'It should be noted that the role of our algorithms is not to diagnose diseases. This is still the role of doctors. We only identify patients at high risk of a given disease', says Marek Dudziński, PhD, MD, from the medical and tech startup Saventic Health. He emphasises that the start-up has created advanced algorithms for about 50 rare diseases, for which drugs have been developed.

In the countries of the European Union, including Poland, a rare disease is defined as a condition that affects a maximum of 5 in 10,000 people. Ultra-rare diseases occur with a frequency of one in 50 thousand or smaller. Some of the se diseases affect only a few people in the world.

Dudziński reminds that - according to estimates - there are currently about 8 thousand rare diseases, of which about 800 have effective treatment. Globally, they affect about 450 million people, in Poland - about 2 million.

Rare diseases are a diagnostic challenge in every country. 'This is due to the fact that, first of all, there are a lot of them, they affect practically all organs and systems, and their symptoms are often non-specific and first raise suspicion of more common diseases', the expert explains. Therefore, it is hard to expect doctors to be aware of all eight thousand rare diseases, he adds.

As a result, making the right diagnosis of a rare disease takes an average of 5-8 years. Thanks to AI algorithms created by scientists from the Polish startup Saventic Health, this time can be significantly shortened. 'Our system allows us to shorten the time to diagnose a disease or enable diagnosis at all - by sending the patient on the right diagnostic path', Dudziński explains. This is important because, according to estimates, only one in eight patients with a rare disease is diagnosed, and a huge number of patients remain undiagnosed.

The expert emphasises that the task of the system created by Saventic Health is not to diagnose the disease, but to select those patients who have a characteristic clinical picture for a given disease. 'It is the doctor who ultimately decides whether it is worth directing the patient flagged by the system to a further diagnostic path. And if that happens, the disease may be confirmed in some of these patients', the expert explains. He adds that without using the system in a given medical facility, no one might even think about diagnostics for a given rare disease, and the patient would remain undiagnosed.

'The group of diseases that our system looks for is large. They include as many as 50 rare diseases - both metabolic diseases, such as Fabry disease, Gaucher disease, as well as haematological diseases, such as Waldenstroem macroglobulinemia, myelofibrosis, diseases of the immune system such as Castleman's disease, or hereditary angioedema', Dudziński says.

The startup team currently focuses on developing algorithms for those diseases for which there are methods of treatment. 'The general rule is that if the disease is detected earlier and therapy is implemented earlier, the treatment results are better. Therefore, we focus mainly on those diseases in which finding a patient results in a real possibility of treating them, and thus increases the patient's chances for a long life in greater comfort', the expert says.

An example of the operation of Saventic Health algorithms are the results of a study conducted in collaboration of researchers from Poland and Brazil, published in the prestigious journal Blood. In this case, the algorithm searched almost 600,000 electronic medical records of patients from a local hospital in Brazil and flagged 102 patients with potential features of Castleman's disease (a rare disease of the lymphatic system). Three of them - as assessed by Brazilian doctors - required further evaluation, which is still ongoing.

The idiopathic form of this disease (idiopathic multifocal Castleman disease - iMCD) is extremely rare and difficult to detect. It occurs in 7-10 patients per million. It is accompanied by very different symptoms related to inflammatory processes and enlargement of lymph nodes, which is often misinterpreted. Other causes of their enlargement are sought first - e.g. neoplastic diseases and diseases such as sarcoidosis, Dudziński explains. In Brazil, the median diagnostic delay for this disease is 18 months. Meanwhile, rapid diagnosis is important because it is one of the few rare diseases for which effective treatment is available, and it not only extends the patients' life, but also significantly improves its quality.

Dudziński points out that over the past year, Saventic Health algorithms have helped diagnose various rare diseases in over 300 patients. 'There are already several hundred patients who, thanks to our algorithms, received an earlier diagnosis of a difficult-to-diagnose disease, and received treatment earlier', he says.

Karol Lis, PhD, from Saventic Health emphasises that the system developed by the start-up's scientists is unique on a global scale, because it analyses the patient's full medical records. 'No one, or almost no one in the world does this. There are many companies that create algorithms that analyse imaging data - from computed tomography or magnetic resonance imaging, or focus on fragments of medical records. And that is much simpler', the expert says.

Dudziński explains that most medical data is not collected in the form of numbers or test results, but unstructured texts. 'These are medical interviews, in which doctors describe symptoms reported by the patient. These are descriptions of tests, including descriptions of imaging tests, tomography, histopathological preparations, these are observations that doctors make in hospitals and physical examinations', the specialist says.

According to him, creating an algorithm that 'understands' extensive medical documentation is very difficult. 'The variety of patients' medical descriptions is enormous. One symptom can sometimes be described in over 100 ways. In addition, medical records are often abbreviated, there are a lot of typos, and they are difficult to interpret', the researcher describes.

Thanks to computers with high processing power, the system created by Saventic Health is able to quickly scan the entire database of electronic medical records of patients in the hospital. 'It can be 500 thousand or 1 million records. And thanks to the use of AI algorithms - advanced natural language processing techniques - we can extract key symptoms and understand their context', Dudziński explains.

He emphasises that the algorithms created by the startup use all possible artificial intelligence techniques. 'This is both machine learning and deep learning, language models, but also rule systems, scoring systems. A language model alone is not enough to analyse medical records. Our experience shows that in medicine, such hybrid solutions, where we use all possible methods, work best. The precision and sensitivity of such a model are much greater', he says.

Due to their uniqueness, the Polish startup's algorithms enjoy huge interest all over the world. 'We are present on four continents, not only in Europe, including Poland, France, Germany, Sweden, Austria, Switzerland, but also in Canada, Brazil, and we are developing cooperation with Taiwan', Dudziński says.

In Poland, the company currently cooperates with 14 centers - from large clinic teams to district hospitals, clinical hospitals and private facilities. 'In general, in Poland, we can say that our algorithms work in a population of about 8 million patients. This is also the population that we use to train our algorithms, because they are not only based on medical, substantive, expert knowledge, but above all on the data of real patients, mainly Polish ones. Of course, the data we use is completely anonymous', the expert explains.

According to Dudziński, in Poland, some algorithms are financed from scientific grants (e.g. the Polish National Centre for Research and Development grant), some from grants from pharmaceutical companies, and some thanks to investors.

In Canada, the startup launched cooperation with the insurer of one of the provinces - Alberta Health Services. 'We will be implementing algorithms for four rare diseases there, with the option of further implementations. We are also implementing algorithms in Brazil, and we have plans for further implementations, for example in Panama, Colombia, and Costa Rica. We are currently preparing an algorithm for Taiwan', Dudziński says.

In his opinion, everyone benefits from implementing the startup's algorithms - primarily patients, but also insurers, taxpayers, pharmaceutical companies.

'We would like to create algorithms for all rare diseases, and not only rare ones, but due to limited resources we are not able to do it all at once. However, it should be emphasized that the process of creating algorithms is developing very dynamically and I suspect that the number of new algorithms for subsequent diseases will grow exponentially', he concludes.

Joanna Morga (PAP)

jjj/ agt/