Interview: Puraite aims to accelerate evidence synthesis with explainable AI

Contents

The medical research community produces over three million new studies every year – but who is supposed to read all of them? Even highly specialized experts can only oversee a fraction of the relevant literature. At the same time, the processing of this knowledge is lagging behind: a systematic review, i.e., the structured summary of the current state of research on a specific question, currently takes between six months and two years. By the time the results are available, the data situation is often already outdated.

The East Westphalian startup Puraite aims to accelerate these processes with an AI platform for evidence synthesis, without sacrificing scientific rigor. The platform searches databases like PubMed, filters relevant studies, extracts data, and evaluates them according to scientific criteria. In an interview, co-founder Schahin Baki explains why the manual approach is reaching its limits and how AI-supported, but methodologically sound processes can accelerate evidence synthesis. While there are similar offerings internationally from companies like "Laser AI", "Nested Knowledge", "DistillerSR", and "Covidence", Puraite is, according to its own statements, the only product of this scope developed and distributed entirely from Europe.

Can you briefly explain what Puraite stands for and what it does?

The name stands for Pure AI Technology. We automate evidence synthesis in the life sciences, i.e., in the pharmaceutical industry, medical technology, clinical care, and related fields. We help to consolidate content from studies to answer research questions.

And what is the difference compared to a good AI language model?

Linguistic elegance is not proof of scientific quality. An AI model can formulate a text that sounds like a thorough summary – but there is no systematic procedure behind it. A real review follows clear methodological steps: a pre-defined research protocol, a traceable literature search, a double independent assessment of studies, structured risk of bias assessment using recognized tools (RoB 2, ROBINS-I, QUADAS-2), an assessment of evidence certainty according to GRADE, and complete traceability of sources down to the sentence level.

International research networks such as Cochrane and Campbell have clarified in a position paper that AI can support these methodological requirements – but it does not replace them. This is precisely where Puraite comes in. We do not automate evidence, but provide the infrastructure in which a methodologically correct, compliant process can be efficiently carried out and demonstrated to regulators.

So Puraite helps to better assess the current state of research?

Exactly. This means extracting information from various studies to answer complex, fine-grained research questions – for example: How effective are GLP-1 receptor agonists compared to placebo or other antidiabetic therapies in reducing serious cardiovascular events in adult patients with type 2 diabetes mellitus? Typically, one would use various databases, screen studies for relevance, extract the data, and finally compile a document from many individual studies that provides the answer. The use cases are very broad – from medical guidelines to price negotiations for drugs. Throughout the healthcare sector, evidence is the foundation for decisions.

We deliberately excluded clinical care at first and focused entirely on industry – meaning the pharmaceutical industry, medical technology, insurance companies, and related sectors like the food industry, wherever literature research is required for regulatory purposes. Product development in our field is cost-intensive, due to AI expertise in terms of personnel as well as infrastructure. Therefore, industry is our primary focus – also because the revenue generated there should enable us to make the platform accessible to the academic and clinical sectors in the long term.

And which databases does Puraite access?

Primarily publicly accessible publications and clinical studies – via databases such as PubMed, Embase, or OpenAlex. About 50 percent of the full texts are freely accessible, the other half is behind paywalls, but the companies and institutions that work with them usually have the appropriate licenses. In addition, there is grey literature. It is crucial to capture everything systematically: Every relevant signal must be included because it can influence the answer to the research question – two clinical studies can lead to different results.

What role do guidelines play in this?

Guidelines are developed based on systematic reviews. These reviews represent the highest form of evidence. They are used to synthesize existing evidence and derive guidelines from it. The problem, however, is that systematic reviews are extremely time-consuming. Given the explosion of knowledge in medicine, information is often already outdated before it is incorporated into a guideline.

So, does this update much faster than conventional AI systems?

What is special is not so much the AI itself, but rather the concept of so-called "Living Reviews". This means that the entire process is re-initiated at regular intervals – quarterly, monthly, weekly, or even daily – to incorporate newly published articles. Because what is the state of evidence today may have changed tomorrow.

Can this also be initiated manually?

That is possible at any time. The current business model is based on a flat-rate principle: there is a licensing model that allows any number of users to be included. We consciously do not charge per seat because we want the tool to be used intensively without having to share accounts. You can think of it like an organizational account. A large pharmaceutical company, for example, has different departments that create or commission systematic reviews. We provide this company with the platform in its entirety – with versioning and task distribution.

You advertise with 95 percent sensitivity?

The process begins with searching for publications. The next step is screening: you read the title and abstract and decide – relevant or not relevant. In this step, we achieve a sensitivity of over 95 percent, which corresponds to the established target value for AI-supported screening tools in this field. The human benchmark with two independent reviewers is around 97 percent – but significantly lower with a single reviewer, as studies show. Humans are considered the undisputed gold standard in the industry, but humans also make mistakes – large studies document error rates of over ten percent in single screening. I am not claiming that AI is perfect, quite the contrary – I am only saying that this side of the debate is often given too little attention.

Is this an agent model?

The platform is primarily modular, in line with scientific methodology. The entire process of evidence synthesis is currently extremely manual: several people work on it in parallel, exchange tables, use different tools – very fragmented. Our platform is intended to be the common evidence infrastructure. You can use AI – explainable AI – but you can also perform every step completely manually, all within one platform without media breaks. This is a significant advantage: no more distributed Excel tables, but versioning, clear task distribution, and collaborative work interface. The AI can be used for the entire process or only for individual steps with sub-agents – as needed.

The product is initially in English. How would that be for the German market? Other databases, adapted models?

That is a very good question. Fortunately, evidence synthesis takes place almost exclusively in English – in the vast majority of cases. Even when used for clinical guidelines in Germany, these are based on publications from various countries. Therefore, we proceeded with English from the outset, as publications in this field almost always appear in English. Translating the platform itself would not be a problem – however, none of our users have expressed a need for it so far, as everyone works in English anyway. And they are always very international teams located in America, Ireland, Germany, or Japan.

Your customers are from different countries?

Yes. However, the majority of our customers come from the USA, Canada, and the United Kingdom. I often had the feeling that in Germany people want to be innovative, but nobody wants to be the first. And in a sense, this is our unique selling proposition and the reason why we started Puraite from Germany: We are the only team worldwide developing this solution entirely from Europe, without branches in the USA or elsewhere.