Montefiore Health System Improves Patient Outcomes and Healthcare Efficiency with Semantic Data Lake and Artificial Intelligence Powered by Intel Technologies


Untold medical breakthroughs are trapped within the data silos that persist even in the current digital era. Unlocking the potential in this data requires uniting disparate databases that may have different nomenclatures or other structural incompatibilities — a significant challenge that typically requires costly manual labor to address.

Montefiore, a premier academic health system in the Bronx, has found a promising solution in its Patient-centered Analytical Machine Learning (PALM) platform, which has been designed to tap into myriad data stores, regardless of where the information is located or how it is structured. Montefiore relies on PALM and the memory capacity offered by Intel® Xeon® Scalable processors to help medical practitioners identify patients at risk for respiratory failure — improving patient outcomes and lowering costs — and is already starting to apply PALM to a variety of other projects.

Untapped Big Data Potential, Some Assembly Required into Data Lake Format

Big data is often a prerequisite to implementing AI algorithms. It also can be one of the biggest challenges when preparing and deploying a solution. Healthcare drives some of the biggest of big data, with images, genomic tests, etc. But health data is also complex and can come from a variety of disparate sources…not just the electronic medical record. As a result, the data itself cannot easily be pulled together. For example, potassium can be identified in at least 47 different ways. Additionally, metadata — companion information researchers rely upon for building accurate models — may be lacking or absent from medical records and other databases used in a health system. Further, personal information must be redacted and patient identities protected in order to insure compliance with privacy regulations and responsible use of the data. These challenges often increase the cost and development time of intelligent healthcare systems, as well as narrowing their scope and reducing their ability to scale broadly.

PALM’s key capability is to enable access to data regardless of where the data is located or how the data is structured. It does not matter if the data is in traditional databases or in newer unstructured data stores that might contain voice, images and sensor inputs. PALM is able to do this through a semantic data lake architecture that combines Montefiore’s multiple stores of medical data with various ontological databases that define over 2.5 million terms and their relationships to one another. For example, there are ontological databases of ailments, medications, and genetic disorders. These ontological databases can be used to define subject-predicate-object relationships between data (“John has hives,” or “Jill takes Ibuprofen”), to associate new data points to known data types (“Ibuprofen” is a drug, “Jack” is a human), and disambiguate terminology (linking “hives” to “rashes” or “Advil*” to “Ibuprofen”).

Additionally, PALM integrates an “Analytic Tapestry” that detects and connects algorithms doing similar work (for example, reducing hospital readmissions) in order to improve algorithm performance. Finally, PALM is ready for what may come next - it can quickly and easily integrate new data and ontologies, including information like voice, images, and sensor inputs from internet-of-things (IoT) devices.

Semantic data lakes like Montefiore’s impose particular system architecture requirements. The tailor-made databases that semantic data lakes create require large amounts of system memory, performant multi-core compute, and super-fast storage. As we’ve seen with other real-world AI applications, these demands mean that PALM requires Intel architecture to work its magic.

Architecture for Demanding AI Applications

Montefiore is using several different Intel-based systems to power PALM. Most interesting is the newest system, which uses two 20-core Intel® Xeon® Gold processors and two 375 GB Intel® Optane™ SSDs, as well as an additional 800GB traditional SSD.

Intel Optane SSDs can be configured for fast storage, or as a method to extend system memory, which means that this system has the potential to be a semantic data lake powerhouse. PALM’s developers have already gotten a glimpse of what Intel Optane SSDs can do for platform performance and are excited for their potential to augment the capabilities of the Intel Xeon Scalable platform.

These systems deliver to Montefiore not only the memory capacity and storage that PALM’s semantic data lake and ad hoc databases require, but also the versatility and flexibility to support multiple server configurations and integrate with legacy hardware, all within the same, well-known technology ecosystem.

Real-world AI Results

PALM was first applied in a pilot to predict respiratory failure, which the team defined by the need to put the patient on a respirator. The consequences of respiratory failure are severe, with as many as 50% of patients dying within 6 months of the event[1], and roughly 35 percent don’t survive hospital discharge.[2]

With support from the National Institutes of Health (NIH), Montefiore researchers used PALM to develop a machine learning model to help identify patients at high risk for respiratory failure or death in the hospital. There’s no room for error when treating critical care patients, so the team then spent months validating the model’s clinical reliability and safety by letting the model practice on live, real-time hospital data. This testing ensured that the model was working as intended and wouldn’t surprise practitioners with unintended consequences.

Following validation, the system was first deployed in early-2017 at the smallest of Montefiore’s three Bronx hospitals. The program has since been expanded to the other two hospitals.

Montefiore’s respiratory failure prediction application is designed to alert practitioners with what is known as a “Best-Practice Advisory,” or BPA, called APPROVE, or Accurate Prediction of Prolonged VEntilation. It recommends, amongst other actions, consulting with a physician from the critical care department. The APPROVE BPA is triggered when the system detects that a patient’s risk score exceeds a 0.25 threshold on a scale of zero to one, a threshold found to identify about two-thirds of patients who end up requiring ventilation or die in the hospital.

One key indicator of this system’s value to practitioners is the uptake of APPROVE BPA messages. Alert fatigue is a well-documented issue in hospitals, not just from these new predictive algorithms but also from monitoring devices and other systems. However, the PALM team has found that at least one clinician interacts with the APPROVE alert about 93% of the time it is issued, a good indication that APPROVE is regarded as a useful alert. Clinician reviews have been consistently positive, with some crediting APPROVE with helping them spot subtle but critical markers. With the APPROVE system and PALM, Montefiore can give its clinicians a head start when treating its sickest patients.

Widely-Varying Applications

Montefiore is now seeking to harness PALM’s predictive AI and the versatility imparted by its semantic data lake to a wide variety of other use cases in the health system. Efforts are underway to use PALM to more effectively route people to the most appropriate care, predict appointment no-shows that waste precious care resources, and forecast and allocate hospital beds to more efficiently house patients and reduce their length of stay. Tapping PALM for a compliance reporting project has led to a pilot to detect the earliest signs of sepsis in patients. The team is even working to find ways to reduce or eliminate irrelevant BPA messages.

Regardless of the application, PALM’s need for large amounts of system memory, incredibly fast storage, and many performant compute cores indicates that Intel Xeon Scalable processors and Intel Optane SSDs will continue to play a critical role in delivering these new AI capabilities. With Intel’s solutions for AI, all of this can occur on the same architecture already in use for so many other traditional enterprise activities, increasing efficiency and improving time to value.

Please consult Embedded Analyst: AI Without Borders, the case study we commissioned, to learn more about Montefiore Health System’s PALM platform. You can also read about the Montefiore’s PALM system in this VentureBeat story. To learn about Intel AI’s possibilities for your organization, please visit

Stay Connected

Keep tabs on all the latest news with our monthly newsletter.