The last time medicine treated people as individuals, our cures were leeches, mercury, and herbal mixtures, while the most fortunate found pain relief from opium. As doctors engaged in trial and error, patients expired.
Medical treatments have been founded on scientific evidence for only a couple of hundred years. Over this time, both technology and means of population control have developed so much that effective treatments can be mass produced. But while the majority of people receive adequate care, this shift has meant giving up on treating people as unique individuals.
The information needed to treat each and every one of us individually exists, however. Data on a person’s genetic heritage, health status and treatment history, lifestyle and activity are available in abundance.
It’s just too much for humans to process.
Algorithms seek out precision vaccines
One person out of a hundred will get a rash from a vaccine. A common pain killer will seriously poison one in ten thousand. Yet both are considered decent medicines, as functionality and harm are always assessed according to average effects.
How a new vaccine, antidepressant or chemotherapy drug is absorbed is unknown. Likewise, how it affects and lives in particular person’s body remains open. To get through the data chaos and find answers, we would need to process superhuman volumes of information with even more superhuman speed.
We need help. We need more intelligence.
For a long time, the University of Helsinki and University of Oslo Professor Jukka Corander has been developing methods and algorithms for statistical inference, which help create vaccines for diseases like diarrhoea, pneumonia and meningitis that afflict the world’s poorest regions.
Corander’s methods can screen and test different candidates for the next vaccines and pharmaceuticals, and examine their combined effects in certain genetic populations. At the same time, his creations learn constantly and begin to invent new combinations – the kinds of combinations that humans would struggle to come up with and that would be impossible to find through laboratory experiments.
“Our AI models are 10 000 times faster than the testing methods currently in use. We can employ them to make reliable forecasts of how bacteria and viruses will react to different vaccines and how long a vaccine will be able to prevent disease. AI essentially functions like a digital assistant that helps drug developers discover entirely new combinations of pharmaceuticals and pathogens.”
There’s an enormous amount of variables in bacterial populations, antibodies and the human body for humans to process. That’s why existing vaccines are not optimal: they are not fully effective for everybody nor are they free of the risk of adverse effects. The effectiveness of a vaccine can also wane; a bacteria population evolves because its genes try to retain their ability to multiply. Corander tries to predict these changes with the help of AI, so that vaccines would retain their bite for as long as possible.
“We can already simulate enormous amounts of information from different sources: laboratory test results, bacteria and virus genes as well as their transformations during infections, and variations in the human immune system.”
Sensors on your skin and in your pocket
Would you worry less about a grandparent recovering from an operation at home if they had a bracelet, or even a carpet, capable of alerting an ambulance? Are the drugs prescribed for after-operation care sure to work? Does your family carry a gene that exposes you to cancer? In a couple of decades, these concerns may feel as antiquated as an old-school, mercury thermometer does now.
Everyday devices, like smart phones and watches, can already be used to collect data on your state of health. The manner in which this information is actually taken into use and made compatible with the data that flows in our body is a task for AI—and the key challenge in Aalto University Professor Simo Särkkä’s research.
“We’re developing models for combining sensor-gathered data with the information available on the human genome, bodily functions or disease properties. There’s a lot of noise and surplus information in all data, so the models need to be taught to extract only the relevant information in
a reasonable amount of time, and to combine information coming from different sources,” Särkkä explains.
A smart phone can provide data on a person’s movements and activity in many ways. Its magnetometer detects the Earth’s magnetic field, from which your position and movements can be deduced. Information on speed, acceleration and the position of the phone is provided by the accelerometer, while a gyroscope detects spin. The location of the device can be determined with the aid of a wireless network, the 4G network, Bluetooth connections and GPS signals.
“For example, pulse can be measured with the motion sensor and gyroscope by placing the phone in your breast pocket. The sensors then detect the movements caused by heartbeat.”
Särkkä and his colleagues are studying how to translate a person into parameters that can be used to construct a comprehensive yet simple model of that individual.
“A lot of different kinds of measurable data can be obtained on human physiology. ECG graphs, which are measured with several electrodes, provide a good picture of the electrical activity of the heart, for example. Cardiac mechanics can also be modelled at the same time – as can multiple bodily functions. In principle, it would be possible to use machine learning methods to automatically create a complete model of a person, but we’re trying to assemble a much lighter model.”
Smart phone and watch apps downloaded from an app store are for self-treatment and, at best, might serve as preventative tools. In order to be accepted for clinical use, a monitoring device must undergo years of studies and testing.
Särkkä is cooperating with the Hospital District of Helsinki and Uusimaa to develop a wearable measuring device for diagnosing heart disease.
“Days spent in hospital and, consequently, treatment costs would reduce, and patients could return home faster to recuperate. ECG analysis takes a lot of time for the physician, and this can be cut through automation. AI could combine different measurement results and make recommendations for the physician to focus on certain parts of the data accumulated on the patient.”
Screening blood on the spot
Data gathering becomes especially difficult and sensitive when sensors need to reach inside the body.
Associate Professor Tomi Laurila at Aalto University employs new carbon materials in the development of extremely sensitive sensors intended for measuring, for example, the concentrations of different neurotransmitters in the brain or the spread and effects of pharmaceuticals in the body.
“In order for the sensor surfaces to react to just the right substances and ignore the wrong ones, we need to know the surface atom by atom and how each will react to various substances and their combinations in the body. Machine learning enables us to model carbon structures up to 1 000 times larger than with current quantum mechanical methods – and not lose precision,” Laurila says.
“Entirely new physical phenomena have been revealed in our sensor materials thanks to AI.”
The effects of pain killers, for example, are highly individual. Even everyday ibuprofen can cause stomach aches and internal bleeding for many people, never mind the harmful effects of powerful opioids. Laurila believes that, in future, pain killer concentrations – and their likely individual adverse effects – could be determined from the blood count without leaving the doctor’s office. AI-utilising analysis tools would screen blood for desired substances and immediately predict the bodily reactions they cause, eliminating the need to wait days or weeks for test results to come back from a lab.
“We have trained machine learning methods to identify carbon in particular, but we’ve also succeeded in teaching AI to detect other substances, like oxygen or hydrogen, as well. The biggest challenge for biological and medical measuring is selectivity: how to extract the right signals from all the noise.”
Laurila notes that the experimental development of sensors is, a little like old-fashioned medicine, still largely based on trial and error. Combining machine learning and computational methods speeds up and rationalise experimental research considerably—while also tailoring the sensors to directly detect the desired substances, pharmaceuticals, proteins and mediators.
A digital twin for everyone
Even though the biggest upheavals AI will introduce to health care are yet to come, the most ambitious vision is clear.
“The personalisation of medicine will peak when we can create a digital twin for everyone,” says Academy Professor Samuel Kaski of Aalto University.
A model would be created for people individually, combining biological data, clinical examination results, sensor data and information on the patient’s lifestyle, environment and bodily system.
“We can now simulate the effects of drugs and treatments on the general level. Digital twins would mean we could do so for each individual. You could in practice query the model tailored for you about anything health-related – just ask, what if?”
AI alone can’t and won’t treat anyone. Determining what to examine and what kind of treatments patients need will remain the job of physicians. A digital twin could provide logical recommendations, evaluate the precision and reliability of its results as well as make an understandable case for its findings and rationale.
“Twins could also help in disease prevention. The model could, for example, give an assessment based on diet, exercise, various risk factors, and genetic disposition. Paired with a smart device, the twin could encourage people to do things that are beneficial to their health – in way that they would actually like to observe.”
Text: Tapio Reinekoski. Illustration: Jaakko Kahilaniemi.
This article is published in the Aalto University Magazine issue 23 (issuu.com) October 2018.