How to harness AI across healthcare

In May 2020, as the COVID-19 pandemic was overwhelming medical systems around the world, researchers at Mount Sinai’s BioMedical Engineering and Imaging institute in New York City showed how artificial intelligence (AI) could lend doctors a hand. Using data from CT scans, an AI model they developed accurately diagnosed the virus in patients who did not yet show obvious lung abnormalities¹. The approach enabled them to isolate and treat patients earlier, saving lives and limiting the spread of the virus.

Now, researchers and physicians at Mount Sinai are applying AI and machine learning across medicine. Their work is transforming diagnosis and treatment in several specialties, empowering healthcare providers to mine patient data to uncover new insights. The results suggest that AI and machine learning could not only enable doctors to make faster interventions that save lives, but also to better allocate resources in hospitals and reduce healthcare costs.

Clinical assistance

David Reich, an anaesthesiologist and president of The Mount Sinai Hospital and Mount Sinai Queens, calls AI “augmented” intelligence. He sees its role in medicine as offering valuable assistance to doctors, nurses, and hospital administrators, not replacing them. “AI can help them do a better job,” he says. “While they remain responsible for patient care and outcomes, AI can help by bringing things to their attention they might not have otherwise considered, such as identifying patients who may unexpectedly require critical care².”

An example of how AI can better monitor patients came when Mount Sinai researchers developed a model to predict intracranial pressure — the blood pressure inside the brain.

“Pressure surrounding the brain is very important because, unlike the rest of the body, the brain can’t expand within the skull,” says Faris Gulamali, a medical student at Mount Sinai, who helped to develop the technique. Excessive intracranial pressure can cause neurological damage and brain deficits, so monitoring how it might change in a patient could allow doctors to act to save them. “If someone suffers a stroke or has an expanding tumour, we can intervene very quickly because we know that they are having an acute change in their intracranial pressure,” he says.

However, unlike blood pressure, which is easy to measure, there’s no non-invasive way to determine intracranial pressure in real time. “The only way to directly measure it would be to place a probe adjacent to the brain or in the spinal canal,” Gulamali says.

Instead, Gulamali and his colleagues explored how intracranial pressure could be predicted using noninvasively collected data. This included electrocardiograms, oxygen saturation levels measured via pulse oximetry, and extracranial waveforms obtained through routine head ultrasounds in intensive care patients.

The researchers leveraged these data to build an AI model, then trained it using real data from patients who had their intracranial pressure measured through invasive means — such as a lumbar catheter or a pressure-sensitive probe inserted through the skull. Finally, they validated the model with clinical data³.

“We have the largest study to date on predicting intracranial hypertension with AI, and the first to provide external validation for an algorithm,” Gulamali says. He hopes that doctors in the ICU could one day use this AI tool to monitor intracranial pressure without having to measure it directly.

And as intracranial hypertension is linked to problems such as glaucoma, and acute liver failure, an AI measure of intracranial pressure could also be useful beyond the ICU.

Medical management

Other research at Mount Sinai shows how hospitals can use AI tools to optimize patient care.

“The human brain does not have the capacity to absorb the sheer volume and different types of data that we generate from our patients right now,” says Eric Nestler, a psychiatrist, director of The Friedman Brain Institute, and dean for academic affairs at the Icahn School of Medicine at Mount Sinai.

Artificial intelligence can recognize patterns within these data that might elude physicians, he says. As an example, Nestler points to all the information that hospital systems collect from blood tests and other medical exams performed on newborns. This can help AI tools predict which infants born with acute problems might have worse outcomes, he says. That can help doctors devote more resources to their care.

Nestler also highlights the growing interest in AI tools to support mental health and emotional well-being. Since patients with mental health issues often experience relapses, continuous monitoring is critical, he says.

Using AI to monitor how a person uses their smartphone can provide clues to their state of mind. “Analysing the tone, volume and cadence of speech and comparing it to their baseline patterns could help detect a patient who is beginning to deteriorate,” he says. Alerts from such automated surveillance tools could prompt caregivers to check on them. Nestler expects AI tools of this kind to assist mental health practitioners in the near future.

Protecting privacy

Harnessing patient data through machine learning demands advanced computing infrastructure, says Patricia Kovatch, dean of scientific computing and data at Icahn Mount Sinai. Mount Sinai has made significant investments to create that infrastructure, she says. It is also essential to protect the privacy of medical data on which AI models are trained, she says — and that has been integral to the health system’s efforts to embrace AI. “We’ve de-identified a lot of the data, to enhance accessibility for everybody,” she says. “We have a million pathology slides here – de-identified, so people can use them, and we have de-identified copies of the structured fields of the medical records.”

Nestler says enormous progress can be made. “We are at the very beginning of mining all the potential advances,” he says.

He envisages a future when doctors will use AI to sift DNA sequences and medical records to find patterns invisible to the human eye and brain. “Putting all that together will yield richer insights into the conditions a patient might present with,” Nestler says. “And that would lead to better precision care.”