Columbia University professor and robotics engineer Hod Lipson knows the importance of artificial intelligence (AI) on a global level.
It permeates everything we do, from the stock market, from predicting the weather to what product youre going to buy, he said Wednesday during the second day of the virtual Ai4 2020 conference. Its even grading essays. You name it.
AI falls into the category of an exponential technology, meaning it accelerates with time. It is not just getting faster, but it doubles at a regular frequency and pace, and there are four major drivers:
Both biopharma and med-tech companies are increasingly pulling the technology into their business operations, working on programs that can assist in everything from drug discovery and clinical trial recruitment to precision diagnostics and patient compliance efforts.
Computing power has doubled every 20 months or so for the past 120 years, Lipson said, moving from mechanical instruments to graphics processing units (GPUs) today. Data-driven AI, however, is growing at a far faster pace than computing power, doubling every six months.
Were not talking about emails and web clicks and transactions. Were talking about things that are hard to capture, Lipson said. Its what you radiate when you walk down the street. Its the DNA when you touch something.
Data-driven AI is different from rule-based AI in that the engineer shows the computer what to do instead of telling it what to do. Rule-based AI systems, for example, employ a rule for suspicious credit card transactions, flagging accounts when a person suddenly spends three times more than in previous months, to spot fraudulent activity. Most AI systems in use today are built on rules.
Data-driven AI is a little more opaque, but the value of this approach is you dont need to understand the rules. You just need to give it examples, Lipson said.
Experts can be slow, expensive and frequently wrong, but with data-driven AI, all you need to improve the system is to give it more data. Its amazing that this simple idea works for everything. Driving a car, it learns, and it learns and it knows what to do from all of these examples.
An even faster exponential technology than computing power and data-driven AI is the machines capacity to learn, which doubles every three months, Lipson said.
Up until a decade ago, despite a surplus of data and machine learning, there were certain things AI was not able to do, like tell the difference between a cat and a dog. The AI community formed a competition to see if anyone could write software to solve this problem, and a group in Toronto presented what is called deep learning, in which the old-fashioned AI neural networks are stacked many layers deep. By 2017, the technology advanced to the point of only having a 3% error in terms of telling the difference between a cat and a dog.
So for the first time in history, machines are better than humans in what they see, Lipson said.
The fourth driver of AI is the cloud, which basically leads to AI teaching other AI systems. A driverless car, for instance, can teach and share its experience with other driverless cars, so the knowledge builds on itself, unlike with humans who do not necessarily learn from other drivers experiences. Doctors, also, do not become better doctors because of the number of doctors in the world and their experiences.
The cloud allows AI to teach AI and that is, at the core, a self-amplifying technology, Lipson said.
Applications in diagnostics, patient compliance
In the field of health care, AI can assist in finding patients for clinical trials and in narrowing down drug candidates to those with the highest potential. It also may play a role in cutting spending and improving care by reducing false diagnosis and overtreatment. Most AI in health care applications focus mainly on data from medical imaging, but convergent AI brings together all relevant patient information.
At Houston Methodist, scientists created a breast cancer risk calculator, pulling in data from PACS (Picture Archiving and Communication System), other raw data, mammography and ultrasound, and breast image features analyses. They then provide free test reports. With 23 million mammograms done annually in the U.S., about 10% are false positives, and 55% to 85% of breast biopsies show benign lesions. The estimated annual cost for over-biopsy and over-diagnosis in the U.S. is about $3 billion, said Stephen Wong, a chair professor and chief research informatics officer for Houston Methodist.
The waste goes much deeper beyond breast cancer. Administrative complexity across the health care system is wasting $265.8 billion each year, he said. But beyond reducing the wasted funds, AI has the potential to catch diseases with more precision than humans.
Many strokes in the U.S. are silent and missed, Wong said. Although computerized tomography (CT) scans are widely available at clinical sites and in emergency rooms (ERs), it is difficult to identify early ischemic changes and non-contrast CT images are noisy. There also are few stroke specialists in community hospitals and ERs. In general, about 30% of diagnosed strokes in ERs are not strokes and about 20% of them are missed.
About half of them are wrong, Wong said, leading his group to ask, Can we leverage AI to look at stroke detection so we can get to the right patient at the right place and right time, without missing anything?
Using deep learning in CT scans to de-noise the images and by employing MRI image mapping, early ischemic changes are more easily viewed, he said.
At King of Prussia, Pa.-based CSL Behring, which is focused on therapies based on human plasma, scientists built a forecasting system over the span of four months using data going back five years. Once COVID-19 hit, however, their single-digit errors rose to double-digits, forcing them to adjust.
When the world changes, your models have to change, said John Thompson, the companys global head of advanced analytics and AI. So we went back and collected data in the new world in the way the world was changing. We waited for four weeks and were retraining constantly, at which point, the models were trending back into single digit error terms again.
The company, which provides treatments for rare diseases such as hemophilia, also began looking to see if there are people who were not being accurately diagnosed with primary immunodeficiency disease (PID).
It looks like an allergic reaction in some cases, but also presents itself in other ways, Thompson said, so patients end up going to a variety of health care provider specialists to find out what is happening to their body. It can take years before a patient is accurately diagnosed.
With doctors records and prescribing behavior, the CSL team came up with a model and ran it through general population data, finding 16,000 people that most likely had the condition but were not diagnosed. About 1,000 of those patients were 80% to 90% through their patient journey.
That leads to a much lower quality of life, Thompson said.
Another area in which CSL has employed AI is in helping patients comply with their plasma-based therapies for several rare diseases. An analysis of different programs that were providing patients with an understanding of how to administer their therapies found that there was one, above all, that when patients had this support program, they complied and persisted on their therapies across the entire age bracket, Thompson said. Some complied at 80% or higher rates than patients not on the program.
Read more:
Artificial intelligence applications in health care on the rise - BioWorld Online
