This story was published in partnership with Ulyces.co.
Machines save lives
At the West of Beijing, next to a busy eight lane road, a cluster of white buildings adorned with gold lettering rise above exhaust fumes. This antenna of the sprawling People’s General Army Hospital is surrounded by pine and guards as solid as trunks. Nicknamed 301, the number it carried when it was still a medical school, the establishment located at 28 Fuxing Road is known for having welcomed many members of the Chinese Communist Party central committee, to which it is related. The old president Yang Shangkun breathed his last breath there in 1998.
Twenty years later, the Chinese military hospital is better armed against death. Thanks to an AI, the hospital’s doctors are now more capable of precisely determining if coma patients will wake up. Against their advice, the machine has already advised doctors to keep seven patients alive who were only deemed fit for death. These seven people did indeed wake up, says a team from the Chinese Science Academy in an article published by eLifeen in August 2018. “We have successfully predicted that a number of patients would regain consciousness even though it had initially been thought that they had no hope.”, the researchers boast.
We contacted professor Tianzi Jiang by e-mail, and he cited the example of a woman whose cortex was examined between January 2014 and May 2016. Admitted at the Chinese hospital 12 months after having received cranial trauma during an accident, she was pronounced to be in a “vegetative state”. This “bad” diagnostic put her between 7 and 23 on the traditional “Glasgow coma scale”. The digital programme gave her a score of 20. After having received care, the patient numbered 003 was allowed back home, and within 6 months doctors noticed signs of recovery.
Two years after being seen to, she could speak with her friends and relatives. “Last June, she came back to see us” confides Tianzi Jiang. “Our current model is optimistic about her recovery.” This is not the only case. A 41 year old woman who was also diagnosed by the AI only stayed in coma for a year, even though doctors had thought her to be a lost cause. After less than a decade in its setup, the program has been 88% correct in its predictions.
To begin with, the Chinese team collected data from 160 patients who were unconscious for at least a month. After using an MRI machine to scan the patients’ brains, they noticed that 11 of them had injuries that were so serious they disqualified them from the experiment. Those whose diagnosis was uncertain or whose conditions varied too much were also excluded. Based on previous studies that detailed the origin of thought, 22 zones of interests were chosen on the cortex of the selected patients. These zones helped evaluate the six states of the decisive neural networks: sensorimotor, auditory, visual, of “salience”, the “default mode”, as well as the “executive control”.
They obtained a map of the brain’s activity, which detailed the connection between different regions, and turned it into data. These numbers were then paired with clinical observations about the origins of coma or the body’s reaction to certain stimulus. On one hand, this was applied to the analysis of MRI scans, and on other hand, to health assessments. The latter already helps researchers make projections by using two methods: the Glasgow scale and the coma recovery scale. The model which was conceived by professor Song Ming draws a better picture because it adds cognitive activity elements, which produce more variables. It is then up to a computer to calculate the probability that the patient awakes.
In summary, the introduction of the article explains that a “computer model was created which predicts recovery based on MRI results, cause of injury, age, and the length of unconsciousness.” It is working rather well. The “multidimensional” approach that “combines clinical characteristics and MRI data improves predictions at an individual level and could conduct an expected identification of patients that are likely to recover consciousness.” It is not an exaggeration to say that the machine saves lives.
Nature imitating science
At the military hospital in the Chinese capital, the hospital beds are surrounded by rich families who came from all over the world to try and save a family member. These relatives “we desperate” notes Yang Yi, who works at the neurology department. “When our AI came back with a high probability in favor of recovery, they were very glad.” As unexpected as it is, the machine’s help is simply caused by an analysis of each of its operations.
Considered a forefather of this discipline, the American Psychologist Walter Pitts was born a century ago and 10,000km away. Son of a Detroit boiler-maker, the young boy grew up under pressure from the other boys in his district, who bullied him whenever they could. In 1995, the then 12 year old found refuge in the local library. Far from external stress and his father speeches about quitting school to join the business, Walter consulted the Principa Mathematica by Bertrand Russell and Alfred Whitehead. The knowledge he had earned in Greek, Latin and Maths during previous hideaways allowed him to spot mistakes in the work.
Sent off in a letter, his remarks caught the attention of Bertrand Russell. The British scholar is so impressed that he invites the young genius to Cambridge. Helas, too young, the latter has to decline. But three years later, having learned that his mentor his visiting the University of Chicago, Walter Pitts runs away to meet him. With Russel’s support, he begins to haunt the halls of the university. You can spot the shy student drinking whiskey and eating ice cream bought with his part time salaries. Before he goes to sleep around 4 am, he still leaves himself plenty of time to study.
Through his classmate Jerome Lettvin, the teenager meets the philosopher William McCulloch. Together, McCulloch and Pitts will in time lay the foundations for AI. These two scholars came together to look for a way to “imitate the human brain” using mathematical formulas. “A Logical Calculus of the Ideas Immanent in Nervous Activity” puts forward the idea of creating artificial neural networks by copying those found naturally. Even better, the machine would be better than man in that it can quickly “calculate, conclude, and operate based on choices; it can also calculate with information” notes the American programmer Edmund Berkeley in “Giant Brains: Or Machines that Think” released in 1949.
Pitts and McChulloch’s arithmetics’ has its limits. The Hungarian physicist John Neumann wrote, “these authors have shown that absolutely everything can be accomplished by the appropriate mechanism, more specifically, by a neural mechanism.” However, “nothing that we can know or learn on the workings of the body gives us any clue on the workings of neural mechanisms, apart from studying brain cells through a microscope.” It will take many decades for us to better understand the brain’s functions. By that time, Pitts and McColloch’s neurons will have vanished.
Alcohol abuse landed Pitts in hospital, and on the 21st of April 1969 he writes a letter from his hospital bed to McColloch, who is also in hospital. “I understand you had a light coronary; … that you are attached to many sensors connected to panels and alarms continuously monitored by a nurse, and cannot in consequence turn over in bed. No doubt this is cybernetical. But it all makes me most abominably sad.” Walter Pitts died on the 14th of May 1969, followed by McCulloch four months later.
The adult age
The sensors that accompanied McCulloch during his last days served a future purpose. By placing an electrode battery on a patient, we can not only observe a coronary insufficiency but also gather indications on the “different states of consciousness, purposeful responsiveness, including comatose and vegetative states.” Doctor Sharbrough presented this theory at the American EEG Society June 9, 1981. That technique is even “particularly sensitive to the gravity of the cerebral malfunction”, Chiappa and Roper added in 1984.
Even as the EEG curve studies evolve and multiply, a new brain activity research field emerges: “Despite its recent introduction, MRI has already become one of the best neuroradiology methods out there” Nadine Martine, clinic leader at the Beaujon hospital, notes in 1986. Soon, MRI and EEG diagnostics will be enriched by algorithms, whose sudden incursion into the hospital was the subject of a 1985 conference in Padova, Italy. Massachusetts Institute of Technology’s professor Peter Szolovits also dedicated an entire article to this change in 1982. But Canadian Physician Edward Shortliffe judges that in 1993, almost a century after its conception, the use of AI in medicine is still in its teenage years. In other words, it is far from being a mature science.
To reach adulthood, the medical AI systems will need “the development of integrated environments for communication and computing that allow merging of knowledge-based tools with other patient data-management and information-retrieval applications” Shortliffe guesses. If it is “difficult to see the transition of a scientific discipline towards adulthood”, data based studies are on the hand “becoming more and more frequent.”
Just two years later, deep learning researcher Piotr W. Mirowski programs a convulsive neural network to predict the chances of epileptic attacks from 2009 onward. The same year, three Chinese researchers bet on MRI in order to evaluate the probabilities of cranial trauma. Thus, there is no longer any doubt that clinical AI is a major help in hospitals. In 2010, Microsoft launched InnerEye, a program able to detect tumours on x-ray images. IBM and Google follow closely with Deepmind and Watson.
AI is articulated in two technologies at coma patient’s bedside, MRI and EEG, to which we can add position relay tomography, Tianzi Jang notes. Each of these is useful. To predict a heart attack, the Dutch Neurophysicist Michel van Putten observes that “Objective Continuous EEG monitoring seems a promising tool”. Also promising is Chinese research on coma caused by cerebral trauma. That said, the first method better traces blood circulation routes. “EEG lets us listen to neurons, but MRI does not measure their expression.” he says.
And so with colleagues he created many networks using specialised literature, so that they could train themselves in analysing encephalographs that belong to coma patients. In 2017, their conclusions appear in an article named “Deep Learning for outcome prediction of postanoxic coma“. They add, “we are trying to make a prediction 24 to 48 hours after heart failure. We are able to have 50% to 60% precision after 24 hours. Which is, to our knowledge, the best result in the world.” The Chinese model itself needs a conscious patient to then determine the chances of waking up.
“Given that some of the brain’s functional activity or connection is not detectable to the naked eye, the link established by our model suggests that AI is pertinent in evaluating the level of consciousness lost and that it is very reliable” Tianzi Jiang brags.
“When we tell families about the AI results, we always tell them that it should account for between 20 and 50% of their decision-making process” advises Yang Yi, who participated in research. “They must also take into account the care that will be given, medical cost, their insurance and the chances of sudden death.” Even if this speech has been generalised for all patients, this way of rationalising an important decision could, according to the researchers, avoid major medical spending in what are deemed desperate cases. But Ming Song assures us that “it will be a tool that will never replace human doctors”.