Algorithm nation

Since the idea of thinking machines first appeared in literature, artificial intelligence has often been portrayed as a future villain or hero. But the fact is, AI is here now and already suffuses our day-to-day lives. AI keeps people safer while driving. It helps around homes through myriad smart technologies. It answers questions from trivial to profound through search engines.

At the Keck School of Medicine of USC, AI has potential for broader impact still — as an enabling tool for investigations to improve human health. Algorithmic brawn fortifies longtime areas of strength at USC, such as research into cancer, neurodegenerative disease and population health, and aids in snooping out the biomedical needle in the hay.

“Our innovative doctors and scientists are using AI to merge clinical medicine and research,” said Steve Shapiro, MD, USC’s senior vice president for health affairs. “By making sense of real-world data, we may learn things no one else knows. The results could help to individualize health care, to make care more affordable, accessible and equitable, and to create new cures.”

“Our academic enterprise focuses on addressing the most complex challenges to health, and to make progress, we embrace the opportunities presented by advances in computation,”
–Dean Carolyn Meltzer, MD

Trojan scientists are penning early chapters of a barrier-breaking tale fusing human ingenuity and computational muscle. The lifesaving and life-alerting implications range from new preventive strategies to more-effective drugs, from earlier diagnosis to enhanced critical care.

“Our academic enterprise focuses on addressing the most complex challenges to health, and to make progress, we embrace the opportunities presented by advances in computation,” said Dean Carolyn Meltzer, MD, holder of the May S. and John H. Hooval Dean’s Chair in Medicine and professor of radiology. “More and more, our researchers are finding creative ways to channel artificial intelligence to uncover new knowledge and cultivate solutions to the problems that threaten life and erode quality of life. I’m excited to see where these avenues of innovation will lead.”

AI’s unique advantages for biomedical research

This story is not about devices or algorithms replacing humans. Rather, it’s about experts deploying powerful tools to do things that they couldn’t otherwise. This “intelligence” doesn’t rival humans, but instead complements them.

Machine learning is a primary enabling AI technique at the Keck School of Medicine. This method’s potency as a force multiplier for research resides in today’s abundance of data. On one hand, electronic health records and devices such as sensors provide mountains of metrics. On the other, insight into the umpteen elements of the invisible world controlling biology only grows.

“Twenty years ago, we were able to measure maybe a handful of genetic variants along the genome, so we had to use prior knowledge to pick which ones to study,” said David Conti, PhD, holder of the Kenneth T. Norris, Jr. Chair in Cancer Prevention, professor of population and public health sciences, and associate director for data science integration at the Keck School. “As technology ramps up, we can measure all across the genome and get to these random variants while still looking at the ones that we think are biologically interesting. We want to build a model that aggregates all of that information, and that’s where machine learning can really come into play.”

Data is fuel for machine learning. The more thoughtfully selected data that is fed to a machine-learning algorithm as it “trains,” the more impact that the results will have. The fundamental application, across contexts, is separating signal from noise, identifying patterns within seeming chaos that confounds both human intellect and most previous technological tools.

“Machines are specifically designed to handle lots of decisions simultaneously,” said Arthur Toga, PhD, holder of the Ghada Irani Chair in Neuroscience and Provost Professor of Ophthalmology, Neurology, Psychiatry and the Behavioral Sciences, Radiology and Engineering at USC. “We can combine an incredibly diverse array of observations in individuals and groups of individuals, feed them into these algorithms, and allow them to derive patterns that may be difficult for us to observe as humans. That’s where the marriage between big data and sophisticated artificial intelligence must occur, because one is dependent on the other.”

Ultimately, machine learning is enabling Keck School of Medicine researchers to approximate reality. Each person’s health and well-being is influenced by layered, interreacting factors: genes and environmental exposures, immunity and metabolism, lifestyle and socioeconomic factors. AI offers researchers the capacity to explore these numerous dimensions and sort what’s important from what isn’t.

“We can combine an incredibly diverse array of observations in individuals and groups of individuals, feed them into these algorithms, and allow them to derive patterns that may be difficult for us to observe as humans. That’s where the marriage between big data and sophisticated artificial intelligence must occur, because one is dependent on the other.”
– Arthur Toga, PhD

“There are just too many variables to deal with,” said Neil Bahroos, chief research informatics officer and associate professor of research in population and public health sciences. “With AI, we can plug in all these factors and see which trends we’re looking at. It’s a total gamechanger.”

Leveraging AI across the spectrum of health care

AI brings the big picture into focus for prevention and community health

For Conti, who also serves as associate director for data science at the USC Norris Comprehensive Cancer Center, machine learning is one important tool in a more comprehensive kit of statistical techniques he uses to tease out causes for cancer and other diseases.

AI helps him mine insights from so-called “omics” — diverse fields of analysis seeking to comprehensively profile details coded in molecules such as DNA, RNA, proteins and metabolic products — and the “exposome,” environmental inputs such as chemicals and air pollution. His discoveries may, in turn, inform policy and empower people to make protective choices.

“Epidemiology is all about identifying a risk factor,” Conti said. “As more dimensions come into focus, it’s less likely we can leverage prior knowledge to pick out that specific risk factor. We need to leverage machine learning to say, How do we select from a massive number of features?”

“If we develop models with data that is limited to individuals of a certain ancestry or ethnic group, they won’t be applicable to others and it could actually increase health disparities”
– David Conti, PhD

Just as his work requires careful choosing of statistical tools, it also calls for conscious curation of data. That means the inclusion of datasets about diverse populations that have traditionally lacked adequate representation in research, such as African Americans. The endeavor may push forward efforts in health justice.

“If we develop models with data that is limited to individuals of a certain ancestry or ethnic group, they won’t be applicable to others and it could actually increase health disparities,” Conti said. “We’ve always been really focused on looking at how risk factors differ, or are similar, across different groups, and then making conclusions that are relevant across them.”

Computation that advances early detection and diagnosis

As Conti demonstrates the predictive power of his work, the results may provide new ways to screen for cancers, one of numerous health threats the medical school’s researchers seek to ferret out before the damage is done.

Among the cruelties of neurodegenerative diseases such as Alzheimer’s is the fact that the brain is not built to recover from them. Early detection must accompany any interventions developed to ameliorate the conditions. That’s why Toga and his colleagues hope to pinpoint markers that precede symptoms.

“We aren’t seeing an index for disease progression until fairly late in the game,” said Toga, who directs the Mark and Mary Stevens Neuroimaging and Informatics Institute and leads the Laboratory of Neuro Imaging. “We need to know earlier, because what you lose, you don’t get back. Machine learning helps us make the literally millions and millions of comparisons that might indicate somebody is in the earliest stages of change to brain tissue.”

He heads up programs that present tremendous assets. The USC Stevens Neuroimaging and Informatics Institute has the world’s largest collection of training data for machine learning, comprising imaging, genetics, biosample, cognitive and electrophysiology data.

With this resource, his team ties together machine vision reading out medical images with potential biochemical or genetic indications for Alzheimer’s risk and signs of disease. Image analysis also drive progress in estimating “brain age,” which sometimes varies significantly from calendar age.

Importantly, USC shares the wealth and coordinates resources across institutions. Toga is a principal investigator of the Data Archive BRAIN Initiative (DABI), a shared repository for brain physiology data. Meanwhile, he also leads the Global Alzheimer’s Association Interactive Network (GAAIN), with data about 560,000 participants with Alzheimer’s disease. If a researcher identifies a pertinent dataset, they can then seek permission from those who gathered it.

“It’s a matchmaking thing that seems to have found an incredible sweet spot to facilitate global cooperation,” he said. “This is important when you have a difficult set of problems such as Alzheimer’s disease.”

Another major international effort with ample feedstock for AI is led by Toga’s colleague and associate director of the USC Stevens Neuroimaging and Informatics Institute, Paul Thompson, PhD, professor of ophthalmology, neurology, psychiatry and the behavioral sciences, radiology and engineering.  

The Enhancing Neuro Imaging Genetics through Meta-Analysis (ENIGMA) Consortium combines data and expertise from more than 2,000 scientists in over 40 countries to tackle a portfolio of 30 brain diseases. The consortium has mapped the effects in the brain of conditions including schizophrenia, bipolar disorder, depression, posttraumatic stress disorder and autism.

Some of Thompson’s own investigations employ machine learning to link images of the brain to genetic changes that could be cues for screening in Alzheimer’s. Other pieces aim to forecast health outcomes. The future may hold algorithms that derive important information from comparing an individual to millions of others.

“We envision that doctors will determine chances of recovery with alternative treatment choices, with help from machine learning,” Thompson said. “AI can deliver a better diagnosis, prognosis and treatment choice. Those are the big three we’d like to happen.”

Breaking a new path for improved critical care

Neurologist Sebina Bulic started with a clinical issue that concerned her deeply. She ended up pioneering the use of AI to extract meaning from the abundance of data in electronic health records at USC.

In the Neurocritical Care Program, fever, high counts of immune cells and accelerated heart rate are common symptoms of brain injury. Elsewhere, these are signs of infection, and automated alerts that would be crucial in any other intensive care unit are often false positives. As a result, it is difficult to differentiate between real infection and an injured brain.

“The stakes are so high in any ICU,” said Bulic, MD, assistant professor of clinical neurology and director of the stroke neurology service. “In infection, the earlier you start treatment, the better. You simply cannot miss the signs.”

She decided to do something about it. She assembled a team to develop an algorithm customized to neurocritical care for identifying actual signs of infections that happen in the hospital.

The pilot program was a classic example of USC collaboration. It pulled in the Keck enterprise data warehouse team, a research team from the Department of Surgery, the SC Clinical and Translational Science Institute and engineers from Amazon Web Services. Bulic is also grateful for the counsel of Daniel Pelletier, MD, holder of the Eric and Peggy Lieber Chair in Neurology, whose own investigations into multiple sclerosis use imaging, genetic data and machine learning.

After two years of hard work, the collaboration’s findings await peer-reviewed publication. Being the first to attempt a clinical application from USC patient data required persistence, especially amid the disruptions of the COVID-19 pandemic. Adjustments were also necessary along the way to gain value from essential data in results locked up in PDF form, using an AI subfield called natural language processing.

“The algorithm took into consideration which data points made the most impact,” Bulic said “I want to continue to cut out the noise and identify data points that are available in any electronic health record, so this can be scaled up to all neuro ICUs in the country.”

Bahroos, the Keck School of Medicine’s chief research informatics officer, identifies her project as a prototype for the future. Applying machine learning to health records for patient benefit is one main objective of the Health Data Innovation Program, his team’s nascent venture to advance USC’s biomedical research.

“We want to make this a part of production for Keck Medicine,” he said. “This can then serve as a template to tackle other priorities such as preventing pressure ulcers and urinary tract infections in the hospital. Predictive analytics could really increase quality and improve health outcomes.”

The vision of a digital guide to the most effective treatments

Studies led by Conti, of the Department of Population and Public Health Sciences and USC Norris cancer center, also set the table for precision medicine approaches to cancer. This is a vision for care custom tailored to each patient’s case, such as the use of drugs that target a specific mutation found in a tumor.

At the same time, neuroscientist Paul Thompson is turning AI toward drug discovery for Alzheimer’s. He is the leader of the NIH-funded project Ultrascale Machine Learning to Empower Discovery in Alzheimer’s Disease Biobanks, known as AI4AD, which involves 40 co-investigators at 11 research centers.

“We excel because we give independent minds all this autonomy to flourish,”
– Neil Bahroos

Alongside efforts to predict an Alzheimer’s diagnosis and advance algorithms themselves, AI4AD nurtures investigations that may identify new drug targets — an urgent need.

“There’s a new FDA-approved drug for Alzheimer’s, and it doesn’t work for everyone,” Thompson said. “It’s very important to find new ones. We’re developing AI methods to identify the hotspots in the genome that promote Alzheimer’s, so we can find a way to switch them off.”

Just the start: lowering the barrier of entry for AI-enabled health research  

AI-enabled research is an area of growth at the Keck School of Medicine thanks to the Health Data Innovation Program. Bahroos is staking out a big tent.

“My goal is to make sure that researchers in all domains can benefit,” he said. “It can be a high burden for a research group to figure out all of this technology. We plan to make it easy.”

One early focus has been building a cloud-based system where data can be consolidated across USC research groups. A catalog comprehensive in scope and searchable detail would be a substantial boon to Trojan investigators. Basing it remotely offers the chance — as with Toga and his GAAIN collaborators — to provide access while protecting researchers’ datasets.

Bahroos sees the approach as well-suited to a fundamental USC advantage in research.

“We excel because we give independent minds all this autonomy to flourish,” he said. “For discovery-driven studies, people want to explore. I ask them, ‘What is your question?’ They say, ‘Let me see what’s there, and I’ll come up with questions.’ Our program is going to help them do that.”