mention. These snippets contain the relevant context needed to interpret the gene’s mutation status or expression.

  
  

By “zooming in” on the relevant parts of the report, the specialized AI agent avoids losing critical details and can accurately determine eligibility for criteria that are only mentioned subtly within the patient files.

Why this matters

Combining structured, stepwise reasoning for medical history with focused context extraction for genetic mutations allows Monsana’s AI agents to navigate the messy realities of clinical data with precision. This leads to more accurate, explainable, and trustworthy eligibility assessments, helping clinicians find suitable trials faster and bringing patients closer to promising new treatments.

If you want to learn more about our AI agents or collaborate, contact us at valerie.vandeweerd@monsana.ai or connect via LinkedIn.