Cardiac amyloidosis is a rare disease characterized by the accumulation of plaques of the amyloid protein in the heart, making it hard (which is why it is also called stiff heart syndrome) and may require a pacemaker or be the cause of death. It has some treatments, but they are very expensive, says Rafael Vidal, a member of the Spanish Society of Cardiology, who believes that it is surely an underdiagnosed pathology, which prevents treating it well. An artificial intelligence development by the company Sopra Steria, which has been tested with the medical records of more than 11,000 patients over 65 years of age who were treated at the San Juan de Dios de León hospital between 2009 and 2019, has allowed the detection of 20 new Biological markers that may indicate that this is the true disease of those affected. The work has been published in the International Journal of Enviromental Research and Public Health.

Marlon Cárdenas, responsible for Data Science at the developer, says that broadly speaking, what has been done is, from patients who were diagnosed, to remove those markers associated with the disease and then, by looking for them in the others medical histories, see if the disease was also present. It is a dynamic process in two rounds. They started from the 60 indicators that the doctors indicated to them (what in a rare disease and with various manifestations are called red flags or red flags), and when reviewing the data in a first screening they detected about 140 possible cases. But when refining the search with the addition of another 20 biomarkers that were found during the process, the sample remained at 90. Cárdenas admits that it could have been refined more, but in such a rare disease (90 possible cases of more than 11,000 people ) ran the risk of arriving at findings of no statistical value.

The machine learning process is the most relevant of this work
This way it sounds very simple, but it is necessary to highlight the biggest problem, to which the developers had to dedicate 60% of the time, according to Cárdenas: that the digitized medical records from which it was started are not uniform. There is no single outline for writing them. “The analytics are all the same, but not the comments of the doctors. The patient who only makes one visit is not the same as the one who is seen in the emergency room or has a longer process, ”says the computer scientist. These annotations in the history are a very important source of data, but more difficult to process, because they are written in natural language, free, and you have to teach the machine to read them. A tabulated information, which is found with a simple program, is not the same as that which is in the middle of a paragraph of annotations. In addition, since they used data from a decade, even coded annotations had changed over time.

The cardiologist Vidal agrees that this machine learning process is the most relevant part of this work, and gives as an example that arterial hypertension, for For example, it may appear in the history as “HBP, hypertension, HT … etc.” And you have to code the program so that it looks for it in all cases and scales it. “The idea is good,” says the doctor, who gives as an example that this pathology has often manifested itself before in other parts of the body, and the affected person has had to undergo surgery for the carpal tunnel or for stenosis in the vertebral canal or a pacemaker had to be implanted early, at age 70 when it is normal to use these devices from the age of 80. The objective would be for the artificial intelligence system to review the patient’s medical history, and warn that they meet the criteria to have amyloidosis. In fact, the article includes a variety of diagnoses associated with this pathology, from the need to use anticoagulants to the appearance of sores on the heels. As you can see, symptoms so varied that they can mislead the diagnosis.

Vidal does not doubt that this type of development will be incorporated into clinical practice, especially as the histories are coded in an increasingly uniform way. To help these processes, the doctor indicates that the Spanish Society of Cardiology is developing guidelines, a kind of dictionary, so that all specialists use the same terms to refer to the same situation. For example, that acute myocardial infarctions appear on all forms as AMI, something similar to what they already have in Sweden, which is why they have a huge amount of information, when in Spain there is not even a record of these events. This will facilitate the work of these artificial intelligence programs in the future.

Developers trust that this type of applications will serve for clinical practice (in this case anonymized data was used so that patients could not be called later suspects to consultation), and propose applying it to much more frequent situations, such as looking for markers that allow predicting how a hip replacement will evolve or who is more prone to suffer it. Also, for example, for the treatment of atopic skin.

Similar approaches have been followed, for example, for pediatric check- ups, and IBM has announced that it has projects underway for the early diagnosis of Alzheimer’s through the study of speech markers, and study of the relationship between changes in the sense of smell and diseases such as covid-19, for example.