Why AI won’t replace doctors yet

by Takamitsu Sawa

A medical doctor diagnoses the patient and writes prescriptions based on interview with the patient as well as blood tests, analysis of image data obtained from magnetic resonance imaging (MRI) and computed tomography (CT), information related to the patient’s genes and so on. In giving the diagnosis, the doctor combines the information obtained through such processes with his or her own knowledge and experience. No matter how reputed a physician may be, the chances of them making a wrong diagnosis can never be zero.

With the recent progress in artificial intelligence, there has been much speculation that artificial intelligence could very well surpass a human doctor’s ability to make diagnoses and write prescriptions.

In August 2016, the Institute of Medical Science at the University of Tokyo released the outcome of a case study to show how powerful AI can be. IBM’s Watson AI program was fed with information contained in nearly 20 million medical articles related to cancer research and more than 17 million pieces of information related to pharmaceuticals.

The program, when it was then provided with data on the examination of a leukemia patient and the gene information of the patient’s cancer cells, offered within 10 minutes the diagnosis of a special type of leukemia that the physician in charge of the patient had never even dreamed of and, moreover, prescribed a combination of anticancer drugs best suited for that disease. Soon after being treated in accordance with the prescription, the patient recovered completely and was released from the hospital.

It is an easy task for any AI program to read and memorize information featured in nearly 20 million cancer-related papers and over 17 million pieces of pharmaceutical information, but it would be an impossible task for a human. Even if a person could read and analyze 10 medical research papers every day, it would still take nearly 5,500 years to complete the task. Moreover, there are limits to what a human brain can remember. It is impossible for anyone to accurately remember all the figures in these papers, and information is bound to slip out of memory as time passes by.

If medical doctors are no match for AI in making diagnoses and writing prescriptions, as this study shows, their jobs could be reduced to asking the patient some questions, informing the patient of the diagnosis given by the AI program and providing the patient with a prescription. The doctor’s skills may be tested only in asking the right questions to the patient. Does that mean that “excellent doctors” have become a things of the past? I will try to prove that is definitely not the case.

The 20 million cancer-related papers include information both relevant and irrelevant to cancer treatment. What determines the level of a doctor’s competence is the ability to determine whether a particular paper is relevant or irrelevant simply by skimming through it. It would be a total waste of time to read through a paper that is irrelevant.

It is said that in the fields of medicine and life science in particular, there are more than a few papers that contain fabricated or falsified images and data. The ability to detect fabrications and falsifications just by glancing at a paper will come from a doctor’s professional intuition — which an AI program cannot emulate. Moreover, the doctor’s ability to make diagnoses and write prescriptions may not necessarily be surpassed just by reading and memorizing all the professional papers and information in them. This is because AI does not possess the type of empirical knowledge that a clinical doctor has accumulated through treating large numbers of patients.

Yet it may be safe to say that AI is superior to radiotherapy physicians and technicians when it comes to analysis of MRI and CT images, which is said to hold the key to cancer diagnosis. AI can concentrate on such analysis over an extended period without fatigue or distracting thoughts.

The type of knowledge a medical doctor gains by reading and comprehending professional books and papers is called “explicit knowledge,” whereas the type of knowledge gained in the form of intuition or senses by conversing with and treating a large number of patients is called “tacit knowledge.”

A doctor hands down a judgment on the condition of a patient by combining these two types of knowledge. A doctor reputed to be excellent possesses a vast amount of tacit knowledge that has been accumulated through numerous cases of clinical experience, in addition to his explicit knowledge.

It is impossible to express or transmit tacit knowledge in writing, drawings, numbers or mathematical formulas. That is to say, tacit knowledge cannot be expressed in a scientific paper. What AI can learn from papers and other forms of information is limited to explicit knowledge that can be expressed in words and numbers. In other words, opportunities to accumulate clinical experiences are closed to AI.

Being a novice in the field, I can’t possibly assess how much weight a doctor’s tacit knowledge accumulated through numerous clinical experiences has in making diagnoses and writing prescriptions. But one thing is certain. AI does not possess an iota of tacit knowledge. That is to say, a diagnosis made or a prescription written by AI is solely based on huge amounts of explicit knowledge and the power of logical thinking.

The AlphaGo software developed by Google DeepMind for the board game of go has studied not only jōseki (standard moves considered to be optimum in the game) but also the records of tens of thousands of actual games played in the past. In addition, it conducts deep learning not only by playing games against professional go players but also playing tens of millions of games against itself.

In May 2017, AlphaGo soundly defeated the most skilled South Korean go player by four games to one. AlphaGo is different from Watson in that the former has carried out deep learning of tacit knowledge based on the experience of numerous games — whose volume is tens of thousands times more than that accumulated by any professional player. In short, AlphaGo possesses tacit knowledge gained through an immense number of games that no human go players could possibly play.

At least in the world of go, there may be no human players who can beat AI. But I am convinced that when it comes to the diagnosis and treatment of cancer, there are outstanding doctors whose ability well surpasses that of Watson, which lacks tacit knowledge.

Takamitsu Sawa is a distinguished professor at Shiga University.