Background
Extensive experience in combining such different professions – a doctor and an IT specialist, allows you to see opportunities where others do not see it. Scientific research and more than ten years of experience in the creation of telemedicine systems have gradually led to a new direction – the use of Artificial Intelligence (AI) in Healthcare. Artificial Intelligence technology holds great promise in Health. For example, everyone knows examples of the use of AI in the diagnosis of X-ray images or the diagnosis of skin diseases, in histology, etc. This industry is developing very quickly and many interesting innovations await us in the very near future.
This project focuses on the use of AI to assess patient complaints. In other words, it is necessary to replace the routine work of a doctor in the early stages of diagnosis, to help primary care (family doctors). I would call it “routing”, when it is necessary to find out to which medical field a given case belongs and in the competence of which specialist. This problem is quite relevant, because the heavy workload on the family doctor reduces the possibility of effective communication with the patient. This was most evident in the era of the COVID-19 pandemic. With the help of AI, we can transfer routine work to machines. Moreover, AI does not know fatigue, and in some areas of medicine, its capabilities exceed those of humans. An important fact is that AI is constantly “learning”. What it can do today, cannot be compared with its capabilities tomorrow.
Idea
The COVID-19 pandemic has helped identify a number of health problems. Its negative impact often leads to high, sometimes critical, burdens on the health care system. I would consider this in the context of primary care. High anxiety, uncertainty and sometimes fear in society, lead to an increase in visits to primary care physicians. Patients interpret their feelings in different ways and require their evaluation, which creates an additional burden on the doctor. Of course, this has negative consequences, both for the patient, the doctor, and for Health in general.
AI can be useful in this situation. Its capabilities make it possible to automate the initial assessment of patient complaints and determine the likely belonging of these complaints to the competence of a specific specialist. AI can recommend further actions, necessary for diagnosis, for example, suggest a typical list of examinations, that a doctor will need or suggest the nearest clinic, where such specialists are available. The above possibilities are based on the AI’s ability to correctly assess the patient’s complaints.
Solution
To assess patient complaints, four Neural Networks (NN) are used, training on two different datasets. In general, data from 360 thousand records were used for training. They were categorized according to the 28 most in-demand medical specialties. The models are based on variations of LSTM and CNN neural network architectures, as well as the capabilities of BERT and Transformers. For natural language processing and other NLP tasks, the Spacy library was used.
The question arises – why not one model, but four? The point is, relying on the predictions of a single, even well-trained model can be risky. We adhere to the principle that information about one object, obtained from several different sources, will be much more reliable. In our case, these are three different NN. The fourth model is used as a control, in the aggregate of the results of the three previous ones. It uses original NERs, the number and variety of which, is constantly increasing during operation. All neural networks, as information accumulates, are regularly trained on updated data, which increases the likelihood of predictions.
The interface uses the Russian language, because the neural networks learned from Russian-language datasets. To search for clinics, we use a database of 8250 Clinics and Medical Centers in Ukraine, that have concluded contracts with the NHS (National Health Service). Of course, the database will be expanded by leading clinics and doctors in other countries.
The application is currently in test mode. You can try it at this link.
Next steps…
Diagnostics is the road to the unknown, where the doctor can meet many intersections and forks. And along the way, any additional information will be helpful. It is for this purpose that some redundancy in the program was provided. This was described above. But this approach has provided additional data, that are useful for differential diagnosis. In the very near future, it will be possible to obtain additional information, in addition to recommendations on the specialty of a doctor.
Further, the possibility of voice input of information is planned. This is primarily to minimize grammatical errors in the text. It is planned to add the ability to receive the service in Ukrainian and English.
Collaboration
If you are interested in this direction, if you see yourself in this process, be sure to let us know about it. We are ready to cooperate. Please write your suggestions to the contact e-mail.
