Machine learning and artificial intelligence (AI) systems have long been touted as the future of medicine. A patient can walk into a doctor’s office and, after a quick scan, find out their risk of various diseases and receive information on how to prevent them from occurring. Patients suffering from diseases like cancer can have treatment decisions made by AI that can optimize care and maximize chances of survival.
Covid has been a chance for AI systems to really shine in the medical field, with increased funding and a spotlight put on them to make the right decisions during surges in cases that threatened to overwhelm hospitals. Instead, the systems failed.
The constant changes of the pandemic, with a new dominant variant appearing every few months, have made it difficult to keep track of machine learning systems. The systems have always proven to be a step behind the virus.
Experts, however, hope to learn from the failures of Covid and be able to improve the systems. More stable conditions like cancer and diabetes also don’t evolve at the same rate as the virus, giving experts a bigger window of time to collect data and build systems.
Machine learning proved to be not up to snuff during the COVID-19 pandemic, as the virus mutated so rapidly that it was impossible for data collection to keep up with the ever-changing landscape.
Dr. Yuan Luo is an associate professor of preventive medicine at the McCormick School of Engineering at Northwestern University. He wrote a viewpoint in JAMA last month highlighting the failures of machine learning during the pandemic.
He noted in his writings that machine learning systems intended to predict which patients were most likely to suffer the most severe symptoms, or even die, from Covid were not as accurate as hoped.
Dr Yuan Luo (pictured), associate professor of preventive medicine at Northwestern University, said the pandemic has taught experts important lessons, but these systems will still be effective against more stable diseases like cancer and diabetes at the future.
“It’s actually kind of a bubble bursting moment, it was machine learning,” Luo told DailyMail.com.
“It prompted us to really look at this issue and then see what is really going on here. Before the pandemic, people [had] so many expectations for machine learning.
He attributes machine learning failures during the pandemic to the ever-changing nature of the virus.
Covid is mutating at a rapid rate and there have been five variants different enough from the original Wuhan strain to be named by the World Health Organization.
Even within the five named variants – Alpha, Beta, Gamma, Delta, and Omicron – there are many bloodlines that emerge with slightly modified traits, such as the “stealthy” Omicron BA.2 bloodline.
Luo explained that the ever-changing nature of Covid prevented machine learning systems from collecting enough data to keep pace.
“The patterns or ideas that we [create] for previous variants or previous populations does not apply to the next variant or the next population,” he explained.
Not all diseases are Covid, and a virus that breaks out like this across the world is a once-in-a-century event.
Despite the recent failures of these systems, Luo remains optimistic that they will work in the future to fight more stable conditions like cancer. More work is needed to get them to this point, however.
The key to building these models, and the reason they failed during Covid, is data.
For an AI system to make accurate predictions of a patient’s health risks and outcomes, it must have thousands of data points to refer to, which it did not have for every new mutation of Covid that has emerged over the past couple of years.
There is a growing effort to collect data from patients around the world to build these systems. Luo mentions the Cancer Genome Atlas and the UK Biobank as specific examples of projects that have increased the availability of data to build systems.
He also mentioned that the National Institutes of Health plans to record data from one million patients in the coming years in a bid to expand data systems.
One thing the pandemic has taught experts is that data collection never stops.
“I think we’re also going to have to catch up to realize that the system is that we’re in a dynamic system,” Luo said.
“So we can’t afford to just be with our role models… [we can’t] I hope it is [accuracy is] will last forever.
He also notes that those building the models need to ensure that people of all types of races and ethnicities are included in the datasets to account for certain genetic differences that could make someone more at risk than others under certain conditions.
This type of data collection could create endless opportunities for data analysts and engineers.
Luo said that within the next 30 years, it’s possible that during an annual checkup, a doctor could quickly scan a patient, take a blood sample, and be able to let them know of any hidden medical conditions that they may have. he might have, and list their risk factors for different diseases.
One of the most ambitious machine learning initiatives of all time has just been announced by the White House.
Last month, the Biden administration announced the “Cancer Moonshot” effort, which the president says could end “cancer as we know it.”
Using data-driven machine learning, the program’s goal is to halve the number of cancer deaths in the United States over the next 25 years and dramatically improve the experience for patients and families. who suffer from cancer.
“The launch of the Cancer Moonshot program is timely given that we are collecting much larger cancer datasets much faster and, more importantly, new types of data,” Luo said, saying that this effort could revive the exact type of machine learning systems. he thinks it is possible in the future.
“I think the Cancer Moonshot Initiative has the potential to jump-start expansion efforts [machine learning] use in health care.
He said these types of programs can optimize care and help doctors make better decisions on behalf of their patients.
There is also concrete evidence of this. A Canadian study published last summer found that cancer treatments generated by machine learning were preferable to those generated by humans and exposed the patient on average to 60% less dangerous and painful radiation.