MedyMatch Technology, a company based in Tel Aviv, Israel, leverages artificial intelligence, deep learning, and computer vision technologies to offer patient-specific clinical decision support. Their application helps radiologists and emergency room physicians to detect signs of intracranial hemorrhages, which are difficult to diagnose by standard analysis of imaging data alone. The Medgadget team recently had an opportunity to speak with Gene Saragnese, Chairman and Chief Executive Officer of MedyMatch, to discuss their technology and its significance in depth.
Prior to joining MedyMatch in January of 2016, Gene was the Chief Executive Officer of Philips Imaging and a member of Philips Healthcare’s Executive Team. A graduate of Rutgers College of Engineering in New Jersey, he has also previously served as GE Healthcare’s Chief Technology Officer and has held management roles with GE, RCA, Martin Marietta, and Lockheed Martin. Our interview with MedyMatch’s lead executive is given in full below.
Zach Kaufman, Medgadget: Gene, thank you so much for taking the time to speak with us on behalf of MedyMatch! To start us off, I thought I would offer you an opportunity to highlight any features or applications of the technology that Medgadget readers might find of particular interest or importance.
Gene Saragnese, MedyMatch: MedyMatch will make available several applications to be utilized in the acute care setting. The first application, for clinical research, is for the detection of intracranial hemorrhage. Upon regulatory approval, it will be utilized as a ‘second read’ by physicians, to identify areas of suspected bleeds. MedyMatch integrates seamlessly into the hospital workflow by extracting medical images from the PACS, processing in the cloud, and returning annotated images to the PACS for viewing by the physicians. In initial preclinical trial results, MedyMatch achieved a ~97% sensitivity with 90% specificity, and a ~96% Negative Predictive Value (NPV) for intracranial hemorrhage detection on hundreds of thousands of non-contrast CT slices.
Medgadget: MedyMatch relies on “Deep Vision” technology, as you term it, to aid radiologists and emergency room physicians in identifying intracranial hemorrhages related to brain trauma and strokes, which are difficult to detect by looking at medical imaging scans directly. What types of abnormalities and indications are medical professionals most susceptible to missing in analyzing imaging data with the human eye?
Saragnese: Extensive studies have been performed on this topic. Research summarized by the National Institutes of Health in January 2012 indicates there are a wide range of causes.
Medgadget: With what frequency do errors in medical imaging diagnosis occur using status quo techniques?
Saragnese: In some cases, the detection of a brain bleed is difficult. Studies have shown that the error rate in the detection of brain bleed by emergency room physicians is estimated to be between 20-30%. Additionally, studies have shown 80% of medical imaging diagnosis errors are evident, but remain unnoticed by the physician.
Medgadget: You mention that MedyMatch will be made available as a prioritization algorithm operating within radiological PACS systems, and I understand that you are aiming to enable implementations with computer assisted detection (CAD) tools as well. Can you tell us a bit about how the different forms the technology might take will vary from one another in terms of their applications, direct users, and benefits?
Saragnese: MedyMatch is a near zero footprint solution. We do not sell workstations, viewers, or imaging systems. Rather, we simply process patient studies that are either sent to a server on the premises or directed to a secure, HIPPA-compliant cloud. Patient studies are annotated and regions of interest are marked before being reinserted into the customers current PACS system for consideration by the physician. It is foreseen that the algorithm can also advise the PACS system in the background of a potential finding. This signal will be considered as part of the prioritization process in PACS.
Medgadget: Are the deep vision algorithms involved in identifying hemorrhages suited only for acute, patient-specific investigation, or can the technology be utilized to offer proactive, population-level information as well?
Saragnese: MedyMatch has developed a generalized deep learning platform. This means that we will be able to apply the company’s technology, techniques, and methods to a wide range of clinical areas. The application of the technology is virtually limitless. However, MedyMatch is not a clinical research company, so applications are to be built where variability in reader performance can be reduced, thereby improving patient outcomes while reducing overall healthcare costs.
Medgadget: Regarding the algorithms themselves, what types of non-imaging data are integrated into the deep learning and computer vision analysis? How does that supplemental data help to inform diagnoses beyond traditional CAD tools?
Saragnese: Traditional CAD considers imaging data, while Deep Vision (deep learning and computer vision) techniques consider the whole patient. Data such as EMR, genomic, lab results, etc., can all be integrated into the data set that the algorithms consider. The more data, the higher the probability for a more personalized assessment.
Medgadget: Once the MedyMatch system produces its insights, what is the role of the radiologist or physician in interpreting the outputs?
Saragnese: MedyMatch will never replace the radiologist or physician. MedyMatch is a visual clinical decision support tool to provide a second set of eyes. MedyMatch can only assess a patient; the final diagnosis will always be the responsibility of the doctor.
Medgadget: Is there an opportunity for users of the system to integrate their feedback in an effort to continue to fine-tune the software once in use? From my basic knowledge of deep learning, I know that the algorithms involved are rarely static and, in fact, generally continuously improve as they encounter new and wide-ranging data sets. Is it reasonable to expect that MedyMatch could be getting ‘better’ as it sees more broad use?
Saragnese: Continuous learning is at the very heart of the intellectual property of the company. In fact, the concept of feedback in A.I. was pioneered by MedyMatch and discussed in our patent filings. Real world feedback is critical for improvement. MedyMatch has been compared to Waze in healthcare; just as user feedback optimizes which driving routes are recommended to users, so too will the MedyMatch feedback mechanism assist in the development of the next generation of algorithms required for patient assessment.
Medgadget: Is MedyMatch designed to be local or cloud-based software, or some combination depending on the specific implementation? Are there limitations due to computational processing power or other factors as to how and where the technology can be used?
Saragnese: MedyMatch can be deployed in either a cloud or on-premises solution. Limitations of performance are self-designed and are usually limited by budget availability and local infrastructure. Customer expectations are reasonable, and patient studies can be processed in just a few minutes.
Medgadget: I imagine that identifying hemorrhages is only the first of many potential applications of the 3D deep vision platform that you have developed. Do you see an opportunity to apply the technology as a patient assessment tool for other diseases? If so, what might those solutions look like and how far away might they be?
Saragnese: MedyMatch has developed a generalized deep vision platform capable of considering the full richness of medical imaging along with any other patient data. This platform and A.I. approach will facilitate rapid discovery and decision support development. In addition, MedyMatch’s has core IP in continuous learning. This IP will allow MedyMatch to harvest only the ‘right data’ to very cost-effectively accelerate continuous learning and accuracy. Platform and continuous learning will rapidly propel MedyMatch into adjacent decision support opportunities. Yes; there are significant opportunities to apply our technology to other acute diseases. Product development is currently underway, and we expect to see our next generation of applications to be made available over the course of 2017.
Medgadget: MedyMatch states one of its main objectives to be reducing healthcare costs. Can you describe how leveraging artificial intelligence based image classification software for clinical decision support might ultimately reduce the cost of care?
Saragnese: Our goal is to deliver A.I.-based, patient-specific, clinical decision support applications to improve quality and outcomes while reducing healthcare costs. To accomplish this, we consider all of the multidimensional patient data (i.e., raw imaging concurrently with other relevant patient data at the leading edge of machine learning technologies). We want every physician to be a life-saving expert, every time. This is what drives us forward every day. MedyMatch is committed to improving real-time decisions in the acute care setting with a laser focused on those decisions in the emergency room which have the largest impact on outcomes and healthcare cost. Stroke decision support is our initial focus area due to the need to make rapid, accurate, and timely decisions. We must create capacity to care for more patients with less resources per patient, as well as access to care everywhere in the world. We at MedyMatch are committed to improving quality and reducing errors through A.I.-based decision support. This in turn will improve outcomes and reduce costs.
Medgadget: As the holiday season approaches and this year comes to a close, what can we look forward to seeing from MedyMatch in 2017?
Saragnese: Our intent is to continue to reapply our A.I. platform capability to new and diverse clinical problems with interest in continuing to build out capability in the acute care ER setting with a natural extension into trauma. Structured problem solving and collaboration are key to realizing the full potential of Deep Vision. With the right partners and data, there is a strong desire and potential to address more chronic diseases, such as neurodegenerative disease, cerebrovascular disease, and PTSD. The pipeline of potential treatments will require definitive complementary diagnosis and prognosis. This is an ideal challenge for deep machine vision and learning.
Medgadget: Thank you again, Gene, for sharing your thoughts with us! Our best wishes to the MedyMatch team, and we look forward to following along with your progress!
Saragnese: You’re welcome!
Link: MedyMatch homepage…