How a young doctor’s paperwork automation system has turned into a clinical decision support system with elements of artificial intelligence.
Is a doctor capable to make correct decisions?
There must be feedback in the decision-making. This feedback should continuously improve decisions. A good doctor makes decisions, gains experience, and tries to improve, but has several human limitations and vulnerabilities
First limitation is a human brain.
The human brain is limited in making decisions in the face of many factors. There is a Miller number. The brain can process 7 (from 5 to 9) factors simultaneously. However, a patient may have several dozen of these factors. It is easy for a doctor to miss something important.
Human decision systems do not produce linear estimates of risk and benefit. The risk of loss has a more substantial influence on decision-making than the expectation of benefit. In addition, a person overestimates the probability of low-probability events and underestimates the likelihood of highly probable ones.
Different hormonal statuses in one person have various effects on brain risk-taking and trust. Therefore, one doctor with the same experience may make different decisions on different days.
Second, the doctor’s experience is flawed.
The difficulty of forming a doctor’s experience is because feedback on the results of his decisions does not come fully. Usual doctor does not track all long-term results. Due to the distorted observation of the world by incomplete feedback data, the doctor falls victim to cognitive distortions such as the “survivor phenomenon.”
Even if we assume that the doctor collects all the data about all his results from patients and the decisions made, he will receive the results much later than he made the decisions. Therefore, the influence of “delayed feedback” will appear. When experience is challenging to form because the brain no longer remembers under what conditions a decision with a given result was made.
Third, doctors learn from biased information
What about instructions and clinical recommendations according to which doctors study?
Hospitals are different, but clinical guidelines are the same. Coronary bypass graft surgery in a village and a well-developed hospital do not give the same results. However, maybe there are experienced interventional cardiologists in this village. Therefore, it is better to consider the local context when making decisions here and now.
How can we help the doctor not make mistakes?
We can overcome the problems of processing more factors, biased risk assessment, hormones, delayed feedback, and lack of objective context assessment using neural network models.
I see the development of decision-making systems like hybrid intelligence. Electronic systems must collect data from factors on which doctors make decisions, collect decisions, monitor feedback, and learn, using the correctness of the decision made in terms of some criterion as a reward parameter. Moreover, the system can assess the result according to various criteria: survival and multiple quality-of-life indicators.
What is the place of a doctor when AI comes?
The future doctor’s task is to mix artificial intelligence and compassion.
Igor Buzaev
The doctor’s task is to listen to the patient, collect factors, and ask the computer system for probabilities and recommendations.
The doctor’s task is to hear from the patient which criterion is more important for the patient. For example, when choosing CABG or PCI, the dilemma often arises: “One candy now or two candies later.” In the case of complex coronary artery disease, CABG requires the patient to have a long adventure in the hospital with incisions and a heart-and-lung machine, but it gives better long-term results. PCI gives results here now, but long-term results may be worse. The patient has the right to choose.
The doctor’s task is to weigh the computer system’s advice, taking into account each patient’s values.
Is it possible to implement this?
This prototype will be the story or how a system for automating the work of one doctor has turned into a system for monitoring long-term results of patient treatment and a clinical decision support system with elements of artificial intelligence. A 17-year project, carried out enthusiastically, without attracting third-party investments, and 80% independently.
How my clinical decision support system prototype grew up?
2001. Automation of my work as a doctor.
In 2001, when the Pentium 4 had just appeared, Microsoft had not released Windows XP, and LCD monitors were beginning to replace the old convex screens on cathode ray tubes; I graduated from university and got an internship. Back then, the usual practice was to delegate almost all of the department paperwork to an incoming intern. At that time, physicians wrote medical case reports by hand or typed the required forms using MS Word. The abundance of duplicate pages in the medical history forced repeated rewriting of the same information from one form to another.
At that time, I was working simultaneously as part-time IT-specialist at the All-Russian Centre for Eye and Plastic Surgery. My programming experience allowed me to start automating my work. I created a database and client interface in VBA in MS Access. The creation of this software enabled me to process medical records automatically, minimising the reentry of already entered data, significantly speeding up my work, and reducing errors.
2004. Automation of the work of a department manager.
That year, Facebook just appeared. At that time I went to work in the interventional cardiology department of the Republican Cardio Centre, where my father was the Chief of Department. At the time, healthcare data was a mess. Almost every week, top managers demanded more and more data in the form of reports from the head of the department. Data collection for these reports was practically impossible to plan. The need for such reports appeared spontaneously and often testified to the rich imagination of the requester. The heads of departments sadly walked through the offices and, drawing sticks on a piece of paper, manually counted patients for these reports.
Gradually, I wrote the software and launched it at the department level. I created a relational database that stores all important formalised data to obtain even a hypothetical report. Using SQL database queries, I received reports quite quickly.
The department’s doctors were happy to use the software to automate their work. My software assembled dozens of different forms for doctors from data entered once. Doctors printed forms, checked them, and included them in the medical record.
2006. MS SQL Server and involvement of other departments.
It was a time before the first iPhone appeared. The number of records in my MS Access database has become impressive, and the file database has become very vulnerable in terms of information security and distribution of rights. I just graduated from law school, where I understand the necessity of information security. Therefore, I transferred the project to MS SQL Server. The hospital’s senior management showed no interest in this product, but department heads adopted my program into their departments. The program almost began to take over hospital computers because it helped both department heads with their needs for reports and planning and ordinary doctors with the pain of paperwork.
At the same time, the leader of the IT department, Ilgizar Khamidovich Sultanaev, moved forward with the development of a system for automating the reception room, laboratory and submitting registers for insurance reimbursement. We managed to connect our two systems, receiving the benefits of automating the processing of data involved in payment on the part of Ilgizar and the clinical part on my part.
2007. Treatment Quality Feedback
At that time, I began working on my doctoral dissertation. I modified the client part so doctors could enter data into my research while maintaining an electronic medical record.
I understood well that not only the number of patients operated on was significant, but also the long-term results of treatment. Together with a colleague from the cardiocenter clinic, Alfiya Akhmetshina, we created an additional client program for analyzing and improving long-term results. The doctor took upon herself the targeted management of all patients operated on in my department. She actively called them, adjusted treatment, and at the same time entered data on results and endpoints.
2013. SaaS system. Web interface for follow-up.
In 2012, my father died, and the Ministry of Health appointed me to take his place as the head specialist in interventional cardiology and radiology of the Republic of Bashkortostan. I realized that a program that works locally only in the cardio center would not solve the problem of the entire republic. I asked two friends, Yuri Lotnik and Anton Kuznetsov, to write a Software as Service online follow-up monitoring system on a web interface.
We developed this system with maximum flexibility. I did not want to contact programmers every time to change forms, so we made a designer where the administrator could create new endpoints for monitoring and fields for data entry. Further in the work process, we refined the necessary information for each group of patients.
2016. Matrix has you.
The Ministry of Health of the Republic of Bashkortostan worked on a unified republican medical information and analytical system, RMIAS. The management of RMIAS contacted me, and I transferred all my listings and developments for inclusion in the system, which united all medical institutions in the region. However, some of my program functions were still unreplaceable even after I left the cardiac center at the end of 2021.
2016. Neural-network clinical decision support model trained on the treatment quality feedback
At the same time, I became interested in neural network models. I realized that we could use feedback data to train a neural network model and later be used to support clinical decision-making.
In the scientific paper I showed a three-stage decision-making model:
1. The first stage uses mandatory algorithms for clinical recommendations.
2. The second stage is essential in ambiguous situations, where a neural network trained with a dataset of factors, decisions, and feedback helps.
3. The third is related to the execution of the decision, tracking the results of this decision and correcting the dataset, followed by training the clinical decision support system.
And place of a doctor here to mix artificial intelligence and compassion.
