Today, researchers at Mount Sinai’s Department of Neurology, led by artificial intelligence (AI)-driven models, are making bold progress in this arena, having developed a sophisticated stroke identification model trained on an array of structured data and unstructured clinical notes available in the electronic health record at the time of stroke alert, offering the potential to optimize stroke triage and reduce false-positive alert activations. The effort is among several AI-driven models being engineered within the department for the detection of perplexing and widespread conditions, such as migraine headache and REM-behavior disorder (RBD).
“The time sensitivity of stroke treatment and the diagnostic uncertainty of evaluating a patient with acute neurological symptoms have generated interest in leveraging AI and machine learning to improve the specificity of stroke alert systems,” says Benjamin Kummer, MD, Associate Professor of Neurology and Director of Clinical Informatics in Neurology at the Icahn School of Medicine at Mount Sinai.
Benjamin Kummer, MD, and Asala Erekat, PhD.
Dr. Kummer recently authored a preprint manuscript on medRxiv describing the groundbreaking stroke alert system he and his multidisciplinary team have developed, including Mark Kupersmith, MD, in the Department of Ophthalmology, and Asala Erekat, PhD, in the Department of Neurology. Dr. Kummer also holds an appointment in the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai.
“We’re close to finalizing and clinically testing a multimodal system that replicates—but doesn’t try to replace—the decision-making process for distinguishing patients with acute stroke from patients without stroke,” says Dr. Kummer. “This type of digital neurology for disease detection is clearly the wave of the future.”
As a neurologist and clinical informaticist, Dr. Kummer has focused the work of his lab on applying emerging large-language model (LLM) techniques to the development of homegrown models such as this one—which is generating robust results. In recent years, LLMs have emerged as a powerful tool for mining clinically relevant insights from unstructured free-text data, which often contain nuanced information about patient presentations that structured data alone may not capture.
“What makes this project so compelling and exciting is that it addresses a really difficult problem faced by many providers,” says Dr. Kummer, who serves as a key subject matter expert for the Neurology department on health informatics initiatives. “In addition, it combines different sources of data, including notes that clinicians, nurses, and Emergency Department triage personnel write, as well as a patient’s lab values, medical history, and medications taken before coming to the hospital.” He explains that the model uses LLM techniques to extract meaning from clinical notes, and combines predictions made by multiple smaller “sub-models” to ultimately distinguish acute cerebrovascular disease from non-stroke causes, such as sepsis, metabolic abnormalities, or respiratory failure.
The team has demonstrated that ensemble machine learning approaches can integrate both structured and unstructured data to provide an alluring pathway forward for enhancing the accuracy and performance of stroke alert systems.
Dr. Kummer describes the stroke alert model’s accuracy as most encouraging at this early stage, while acknowledging that the model remains a prototype that will continuously improve in precision as it moves from the lab—where it has trained on 10-year retrospective data from Mount Sinai Health System stroke alerts—to the clinical arena. “Our goal is to initially run this model in the background so no one knows it’s running and it’s not actually showing clinicians its predictions, but the model is learning from its own mistakes going forward,” he explains. “What we’ve demonstrated so far is that ensemble machine learning approaches can integrate both structured and unstructured data to provide an alluring pathway forward for enhancing the accuracy and performance of stroke alert systems.”
The vast promise of machine learning and AI is being further applied to migraine headache. Predicting a migraine attack is one of the most challenging arenas for automated systems given the heterogeneity of underlying headache triggers for patients.
Under the leadership of Bridget R. Mueller, MD, PhD, Assistant Professor of Neurology and Director of Research for the David S. and Ruth L. Gottesman Center for Headache Treatment and Translational Research, a prototype model has been developed based on data from high-quality wearable devices that continuously monitor patients’ autonomic nervous systems. From this information are emerging personalized predictive algorithms to forecast the likelihood of a next-day migraine attack based on physiological signals originating from the autonomic nervous system from the preceding night.
Moreover, data that’s passively gathered by nocturnal video camera monitoring is figuring in the development of machine learning algorithms for detecting idiopathic REM sleep behavior disorder, a known precursor for Parkinson’s disease. This project, led by sleep specialist Emmanuel During, MD, Associate Professor of Neurology, developed models to analyze videos of sleeping patients, and from that data identify RBD with high accuracy in hopes of developing a plan for patients deemed at risk of developing Parkinson’s disease [see related story here].
“Along with other major institutions that are pioneering work in this area, we’re taking some of the critical first steps toward realizing the enormous potential of AI across the diverse sub-fields of neurology,” sums up Dr. Kummer. “We’ve always been a very traditional discipline, but we’re digitizing quickly now in ways that promise to dramatically change how we diagnose and treat neurological disorders.”