The ERACLITO Project Between Clinical Opportunities, Risks, and New Skills

At a time when artificial intelligence is entering clinical practice at an increasingly rapid pace, the Eraclito project was created to explore how this transformation may specifically affect Physical and Rehabilitation Medicine. Not merely as a technological innovation, but as a paradigm shift: a new way of interpreting data, building personalized pathways, supporting clinical reasoning and, potentially, giving back more time and quality to the relationship with the patient.

As Dr. Molteni stated, “the topic is extremely hot.” And there is more. Artificial intelligence changes from week to week: for this reason, as Professor Zampolini observed, rather than offering definitive answers, it is necessary to build cultural, operational and educational tools that enable clinicians to use it consciously.

The ERACLITO Project

As Professor Mauro Zampolini explained, Eraclito is a project developed within the intersociety group SINFER and SIRN, with the aim of producing a guidance document for the implementation of artificial intelligence in Physical and Rehabilitation Medicine. The acronym stands for Artificial Intelligence in Rehabilitation for Conscious Involvement on Clinical Opportunity and Adoption: a formula that already contains the core of the initiative, namely promoting the use of AI in a way that is conscious, clinically useful, ethically grounded and safe.

The group’s first output was a guidance document, freely available in the Giornale Italiano di Medicina Riabilitativa, designed to provide operational guidelines and support professionals, patients and decision-makers in a phase in which the use of artificial intelligence is already widespread, but often not yet governed with full awareness.

There are many potential applications: from personalized treatment plans to early diagnosis, from prognosis prediction to the integration of clinical and laboratory data and data from sensors; through to decision support, robotics, assistive technologies and the operational management of workflows. This last aspect, Zampolini emphasized, is crucial: artificial intelligence is not only about “what” can be done, but also about “how” it can be integrated into daily practice without compromising safety, privacy, professional responsibility and clinical quality.

The Eraclito group is also working on a survey to map the current landscape, identify information needs, foster scientific networking and progressively develop more structured guidance. The aim is also to provide decision-makers with a competent voice at a time when European regulations on artificial intelligence are entering their more operational phase.

The risk: From hallucination to de-skilling

For Professor Zampolini, the first danger is so-called sycophancy: the tendency of artificial intelligence systems to be accommodating, confirmatory, kind and non-judgmental. This may appear to be a positive trait, but it can become problematic, especially when AI is used as psychological support or as an interlocutor in situations of vulnerability.

The second risk is semantic pareidolia: the system’s ability to produce formally convincing answers, with perfect syntax and an authoritative tone, even when the content is not entirely correct. In other words, the system can construct a response that “sounds right” even when it is not fully accurate. In the medical field, this represents a critical issue, because inaccurate information can influence both clinicians and patients.

Alongside this is the issue of hallucinations, namely invented or unfounded responses, as well as bias, which may be embedded in systems through the data on which they have been trained.

But one of the most relevant points in the presentation concerned de-skilling. If professionals delegate reasoning to artificial intelligence too quickly, they risk progressively losing their ability to analyze, interpret and make decisions. Professor Zampolini referred to studies in the medical field showing that AI-assisted use, if not properly governed, can reduce operational skills when technological support is removed. The risk, especially for younger generations, is not only de-skilling, but even “no-skilling”: failing to fully develop certain skills because they are replaced too early by automated support.

How to protect against these risks: prompts, rag and vertical systems

To reduce these risks, Zampolini identifies three levels of use.
The first is prompt engineering, namely the ability to formulate precise, contextualized and structured requests. It is not enough to ask generically how stroke rehabilitation works: it is necessary to define the role the system should take on, the clinical context, the type of sources to be used, the required level of accuracy and the limits within which it must operate.

The second level is represented by RAG systems and tools such as NotebookLM, which make it possible to upload selected sources — articles, guidelines, slides, audio, video and documents — and query the system only on that body of information. In this way, AI does not respond on the basis of the entire information universe on which it was trained, but works within a controlled perimeter. Professor Zampolini gave the example of his own physiotherapy lectures, transformed into a searchable environment for students, and of the European position papers of the UEMS PRM Section, uploaded to obtain answers consistent with those sources.

The third level is that of vertical systems, trained on peer-reviewed scientific literature, guidelines, systematic reviews and validated documents. These are tools that reverse the traditional relationship with the literature: instead of starting with a PubMed search and manually selecting articles, the system can produce a reasoned synthesis, connected to the sources, which the physician can then explore in greater depth.

AI and the humanization of care

One of the most interesting aspects of the presentation concerned the use of artificial intelligence to improve, rather than weaken, the doctor-patient relationship. Today, Professor Zampolini observed, the consultation is often mediated by a screen: the doctor looks at the computer, types, takes notes and organizes information. An AI system could instead act as a “clinical secretary”, recording and reorganizing the conversation into a structured report, freeing the professional from part of the bureaucratic burden and restoring centrality to eye contact, listening and the relationship itself.

The key issue, however, is privacy. The use of general-purpose systems with clinical data, even when anonymized, is extremely delicate. But there are two possible paths: enterprise solutions, with controlled servers and no use of data for training, or local models installed on the clinician’s computer, isolated from the internet and therefore safer for certain basic functions, such as reorganizing clinical texts.

From the logic of prompts to the age of agents

The final part of the presentation introduced a further leap: the agentic era. This is no longer just about querying a system, but about building agents capable of carrying out specific activities, planning, observing, correcting themselves and achieving defined objectives. In medicine, and particularly in rehabilitation, these agents could deal with medical history-taking, differential diagnosis, pharmacological monitoring, organization of the clinical pathway or support for assessment.

But precisely because they are more autonomous, agents require even clearer limits: it is necessary to define what they should do, where they should stop, which success criteria they must meet and what level of human supervision is required.

The ultimate goal, for Professor Zampolini, is to move towards a new clinical humanism: using artificial intelligence to automate what can be automated, enhance professional reasoning and improve the relationship with the patient.

The Q&A: The exchange between Franco Molteni and Mauro Zampolini

The professor’s presentation was followed by a Q&A with Dr. Franco Molteni, an excerpt of which is reported below.

Franco Molteni: “We are all very pleased if artificial intelligence can take care of boring tasks for us. However, there is a crucial point: decision support requires competence on the part of the person using it. At the moment, there is still no defined level of professional competence that enables artificial intelligence to be used consciously. Is this a solvable problem?”

Mauro Zampolini: “More than solvable, we have to solve it. As a group, we need to focus strongly on this. European legislation provided for basic training on artificial intelligence, but in practice very little has been done. Today there is a great deal of ‘grey’ use: everyone uses it, but few really know what it means and what is happening”.

Franco Molteni: “Does this mean that artificial intelligence is already being used today, even if not always openly declared?”

Mauro Zampolini: “Yes. Formally, today it is not possible to make a clinical decision supported by artificial intelligence. In practice, however, just as we go online or use PubMed to look for information, many people use artificial intelligence as if it were an advanced version of PubMed. So our clinical decisions, even if not supported, are still influenced by what artificial intelligence tells us. The point is to avoid remaining in this grey area and move towards specific systems, trained on validated medical content”.

Franco Molteni: “Can you give a concrete example of clinical use?”

Mauro Zampolini: “I once had a patient in the outpatient clinic with unusual symptoms that I could not interpret. At one point she told me she had Niemann-Pick disease, which of course is not something we encounter every day. So, after telling the patient, I used a specific system to obtain a precise overview of the disease. At that point, those symptoms became explainable and I was able to interact better with her and plan the pathway”.

Franco Molteni: “What would have happened without this support?”

Mauro Zampolini: “I would have had to tell the patient that we would see each other again later, then I would have gone away to study the disease, losing hours to probably reach the same kind of synthesis. The point, however, is to avoid asking delicate medical questions to general-purpose systems, such as Gemini, ChatGPT or Claude. There may be something hallucinatory in those answers. If it is marginal, it is not a problem; if it is relevant, it becomes a problem”.

Franco Molteni: “In medicine, speed also matters: how quickly we carry out certain operations and make decisions. But clinical experience teaches us that speed is not always useful, especially when we deal with complex systems. Sometimes patience is almost more important. Will we also have to learn how to manage this balance between speed and patience in our relationship with artificial intelligence?”

Mauro Zampolini: “Certainly. If I start using these systems, even the most accurate ones in the medical field, and directly apply what they tell me, I quickly risk de-skilling. It is a bit like the Google Maps effect or the calculator effect. If you lose the ability to reason, to process things in a complex way, to analyze nuances, you lose everything”.

Franco Molteni: “When we query artificial intelligence, then, we need to understand not only how its mental map works, but also what our own mental maps are when we enter this logic of interaction. Training begins to become substantial. Are we ready to provide training in order to develop complex mental maps?”

Mauro Zampolini: “We absolutely have to do it. It is urgent. At the same time, we need to move beyond the idea that these systems are merely ‘statistical parrots’. The most recent studies are showing that artificial intelligence systems self-organize through semantic networks. This does not mean attributing consciousness to them, but recognizing that there is something more complex than simple statistical prediction”.

Franco Molteni: “This becomes even more relevant when we talk about agents. At that point, the system takes on different characteristics: I am no longer asking someone to do something, but working with a system that can act. Agents, in medicine and particularly in rehabilitation medicine, could do a great deal. How do you see them?”

Mauro Zampolini: “It is a trend that cannot be stopped, because the use of agents is so advantageous. One can imagine an agent that acts as a physiatrist, but not simply as a physiatric prompt: an agent that behaves like a physiatrist, does things and then self-corrects if it makes a mistake. This opens up an entire world”.

Franco Molteni: “What is the main risk?”

Mauro Zampolini: “The risk is that agents do whatever they want if boundaries are not clearly defined. We need to say what they must do, where they must stop and what the success criterion is. Without rules, the agent acts as it wishes”.

Franco Molteni: “In clinical management, we can truly be helped, provided that we enter the information needed to perform complex reasoning into the medical record. The more we realize how much artificial intelligence can intervene as a support tool, the more we will understand the importance of how we write the medical record. Text will also become an integral part of the information. What observations can you make with regard to text analysis?”

Mauro Zampolini: “I carried out some tests, respecting privacy criteria, by copying the clinical diary in anonymized form and asking the system to summarize the main disease, secondary conditions, problems and priorities of the day. The system did this very reliably”.

Franco Molteni: “Have you encountered cases in which this type of analysis was particularly useful?”

Mauro Zampolini: “In one case, the system indicated lower-limb Doppler ultrasound as the absolute priority because of the risk of venous thrombosis. I went back to check the medical record, verified that the reasoning made sense, requested the Doppler examination, and the patient had bilateral femoral deep vein thrombosis. It is an anecdotal example, but it shows how artificial intelligence can work as an assistant that helps us avoid missing important elements.”

Franco Molteni: “Since many people may think of copying the clinical diary into an artificial intelligence system, it is worth saying a few words about data ownership, privacy and consent from the administration. It is easy to do, but it is also extremely delicate”.

Mauro Zampolini: “Doing this, even in anonymized form, with general-purpose systems is not acceptable. The data enters the system and may be used for training. There are two possible paths: using corporate systems, such as enterprise versions with controlled servers and no training on the data, or using local models installed on one’s own computer and isolated from the network”.

Franco Molteni: “So even less powerful systems can be useful, if used in the right way?”

Mauro Zampolini: “Yes. For some basic activities, such as interpreting, reorganizing and summarizing a clinical text, a large amount of computing power is not necessarily required. If you need to do more advanced things, you need more complex systems, but simpler local models may be enough to systematize a medical record”.

Franco Molteni: “Describe more clearly what a RAG is”.

Mauro Zampolini: “A RAG is a system in which I build specific knowledge. The most tangible example is NotebookLM: I create a notebook, upload texts, PDFs, audio, video and even YouTube videos, which are transcribed, and then I query the system only on that material. In this way, artificial intelligence responds within a defined perimeter”.

Franco Molteni: “Can this model also be applied to community-based medicine or regional care pathways?”

Mauro Zampolini: “Yes. It is the same concept as the MIA project for general practitioners: guidelines, documents and regional diagnostic-therapeutic care pathways are uploaded, and the doctor can query the system on the basis of those sources. If a doctor in Lombardy asks what should be done according to the pathway, the system personalizes the answer to that context. This greatly reduces the risk of hallucinations”.

Franco Molteni: “How much will artificial intelligence change the way we use robotic devices in rehabilitation medicine?”

Mauro Zampolini: “I think it will do so substantially. It will change the control of robotic phases, the capacity for self-adaptation and personalization. But what will really change the game is the robot’s ability to explore the external world and learn from the external world”.

Franco Molteni: “How will the robot be able to learn from the external world?”

Mauro Zampolini: “Today, the robot is no longer trained only with typed algorithms. The system observes how the world works and learns, somewhat like mirror neurons. This will substantially change the way the robot interacts with the environment”.

Franco Molteni: “We are facing epochal changes. The need for the intersociety group is to activate initiatives that increasingly expand training and the exchange of experiences at a widespread level. It will be important to discuss these issues among people who are using artificial intelligence in rehabilitation medicine, in order to understand what is emerging. Do you have any appointments to propose to those who followed this webinar?”

Mauro Zampolini: “We will carry out the survey, and then it would be useful to build a series of webinars, both synchronous and asynchronous. I also strongly believe in short content, perhaps ten-minute videos on YouTube, to explain specific topics. It is a form of cultural dissemination that can work”.

Franco Molteni: “I think today was a way to open up the topic. Scientific societies will have to step in using different approaches compared with the past. Face-to-face discussion will remain a fundamental element for advancing knowledge on these issues. Webinars are useful, but we must also find opportunities to meet in person and discuss in depth problems that we can then further develop online”.