I’ve been doing a lot of research on healthcare AI applications lately. Not surprisingly, while people find the abstract issues involved to be intriguing, most would prefer to hear news of real-life projects, so I’ve been on the lookout for good examples.
One interesting case study, which appeared recently in Health IT Analytics, comes from Montefiore Health System, which has been building up its AI capabilities. Over the past three years, it has created an AI framework leveraging a data lake, infrastructure upgrades and predictive analytics algorithms. The AI is focused on addressing expensive, dangerous health issues, HIA reports.
“We have created a system that harvests every piece of data that we can possibly find, from our own EMRs and devices to patient-generated data to socio-economic data from the community,” said Parsa Mirhaji, MD, PhD, director of the Center for Health Data Innovations at Montefiore and the Albert Einstein College of Medicine, who spoke with the publication.
Back in 2015, Mirhaji kicked off a project bringing semantic data lake technology to his organization. The first pilot using the technology was designed to find patients at risk of death or intubation within 48 hours. Now, clinicians can also see red flags for admitted patients with increased risk of mortality 3 to 5 days in advance.
In 2017, the health system also rolled out advanced sepsis detection tools and a respiratory failure detection algorithm called APPROVE, which identifies patients at a raised risk of prolonged ventilation up to 48 hours before onset, HIA reported.
The net result of these efforts was dubbed PALM, the Patient-centered Analytical Learning Machine. PALM “represents a very new way of interacting with data in healthcare,” Miraji told HIA.
What makes PALM special is that it speeds up the process of collecting, curating, cleaning and accessing metadata which must be conducted before the data can be used to train AI models. In most cases, the process of collecting data for AI use is largely manual, but PALM automates this process, Miraji told the publication.
This is because the data lake and its graph repositories can find relationships between individual data elements on an on-the-fly basis. This automation lets Montefiore cut way down on labor needed to get these results. Miraji noted that ordinarily, it would take a team of data analysts, database administrators and designers to achieve this result.
PALM also benefits from a souped-up hardware architecture, which Montefiore created with help from Intel and other technology partners. The improved architecture includes the capacity for more system memory and processing power.
The final step in optimizing the PALM system was to integrate it into the health system’s clinical workflow. This seems to have been the hardest step. “I will say right away that I don’t think we have completely solved the problem of integrating analytics seamlessly into the workflow,” Miraji admitted to HIA.