Doctors and computers
A 49-year-old man notices a painless rash on his shoulder, but does not seek medical help. A few months later, during a routine examination, the therapist also notices this rash and diagnoses a benign neoplasm of the skin. After some time, during routine screening, the nurse points out this rash to another doctor, who strongly recommends that the patient consult a dermatologist. A dermatologist performs a biopsy. According to the results of histological examination, the patient has a benign neoplasm. The dermatologist requests a repeat examination of histological preparations. This time, a different verdict was issued: invasive melanoma. The patient immediately begins a course of chemotherapy. A few weeks later, a therapist friend asks why he doesn't get immunotherapy instead.
Various versions of this hypothetical scenario quite often unfold in the conditions of modern healthcare, and the reason for this is not negligence, but the human factor and system errors.
Experts from Harvard University Medical School and Google claim that with the right approach, artificial intelligence can significantly reduce the frequency of both system errors and erroneous conclusions of individual clinicians.
The article they published presents a project for integrating machine learning into clinical practice, as well as describes the prospects and pitfalls of a technological breakthrough that has captured the imagination of both bioinformatics specialists and clinicians, as well as people who are not related to science.
The huge computing power and analytical capacity of artificial intelligence can expand the unique abilities of the human brain to make decisions based on common sense and the ability to notice nuances. The authors believe that the combination of these two approaches will optimize clinical practice.
Machine learning is a form of artificial intelligence that is not based on predefined parameters and rules, but implies adaptive learning. Thus, new data progressively improves the algorithm each time, the ability of which to identify patterns improves over time. In other words, machine learning demonstrates a kind of neural plasticity similar to the cognitive plasticity of the human brain. However, while the human brain can identify complex associations based on small pieces of data, machine learning requires significantly more examples to acquire a similar skill. Computers learn much slower, but they have more productive power and make fewer interpretation errors.
According to Isaac Kohane, dean of the Faculty of Biomedical Informatics at the Blavatnik Institute at Harvard Medical School, a machine learning model can be trained on tens of millions of medical records containing hundreds of billions of observational data without episodes of attenuation. At the same time, a human clinician throughout his career is not able to examine more than several tens of thousands of patients.
Thus, the authors claim that the use of machine learning can provide clinicians with the collective wisdom of billions of medical decisions, billions of medical records and billions of disease outcomes, which will allow them to make informed decisions when making a diagnosis and select therapy according to the patient's characteristics.
In situations for which the accuracy of the forecast is critical, the ability of a machine learning system to identify indicative patterns among millions of samples can endow a doctor with "superpowers".
In 1999, the Institute of Medicine, now known as the National Academy of Medicine, published an article entitled "To Err is Human" ("To Err is Human"). It describes the imperfection of the inherent human decision-making mechanism and the limitations of the clinical knowledge of an individual doctor. The latter was predicted to play the role of an aggravating problem of ordinary doctors who have to synthesize, interpret and apply an ever-growing mass of biomedical data generated by an exponentially increasing number of new discoveries.
According to Kohan, we should recognize and accept the fact that a practicing physician is physically unable to keep up with the new biomedical data and discoveries that are emerging at a tremendous speed. Artificial intelligence and machine learning can help reduce the number of errors or even eliminate them, optimize productivity and provide support in clinical decision-making.
According to the publication of the Institute of Medicine, clinical errors are divided into four broad categories:
- Diagnosis: inability to prescribe adequate examinations or interpret their results; use of outdated examination methods; erroneous diagnosis or untimely correct diagnosis; inability to act in accordance with the results of the examination.
- Therapy: the choice of non-optimal, outdated or unsuitable methods of therapy; errors in the course of therapy; errors in the dosage of drugs; untimely start of treatment.
- Prevention: ineffectiveness of preventive control and application of preventive measures, such as vaccination.
- Other errors related, in particular, to communication or equipment failures.
The authors believe that machine learning has the potential to reduce the frequency of many of these errors and even completely eliminate some of them.
A properly designed system can notify clinicians about the choice of a suboptimal drug, exclude dosing errors, and also send medical records of patients with unclear and strange symptoms to experts in rare diseases for remote consultation.
Machine learning models have the most pronounced potential in the following areas:
- Prediction is the ability to identify patterns that predict the outcome of the course of the disease based on a huge number of previously recorded outcomes. For example, what is the most likely course of the disease? How soon will the patient be able to return to work? How fast will the patient's disease progress?
- Diagnosis is the ability to help identify the most likely diagnoses during clinical visits and indicate possible future diagnoses based on the patient's medical history and the whole set of results of earlier laboratory tests, diagnostic imaging and other available data. Machine learning models can be used as an additional source of information that encourages doctors to consider alternative conditions or ask leading questions. This may be most valuable in cases with high diagnostic uncertainty or when patients exhibit particularly contradictory symptoms.
- Therapy: Machine learning models can be "trained" to identify the optimal therapy for a specific patient with a specific disease based on extensive data sets on outcomes for patients with a similar diagnosis.
- Clinical workflow: Machine learning can improve and simplify the existing electronic medical records storage system, which significantly complicates the work of clinicians. Improving work efficiency and reducing the amount of time spent working with electronic medical records will allow doctors to devote more time to working directly with the patient.
- Expanding access to consulting services is the ability to improve access to medical care for patients living in remote geographical areas or regions with a shortage of specialists in the field of medicine. Such models can provide patients with information about the possibilities of getting help near home or warn about the appearance of symptoms that require immediate attention or treatment in the emergency department.
Artificial intelligence and machine learning are not perfect and will not allow you to cope with all the mistakes in clinical practice.
Machine learning models can be as good as the data entered into them is good. For example, a machine learning model for choosing therapy methods may be as good as the exact treatment methods are entered into the database on which this model was "trained".
The authors note that the most significant barrier to the development of optimal machine learning models is the lack of high-quality clinical data, including populations that differ in ethnic, racial and other characteristics. Other difficulties are more of a technical nature. For example, the modern separation of clinical data both between and within institutions is a significant, albeit surmountable obstacle to the creation of reliable machine learning models. One solution is to put data in the hands of patients to create patient-controlled databases.
Other difficulties include different legal requirements and legal policies, as well as a variety of technical platforms used in healthcare and technical providers, which may be poorly compatible with each other, which makes it difficult to access data.
The authors also warn that one of the undesirable consequences of using machine learning may be excessive trust in computer algorithms and a decrease in the vigilance of doctors, which will lead to an increase in the number of clinical errors.
One of the methods to minimize such risks is to introduce confidence intervals for all machine learning models, indicating to clinicians the estimated accuracy of the model.
More importantly, all models must undergo periodic inspections, just as practitioners periodically take exams to update certificates of compliance with their field of activity.
If the question is posed correctly, machine learning will act as a support that increases the effectiveness of interaction between a doctor and a patient, and not as a substitute for a human doctor.
The authors note that the human attitude and responsiveness of the doctor, as well as his attention to the small details and complexity of human life will never lose relevance. Therefore, in this case, we are not talking about competition between a computer and a person, but about optimizing medical care through the use of artificial intelligence capabilities.
Article by Alvin Rajkomar et al. Machine Learning in Medicine is published in The New England Journal of Medicine.
Evgenia Ryabtseva, portal "Eternal Youth" http://vechnayamolodost.ru based on materials from Harvard Medical School: The Doctor and the Machine.