Earlier this year, we rolled out AI-powered patient pre-screening capabilities in the Inato platform. Research sites can use these capabilities to help them quickly find, assess, and enroll the right patients for their trials, transforming a historically burdensome process.
Anatole Callies, Inato’s Lead AI Engineer, was one of the driving forces behind our efforts to reimagine patient identification and pre-screening. He recently co-authored a research paper offering an in-depth look at how Inato uses AI to tackle recruitment and enrollment challenges, with robust analysis on the impact for sites. That paper, Real-world validation of a multimodal LLM-powered pipeline for High-Accuracy Clinical Trial Patient Matching leveraging EHR data, is out now.
We sat down with Anatole to learn more about his work at Inato, the AI behind Inato patient pre-screening, and the impact on site efficiency.
Before we discuss patient pre-screening, tell us a bit more about your role and what brought you to Inato.
I joined Inato because I was excited about the opportunity for AI in clinical trials. Trials are the number one bottleneck in bringing new drugs to market, and there are so many ways that AI can make the process better.
We’ve already identified impactful use cases for AI at all the stages of the clinical research lifecycle–from optimizing protocols by finding and eliminating unnecessary restrictions, to identifying and recommending the most relevant sites to sponsors. Additionally, there are endless opportunities to accelerate and improve patient recruitment, starting with pre-screening, which we’re going to talk about in-depth today. In my role, I’m focused on determining where AI can add the most value for sites and sponsors, to ultimately make research more accessible and efficient.
You’ve played a key role in developing and honing Inato’s patient pre-screening capabilities. Given all of the possibilities for AI in clinical research, how did you and the team decide to focus on this challenge?
We repeatedly hear from both sites and sponsors that patient recruitment is the number one challenge in clinical trials. Today, many research sites find eligible patients by manually evaluating patient records–which are hundreds of pages long–against inclusion and exclusion criteria. One study found that for Phase III trials, sites review an average of eight patients to find just one eligible participant. With review times averaging 50 minutes per patient, this process can take nearly seven hours to identify a single patient, and that patient may or may not choose to participate.
It’s a very labor-intensive process, and it’s actually becoming more difficult as inclusion and exclusion criteria become more stringent. This is a major barrier to improving clinical research, and one that AI is very well-suited to address.
How did you go about developing a solution?
Previous studies demonstrated the potential to use LLMs for patient pre-screening, but those applications fell short in real-world settings due to some technical and operational limitations at the time. We leveraged a couple of recent technical advances to overcome these shortcomings.
First, recent advances in large visual language models (VLMs) and visual retrieval enabled us to implement a fully visual pipeline. This is important because patient records don’t just contain text–many of them also include visual information like doctor handwriting, charts, or graphs. Historically, converting visual elements into machine-readable text required optical character recognition (OCR), which is both expensive and error prone. Though AI is on its way to replacing OCR, this conversion is never neutral and always results in a loss of information that’s difficult to assess and mitigate. That’s why we went with a third approach, bypassing conversion altogether by using an AI that’s able to interpret images of documents at face value.
Additionally, we used a new reasoning model paradigm, introduced by OpenAI’s o1. Previous models tended to struggle with logic or even basic arithmetic, which of course created issues when they were applied in real-world settings. For example, single-turn models had a hard time calculating a patient’s age when given the birthdate and the current date. The new model can assess complex criteria and provide an accurate, sophisticated rationale for its decisions.
Can you tell us more about how Inato patient pre-screening works?
We explain our approach in detail in the new paper for anyone who wants the technical specifics, but I can provide a high-level overview. Essentially, we’ve assembled an AI agent that can quickly and accurately evaluate patients for trial opportunities.
Our pipeline starts by embedding each page of a medical record using VoyageAI’s multi-modal embedding model. This basically makes the charts easily searchable during the next steps, and is crucial to ensure we can retrieve the right information from lengthy patient records. From there, the AI conducts a two-step assessment for each patient.
First, it evaluates whether a patient is relevant at all for the trial by checking whether the main condition targeted by the trial appears in the record, or is suggested by the patient’s symptoms. For example, the AI will quickly rule out asthma patients for an arthritis trial. Then, if a patient is deemed eligible, the reasoning model conducts a thorough assessment against each eligibility criterion. The model provides not only its assessment for each criterion, but also a rationale and a link to the source within the patient record to back up its decision. Ultimately, this enables site staff to review the evaluation and quickly make a decision, rather than spending hours reading the full patient record.
You recently co-authored a technical paper with detailed, robust analysis of the impact for sites. Can you tell us about your findings?
I mentioned earlier that previous research demonstrated the theoretical potential of LLMs for this use case, but ultimately didn’t meet the real world needs of sites. It was really important that our application meaningfully improved the site experience, without requiring a time-consuming integration with other site systems.
To find out, we evaluated our pipeline using both a public dataset and data from our own platform. We found that the AI not only assessed patients with a high level of accuracy–actually outperforming previous state of the art methods on the public dataset–but also significantly improved site efficiency. On average, sites using Inato decide on patient eligibility in just a few minutes, which is more than 80% faster than traditional methods.
We’d heard anecdotally from sites that our new AI capabilities have been a game changer for pre-screening, but it was encouraging for us to see the hard numbers on this as we continue rolling this out to help more sites and sponsors address the enrollment bottleneck.
“Real-World Validation of a Multi-Modal Pipeline for AI-based Clinical Trial Patient Prescreening,” written by Anatole Callies, Quentin Bodinier, Philippe Ravaud, and Kourosh Davarpanah, is out now. To read the full paper, featuring more detailed insights on Inato’s AI pipeline, approach to patient pre-screening, and results, click here.