How medical chatbots are arranged - let's look at the example of the DOC + bot
History taking is an essential step in examining a patient. The doctor interrogates the patient about pathologies, surgical interventions, injuries received, the course of the disease. Often he writes down the information that the patient tells him, but sometimes he asks to fill out a special questionnaire.
Usually, taking anamnesis takes up to 60% of the consultation time and lasts about 15 minutes. Therefore, in the West, in connection with the high cost of working time of the doctor, there is a practice in which the nurses conduct the initial pre-medical examination of the patient. They fill out special questionnaires on which the doctor further relies. However, this approach only shifts the process of collecting anamnesis from one person to another.
Therefore, today technologies are developed and implemented on the basis of artificial intelligence in the format of chat bots, which reduce the history taking time several times and reduce the likelihood of error. This saves the resources of the clinic and "unloads" nurses and doctors, giving the latter the opportunity to further study the symptoms of the disease and make a more accurate diagnosis.
')
About what chat bot solutions already exist today, what they can do, and how the DOC + chat bot works, let us describe further.
“Write to the doctor around the clock in the DOC + application ”
In the world of medical services there are many different symptom checkers. Most of them are based on the traditional question-answer system - they do not have artificial intelligence, and all asked questions are “scripted”. Such decisions may collect complaints and anamnesis of patients in the most common cases. An example of such a system might be WebMD. The application collects data and shows the user information about the diseases associated with the specified symptoms.
With the development of technology and the accumulation of medical data, it became possible to implement smarter questionnaires. They adapt to the situation, changing the "line of conduct" depending on the responses of users.
Of these chat bots, you can select the application of German developers - ADA. The system formulates personalized questions based on information that the patient tells her. After that, he asks to clarify a few points, in particular, related to the localization of the source of pain (for example, “do painful sensations arise behind the eyeball when the eyes move”). Based on the data obtained, the system offers 4–5 possible diagnoses. If necessary, the application will connect the patient with a real doctor in a chat for consultation.
The Babylon English telemedicine service works in a similar way. After analyzing the patient's answers to questions, the application says what to do next: go to a pharmacy, book a telemedy consultation, go to an appointment with a general practitioner or a specialized clinic.
The chat bot application is even being tested by the National Health Service of England (NHS) as the first line for patients requiring medical care or consultation. More than a million people in North London were able to access the Babylon AI system instead of the call center operator NHS. Operators, though supervised by general practitioners and nurses, may not have knowledge in a particular medical field. In the six months of the experiment, Babylon showed higher accuracy and speed in making the diagnosis — on average, one patient “left” a minute (for comparison, nurses and doctors spent from 2 to 3 minutes to do this).
Another smart system on the market is Your.MD. She works in instant messengers and allows the patient to enter symptoms in a solid text. Your.MD asks additional questions about possible associated symptoms. As a result, the bot issues one diagnosis with a description and treatment options or a warning about the need to call an ambulance.
In the Russian market, such solutions are also being introduced - one of them is the DOC + chat bot. Intellectual algorithms of our application collect symptoms, medical history and prepare data for the doctor. A chat bot can ask whether a person has been ill for a long time, clarify whether he has taken any drugs, etc.
So far, the accuracy of diagnosing all such systems is far from ideal. This is mainly due to the fact that it is impossible to make an accurate diagnosis based on complaints from the patient alone. Other information is needed for this — inspection data, analyzes, instrumental studies, etc.
However, existing algorithms allow you to train the system to ask the right questions and collect more complete information about the disease, compared to what the patient describes. That is why we decided to start laying in our bot only this functionality.
As part of the Data Science direction, we at DOC + are developing an infrastructure for machine learning algorithms. The bot is the closest product to the end user, but the technologies underlying it are used in our other solutions (we wrote about some of them in previous articles).
The basis of the bot - algorithms, trained on the data of anonymous electronic DOC + patient records , as well as information from open reference books and medical data bases. At the same time, since we are not talking about simple recommender algorithms, but about a complex medical system, whose work is connected with the health of people, we have involved practicing doctors for its development.
In total, more than 30 doctors took part in the project: they helped develop rules for the NLP system (after all, patients can describe the same symptom in dozens of different ways), developed lists of clarifying questions for the most popular symptoms, tested the bot itself and gave (and continue to give ) feedback about his work.
The key goal of the bot is to reduce the time that the patient and the doctor spend on consultation. The application collects anamnesis in 1-3 minutes, which is much faster than answering similar questions to a doctor.
It is important to understand that the bot does not make decisions, does not diagnose and does not prescribe treatment - the last word always remains for the doctor. Therefore (this is critically important in the field of medicine) the intellectual system does not bear any risks for the patient. In the worst case, the application will ask the user about the complaint, which he does not have, having spent an extra few seconds.
The second goal of the bot is to improve the quality. The system asks a lot of questions, aggregating a large amount of information and reducing the risk that the doctor will miss something.
The third goal is a more structured collection of data on complaints and symptoms. If earlier they were gathered in a text form in a free form, then thanks to the bot they are structured and have a context, since the patient answers additional questions for most of the symptoms. Structured data allows us to improve the quality of all our machine learning algorithms.
The “brains” of a bot consist of 4 key parts: a natural language processing (NLP) system, a recommender system, a module for determining the group of diagnoses, and a module that forms additional questions about the history of life and disease.
Natural language processing. This block is based on the DLP + NLP system, which we wrote about in more detail earlier. In the first question, the bot asks the patient to describe what is troubling him. The algorithm analyzes the information received, structures the data about the symptoms and asks clarifying questions. For example, if the user mentioned an elevated temperature, but did not indicate a specific value, the bot will ask him to enter. If the patient said cough, the bot will check if it was dry or wet. After that, the collected data is transmitted to the input of the recommendation system.
/ Bot asks patient to specify temperature
Note that plain text introduces only primary complaints. Virtually all other issues to speed up the process are implemented in a special interface - buttons, drums, lists with multiple selections, etc.
Recommender system. It is based on the neural network. The process of interviewing and determining the symptoms of a disease resembles systems used, for example, in e-commerce or appraisal services.
Trained in 100,000 real complaints from the DOC + EHR, our module of the recommendation system is able to find symptoms that often occur together. For example, if the patient said about cough and runny nose, the bot would check with him about the sore throat and fever, and also ask clarifying questions about these symptoms.
Definition of a group of diagnoses. When questions about all the most likely symptoms are already asked, the third stage of the survey is included. As noted, making a diagnosis on complaints alone is problematic. But in many cases this is enough to establish with great precision a group of diseases - for example, diseases of the respiratory system or diseases of the gastrointestinal tract. This is exactly what the classifier based on the Gradient Boosting and Random Forest algorithms does .
Within each group, we know in advance the most important symptoms for making a diagnosis and we can clarify their presence with the patient. For example, a headache is not the most common symptom of a respiratory illness, and the recommender system may not be asked about it, but it may be important to distinguish flu from the common cold. Thus, the system asks questions about not the most frequent, but nonetheless important symptoms, complementing the information collected earlier.
Additional questions on the history of life and disease. When all complaints are collected, the patient answers several questions - when the illness began, whether he took any drugs, etc. All of them are formulated based on the analysis of online consultations and interviewing of doctors.
/ An example of the work of the bot in collecting life history
The system may also ask the patient to attach test results or other information useful to the doctor, for example, a history of life: allergies, chronic diseases, traumas and operations. All of these queries are selected from detailed references.
At the end of the bot's work, all information in a structured form gets into the interface of the doctor, where he can familiarize herself with it immediately before the consultation. These data are automatically transferred to the medical card, also reducing the time spent by the doctor to keep records.
We will expand the functionality and scope of the chat bot. In our plans to create a system that will be useful in consultation with highly specialized specialists, and in the work of the contact center.
In the future, our bot will learn not only to collect complaints, but also to pre-route patients, automatically detect critical situations when an ambulance call is needed, and also identify risk groups for chronic diseases in which patients fall.
Additional reading: some useful articles from our blog “ Just Ask ”:
Usually, taking anamnesis takes up to 60% of the consultation time and lasts about 15 minutes. Therefore, in the West, in connection with the high cost of working time of the doctor, there is a practice in which the nurses conduct the initial pre-medical examination of the patient. They fill out special questionnaires on which the doctor further relies. However, this approach only shifts the process of collecting anamnesis from one person to another.
Therefore, today technologies are developed and implemented on the basis of artificial intelligence in the format of chat bots, which reduce the history taking time several times and reduce the likelihood of error. This saves the resources of the clinic and "unloads" nurses and doctors, giving the latter the opportunity to further study the symptoms of the disease and make a more accurate diagnosis.
')
About what chat bot solutions already exist today, what they can do, and how the DOC + chat bot works, let us describe further.
“Write to the doctor around the clock in the DOC + application ”
Who develops chat bots
In the world of medical services there are many different symptom checkers. Most of them are based on the traditional question-answer system - they do not have artificial intelligence, and all asked questions are “scripted”. Such decisions may collect complaints and anamnesis of patients in the most common cases. An example of such a system might be WebMD. The application collects data and shows the user information about the diseases associated with the specified symptoms.
With the development of technology and the accumulation of medical data, it became possible to implement smarter questionnaires. They adapt to the situation, changing the "line of conduct" depending on the responses of users.
Of these chat bots, you can select the application of German developers - ADA. The system formulates personalized questions based on information that the patient tells her. After that, he asks to clarify a few points, in particular, related to the localization of the source of pain (for example, “do painful sensations arise behind the eyeball when the eyes move”). Based on the data obtained, the system offers 4–5 possible diagnoses. If necessary, the application will connect the patient with a real doctor in a chat for consultation.
The Babylon English telemedicine service works in a similar way. After analyzing the patient's answers to questions, the application says what to do next: go to a pharmacy, book a telemedy consultation, go to an appointment with a general practitioner or a specialized clinic.
The chat bot application is even being tested by the National Health Service of England (NHS) as the first line for patients requiring medical care or consultation. More than a million people in North London were able to access the Babylon AI system instead of the call center operator NHS. Operators, though supervised by general practitioners and nurses, may not have knowledge in a particular medical field. In the six months of the experiment, Babylon showed higher accuracy and speed in making the diagnosis — on average, one patient “left” a minute (for comparison, nurses and doctors spent from 2 to 3 minutes to do this).
Another smart system on the market is Your.MD. She works in instant messengers and allows the patient to enter symptoms in a solid text. Your.MD asks additional questions about possible associated symptoms. As a result, the bot issues one diagnosis with a description and treatment options or a warning about the need to call an ambulance.
In the Russian market, such solutions are also being introduced - one of them is the DOC + chat bot. Intellectual algorithms of our application collect symptoms, medical history and prepare data for the doctor. A chat bot can ask whether a person has been ill for a long time, clarify whether he has taken any drugs, etc.
So far, the accuracy of diagnosing all such systems is far from ideal. This is mainly due to the fact that it is impossible to make an accurate diagnosis based on complaints from the patient alone. Other information is needed for this — inspection data, analyzes, instrumental studies, etc.
However, existing algorithms allow you to train the system to ask the right questions and collect more complete information about the disease, compared to what the patient describes. That is why we decided to start laying in our bot only this functionality.
How the DOC + chat bot works
As part of the Data Science direction, we at DOC + are developing an infrastructure for machine learning algorithms. The bot is the closest product to the end user, but the technologies underlying it are used in our other solutions (we wrote about some of them in previous articles).
The basis of the bot - algorithms, trained on the data of anonymous electronic DOC + patient records , as well as information from open reference books and medical data bases. At the same time, since we are not talking about simple recommender algorithms, but about a complex medical system, whose work is connected with the health of people, we have involved practicing doctors for its development.
In total, more than 30 doctors took part in the project: they helped develop rules for the NLP system (after all, patients can describe the same symptom in dozens of different ways), developed lists of clarifying questions for the most popular symptoms, tested the bot itself and gave (and continue to give ) feedback about his work.
Why we created a bot
The key goal of the bot is to reduce the time that the patient and the doctor spend on consultation. The application collects anamnesis in 1-3 minutes, which is much faster than answering similar questions to a doctor.
It is important to understand that the bot does not make decisions, does not diagnose and does not prescribe treatment - the last word always remains for the doctor. Therefore (this is critically important in the field of medicine) the intellectual system does not bear any risks for the patient. In the worst case, the application will ask the user about the complaint, which he does not have, having spent an extra few seconds.
The second goal of the bot is to improve the quality. The system asks a lot of questions, aggregating a large amount of information and reducing the risk that the doctor will miss something.
The third goal is a more structured collection of data on complaints and symptoms. If earlier they were gathered in a text form in a free form, then thanks to the bot they are structured and have a context, since the patient answers additional questions for most of the symptoms. Structured data allows us to improve the quality of all our machine learning algorithms.
Bot components
The “brains” of a bot consist of 4 key parts: a natural language processing (NLP) system, a recommender system, a module for determining the group of diagnoses, and a module that forms additional questions about the history of life and disease.
Natural language processing. This block is based on the DLP + NLP system, which we wrote about in more detail earlier. In the first question, the bot asks the patient to describe what is troubling him. The algorithm analyzes the information received, structures the data about the symptoms and asks clarifying questions. For example, if the user mentioned an elevated temperature, but did not indicate a specific value, the bot will ask him to enter. If the patient said cough, the bot will check if it was dry or wet. After that, the collected data is transmitted to the input of the recommendation system.
/ Bot asks patient to specify temperature
Note that plain text introduces only primary complaints. Virtually all other issues to speed up the process are implemented in a special interface - buttons, drums, lists with multiple selections, etc.
Recommender system. It is based on the neural network. The process of interviewing and determining the symptoms of a disease resembles systems used, for example, in e-commerce or appraisal services.
Trained in 100,000 real complaints from the DOC + EHR, our module of the recommendation system is able to find symptoms that often occur together. For example, if the patient said about cough and runny nose, the bot would check with him about the sore throat and fever, and also ask clarifying questions about these symptoms.
Definition of a group of diagnoses. When questions about all the most likely symptoms are already asked, the third stage of the survey is included. As noted, making a diagnosis on complaints alone is problematic. But in many cases this is enough to establish with great precision a group of diseases - for example, diseases of the respiratory system or diseases of the gastrointestinal tract. This is exactly what the classifier based on the Gradient Boosting and Random Forest algorithms does .
Within each group, we know in advance the most important symptoms for making a diagnosis and we can clarify their presence with the patient. For example, a headache is not the most common symptom of a respiratory illness, and the recommender system may not be asked about it, but it may be important to distinguish flu from the common cold. Thus, the system asks questions about not the most frequent, but nonetheless important symptoms, complementing the information collected earlier.
Additional questions on the history of life and disease. When all complaints are collected, the patient answers several questions - when the illness began, whether he took any drugs, etc. All of them are formulated based on the analysis of online consultations and interviewing of doctors.
/ An example of the work of the bot in collecting life history
The system may also ask the patient to attach test results or other information useful to the doctor, for example, a history of life: allergies, chronic diseases, traumas and operations. All of these queries are selected from detailed references.
At the end of the bot's work, all information in a structured form gets into the interface of the doctor, where he can familiarize herself with it immediately before the consultation. These data are automatically transferred to the medical card, also reducing the time spent by the doctor to keep records.
Future plans
We will expand the functionality and scope of the chat bot. In our plans to create a system that will be useful in consultation with highly specialized specialists, and in the work of the contact center.
In the future, our bot will learn not only to collect complaints, but also to pre-route patients, automatically detect critical situations when an ambulance call is needed, and also identify risk groups for chronic diseases in which patients fall.
Additional reading: some useful articles from our blog “ Just Ask ”:
- How to get medical care in Moscow
- 7 patient rights that everyone needs to know
- How to attach to the clinic in Moscow
Source: https://habr.com/ru/post/374493/
All Articles