An AI Morphology Array’s Potential to Improve Parkinson Diagnosis
John F. Crary, MD, PhD, and Jerry Fernandez, MD, offer insight into PreciseDx's AI Morphology Feature Array and how this technology can improve neuropathology and diagnosis for neurodegenerative diseases, such as Parkinson disease.
John F. Crary, MD, PhD
Gerardo “Jerry” Fernandez, MD
Recently, a study showed that an AI algorithm built by the Mount Sinai spinoff company, PreciseDx, is capable of accurately detecting Parkinson disease (PD) pathology in biopsy sample image patches with 99% sensitivity and 99% specificity as compared with expert annotated ground truth. All told, PreciseDx's AI Morphology Feature Array outperformed human pathologists with an accuracy of 0.69 compared with 0.64 in the prediction of clinical PD status.1,2
Conducted by John F. Crary, MD, PhD, professor of Pathology, Neuroscience, and Artificial Intelligence (AI) & Human Health, Icahn School of Medicine, Mount Sinai, and colleagues, the study assessed the algorithm’s design to immunohistochemically detect α-synuclein within peripheral nerves of salivary glands.
To find out more about the array and algorithm, and how this may play a role in PD care and diagnosis, NeurologyLive® sat down with Crary and Gerardo “Jerry” Fernandez, MD, cofounder and chief scientific officer, PreciseDx.
NeurologyLive®: For anyone who is not familiar with it, could you offer a little bit of background on this AI and how we got to this point?
John F. Crary, MD, PhD: Broad strokes, the major issue is that in diagnosis of Parkinson disease, and most movement disorders, the gold standard is still—despite all of the progress in neuroimaging and biomarkers—the autopsy. This is frustrating, and it makes it extremely difficult for families to get an accurate diagnosis, but also to run clinical trials to adequately stratify your patients. So really good biomarkers—neuroimaging, blood, or otherwise—are extremely valuable. What makes Parkinson disease different from essentially every other neurodegenerative disorder is there's this unique thing in the pathology that we see in the central nervous system and can also be observed in the peripheral nervous system. This has been recognized for many, many years—you can see Lewy body pathology, which is the telltale lesion. These are inclusions in the abnormal cytoplasmic changes made of the α-synuclein protein. You can find abnormal pathology in the colon, you can find it in the skin, you can find it in glands all throughout the body.
There is a large study that was conducted by the Michael J. Fox Foundation called the S4, where they looked at these peripheral biopsies. The idea of biopsy is seeing these peripheral tissues as a diagnostic and prognostic biomarker for Parkinson disease. They did it in a standardized way using rigorous clinical trial type protocols and compared it head-to-head with cerebrospinal fluid and other biomarkers to see the extent to which the peripheral biomarker, the biopsy, stained immunohistochemically for α-synuclein—whether the pathology could be detected there and whether it could be used to diagnose Parkinson disease. And, actually it can.
One of the major issues with that is that the pathology is very, very hard to find. Screening the slides and finding these peripheral axons is super time consuming and very difficult. I was trained to do it, along with the other neuropathologists on the project, but throughout the entire time I was thinking, “There's got to be a better way to do this. It'd be amazing if we could if we could train an AI to do it.” That's where this project came from.
As far as the clinical community is concerned, what should they know about this AI and what’s been done with it?
Jerry Fernandez, MD: The group that I worked for at Mount Sinai now became a company called PreciseDx. Our mission really was to develop a platform that was able to interrogate malignancies, cancer tissues, and generate more robust and more reproducible grading systems that would add prognostic information to a patient's management course. We've developed a breast cancer test and a prostate cancer test that was targeted to give added information to the oncologist, so the oncologist and the patient can better select therapies for the treatment.
When I came to Mount Sinai, I met John and we were chatting it up, and we hit it off pretty well early on, and I was telling him about what I did. And as John is a specialist in Alzheimer, he started putting two-and-two together and he said, “Hey, why don't we try using your oncology platform to characterize tauopathies in autopsies of patients with Alzheimer's disease?” So, we did this study a couple of years ago, where we applied this oncology platform to Alzheimer disease, and were very, very successful in characterizing an accurately assessing tauopathies in these patients that had died with Alzheimer disease.
The second part of the relationship came up when he had finished this S4 project and said, “We did it with tau, let's do it with α-synuclein in these peripheral biopsies.” That's how we came together and started working on this from an AI standpoint.
Could you describe the technology a little further? How does it work?
Jerry Fernandez, MD: We have a platform, which we call the morphology array. Basically, it's analogous to a gene array, where we generate about 10,000 to 15,000 features every time we run a tissue through our platform. Then, we're able to correlate any one of those features with whatever clinical data that we have available to us. In some cases, we correlate it to cancer tumor grade. In other cases, most of the time, we correlate to outcomes. So, for a patient with cancer who dies in 3 years versus one that dies in 15 years, we're able to find features in that morphology array that are able to predict which is which. We build predictive models that use these aspects of morphology to predict outcomes.
When we tried it with immunohistochemistry, and tau, that was the first time we had tried that, and there is no reason to believe it wasn't going to work. But you know, we built similar features that we built for cancer prediction. We built the same features using the same platform except for instead of looking at H&E stain sections, we were looking at tau-stained sections. We did it so successfully with Alzheimer, that we did the exact same thing with Parkinson—the features that come out are different, but the platform is based on what we call components. We identify histologic components within the tissue, and then the features that are extracted are basically either mathematical characterization of those objects or are mathematical derivations of those objects. By that, I mean that we look for relationships. So we look for distribution in tissue, we look for shapes and sizes, and the consistency of shapes and sizes throughout the tissue. It's a variety of different approaches, which all are essentially visual characteristics, but that the human can't quantitate them. They can visualize them, but they can't quantitate them, and the AI does a good job of not only detecting them, but also quantitating those visual objects. The platform is basically based on extracting the relationship information from the data.
Transcript edited for clarity.
