AI can accurately predict the time of death. Is it reliable?
"I don't want to be speechless when I finally leave. I hope I can say goodbye to him when I am still pretty." In the latest issue of "Exquisite Convention", Ding Rui, owner of the death experience hall, shared a touching feeling. The story of a woman in the late stage of cancer, in order to be able to say goodbye to her husband and to participate in the death experience with her husband, the husband uttered tears when she said good-bye to her “dead” wife.
For a cancer patient, although the topic of "death" is very heavy, it also needs space for exploration and adaptation. And if you can accurately predict the patient's "death", can you give the patient this space?
Stanford University in the United States has developed a "predicted life" AI system. The AI system collated the electronic health record data of nearly 2 million adults and children, as well as relevant medical diagnostic information, to get big data on the condition. Then through data collection and system autonomous learning mechanism, to predict the specific time of death of patients.
For the end of life care, but also to save lives
In China, about 7 million people come to the end of life each year, but hospice care provided by society can only meet about 15% of the demand. The emergence of this system indicates that doctors can more accurately arrange the hospice care of patients. In addition, using "predicting death" we can also explore a new path.
For late-onset patients, we can determine the probability of death through long-term data tracking. As for the sudden symptoms of some special diseases, we can also learn the changes in some vital signs of the patient through machine learning so as to send out dangerous warnings.
The FDA (United States Food and Drug Administration) recently approved the first death-predictable AI product, the Wave Clinical Platform, developed by medical technology company Excel Medical.
Wave is an online, always-on, remote monitoring platform that integrates real-time data from hospitals, including patient medication status, age, physical condition, past medical history, and family history.
This system can sense subtle changes in vital signs, giving an early warning six hours before a fatal event occurs. In other words, the AI system can predict sudden death time by 6 hours in advance by comparing sudden death cases in the database and earn time for medical personnel.
British scientists also published an article in the Journal of Diagnostic Imaging, saying that artificial intelligence can predict when heart disease people die. This technology allows healthcare workers to find patients who need more intervention to save more lives.
AI predicts death - unable to escape the cage
For the problems encountered by AI's prediction of death, Yan Xi, an analyst with the Intelligent Relativity Theory (ID: aixdlun), thinks that these three aspects can be considered.
- "Predicted death" means "death of death". Can the patient accept it?
It is undeniable that predicting death can indeed allow doctors to more rationally allocate medical resources. But "death" is not so easily accepted by everyone.
Siddhartha Mukherjee told a story he had witnessed in his article. He once treated a patient with esophageal cancer. The treatment of this patient is very smooth, but there is still the possibility of recurrence. So the doctor put forward the topic of hospice care. But this patient refused. This patient thinks that his physical condition is getting better and better, and his mental state is very refreshing. Why does he have to say these disappointments?
Unfortunately, the patient’s cancer still relapsed. Before his death, he was in a coma and couldn't respond to his family beside his bed.
From this story we can see that not every patient can accept the topic of "death" indifferently. When patients struggle with the disease and death, doctors use a set of scientific and sophisticated AI systems to predict the patient’s death. For patients, the cancer journey is difficult and it hangs overhead. The blade of death that will drop on time is also too cruel.
- There are individual differences in the condition, difficult to determine complex cases
AI predicts death depends mainly on medical big data and deep learning. The research team stated that the AI system collected patient data from the discovery of the disease and died within 12 months, and then calculated the weight and intensity of each information using deep data through deep neural network to generate a given patient in 3 to 12 The probability of death within months, using scores to predict whether the patient will die within 3-12 months.
The variety of medical data and the uneven quality are very personal information. The course of the disease has a certain law, but the specific symptoms of the disease must vary from person to person. Individual physical conditions, the surrounding environment and other factors will affect the outcome of the disease. In addition to individual differences, the disease itself is difficult to clearly recognize.
For example, the initial symptoms of almost any infectious disease are similar to colds. In other words, the disease itself is deceptive. In hospitals, doctors often need help tools. In the face of complex cases, doctors even need to hold a disease discussion meeting, and several parties can determine the treatment plan.
Moreover, AI predicts a profound and inexplicable part of the deep learning of death, the "black box" problem - it can calculate a patient's death probability score, but cannot express the logic behind it.
Therefore, there are still many problems in predicting the patient's death time through the probability score. The prediction of death for certain diseases alone may be effective, but the probability of predicting the death probability of major diseases is minimal.
- Medical big data sharing is difficult
AI+ Medical mostly starts with an algorithm, but eventually it will return to the data. Although data is a problem for all AI projects, data from the medical industry, especially those related to life and death, are more difficult to obtain.
Medical information is different from information in other fields. The types are very complicated and the standards are not uniform. In particular, many diseases involve patient privacy. Some patients are reluctant to use their medical data for AI research.
In terms of quality, medical data also have higher requirements. For example, all medical data require the doctor's manual identification.
In addition to patients, there are also resistances from obtaining data from hospitals. When it is uncertain whether a study will benefit hospitals, the hospital may not be willing to contribute all the work data. How the technicians work with doctors to obtain high-quality big data is a common problem faced by most artificial intelligence medical companies.
So is this technology really worthless?
"The accuracy of AI predicting deaths as high as 90%" is more like an over-hyped gimmick. It is predicted that human deaths are only more convenient for palliative care, but there are still some ethical issues. For example, do you want to inform the patient and his family about the date of death? Is a machine qualified to declare the death of humans?
And if you change another forecasting object? Imagine, as the owner of a pet dog, when the dog's bodily functions gradually become old, does the owner want to know when the dog will die? Because of language barriers, humans would like to use some aids to better understand pets and hope to have a more accurate medical assistance system to diagnose the condition of pets so as to make better arrangements for pets. In the face of pets, AI predicts that death seems to be more acceptable to humans.
AI predicts the development process of the death system should be a process of continuous improvement of value. On the one hand, a database of more objects should be established, relying on deep learning to make more choices of application scenarios. First, a class of objects (mostly pets) is selected as a teaching material for a training and learning model. Then, the accumulated “experience” is used to determine the probability of death of such objects during the onset of illness. Finally, the subject is treated with intervention.
On the other hand, predicting death becomes a predictive course. From the vertical domain to the horizontal domain, the prediction scenario is to build an intelligent prediction system that includes both the duration of the disease course and all the stages of the early stage of the disease, and finally, it can be used to model the user's personality.
At AI Medical, we have broken down more and more titles. The "prediction of death" seems to involve human life and death, but it only touches the surface of the matter. After the "death prediction" bubble is punctured, how to make AI medical prediction become a truly beneficial project that touches medical care. Pain points, I am afraid that is the most AI medical companies to think about.