Researchers from the University of Toronto and the Universidad de las Américas Puebla have developed an advanced machine learning methodology to improve the detection of Parkinson’s disease using brain imaging techniques that track brain activity during rest. This study, led by Dr. Gabriel Solana-Lavalle and colleagues, applies a Causal Forest machine learning algorithm to analyze patterns of brain activity, offering a highly accurate method for identifying Parkinson’s disease while also revealing the brain regions most impacted by the disease. The findings were published in Tomography.
Dr. Solana-Lavalle, along with Professor Michael Cusimano, Dr. Thomas Steeves, Professor Roberto Rosas-Romero, and Dr. Pascal Tyrrell, devised a machine learning model that processes brain scan data to accurately classify Parkinson’s disease patients. “Our method focuses on a unique combination of reducing unnecessary data while making sure that we can still clearly understand which brain areas are affected by Parkinson’s disease” Dr. Solana-Lavalle explained.
The research team analyzed data from the Parkinson’s Progression Markers Initiative and additional control data from another public database that collects brain scans from various research sites. They processed brain scans from more than two hundred individuals, applying Causal Forest and Wrapper Feature Subset Selection algorithms to filter out noise and unnecessary information and focus on the brain regions most strongly associated with Parkinson’s disease while optimizing classifier performance.
To manage variations in data quality and acquisition conditions, the team used advanced data processing techniques, including image alignment and standardization. “This data-driven approach provides interpretable insights into brain regions strongly associated with Parkinson’s disease, which can help clinicians better understand disease progression and personalize treatments,” Dr. Solana-Lavalle added.
The study identified specific brain regions showing significant changes in Parkinson’s disease patients compared to healthy controls. The Causal Forest algorithm ranked these regions by their relevance, enabling the use of statistical tools for visualization and interpretation of the activation patterns that differentiate between Parkinson’s disease and non-affected groups. The method was effective across different subsets of the population, showing strong accuracy for both men and women.
The potential of this approach extends beyond diagnosis, offering insights into how Parkinson’s disease affects different brain areas. The method also identified a correlation between the activations in certain brain regions and the motor section of the UPDRS, a clinical assessment tool that measures various motor functions.
This research lays the groundwork for future studies aimed at improving machine learning models for other neurodegenerative diseases. By emphasizing interpretability alongside performance, the method could help clinicians diagnose Parkinson’s disease more effectively and understand its varied impacts across patients.
The study represents a significant advancement in applying machine learning to medical imaging and neurodegenerative disease detection. Moving forward, Dr. Solana-Lavalle and his team plan to expand their approach to include long-term studies, hoping to track the progression of Parkinson’s disease over time.
Journal Reference
Solana-Lavalle, G., Cusimano, M. D., Steeves, T., Rosas-Romero, R., & Tyrrell, P. N. (2024). “Causal Forest Machine Learning Analysis of Parkinson’s Disease in Resting-State Functional Magnetic Resonance Imaging.” Tomography. DOI: https://doi.org/10.3390/tomography10060068
About the Authors
Gabriel Solana-Lavalle earned his Ph.D. in Smart Systems from the Universidad de las Américas, Puebla, México, in 2023. His research interests encompass signal processing, medical imaging analysis, forecasting, and machine learning. In 2022, he was an international visiting graduate student at the Institute of Medical Science, University of Toronto. He is currently collaborating with industry partners on projects aimed at developing and implementing innovative technologies in signal processing for medical imaging.
Pascal Tyrrell, an accomplished data scientist, is the Director of Data Science and an Associate Professor at the University of Toronto’s Department of Medical Imaging. He founded the MiDATA data science program and holds appointments in the Institute of Medical Science and the Department of Statistical Sciences. His research applies innovative Artificial Intelligence to medical image analysis for improved health outcomes. Pascal is also a serial entrepreneur with experience spanning computer software, medical devices, and agri-tech.
Professor Roberto Rosas-Romero received a Ph. D. Degree in Electrical Engineering from University of Washington. He holds the position of Professor at the Department of Electrical & Computer Engineering, Universidad de las Américas-Puebla (México). He was a Visiting Professor at the Department of Diagnostic Radiology at Yale University. He has been a Fulbright Scholar twice, as a student at the University of Washington and as visiting professor at Yale, respectively. His research interests are Signal Processing, Computer Vision, Pattern Recognition, Machine Learning, and Medical Image Analysis. His research has been applied to ultrasound image segmentation, forest fire detection from video signals, micro-aneurysm detection in fundus eye images to assist in the diagnosis of diabetic retinopathy, prediction of epileptic seizures based on brain waves, detection of deafness in newborn cries, detection of micro-calcifications on mammograms, Parkinson’s disease detection by analyzing voice, classification of magnetic resonance images to assist Parkinson’s disease diagnosis, classification of skin burns in color images.
Michael D. Cusimano is a neurosurgeon and Professor of Neurosurgery and Public Health Sciences at the University of Toronto. As Canada’s first formally trained skull base surgeon, he developed the now globally adopted dual-nostril fully endoscopic approach in 1993. His body of published work comprises three books, including the co-authored Handbook of Skull Base Surgery, and more than 450 publications in all fields of neurosurgery from basic science to clinical outcomes. In addition to being one of the foremost and most sought-after neurosurgeons in the country, he is an internationally recognized expert in traumatic brain injury and his work has helped to transform public awareness of concussion in the general public and contributed to changes in policies and rules at all levels of sports world-wide. His highly collaborative work also highlights the importance of patient quality of life assessment and a career long use of the latest advanced data analytics, particularly in applying measurement, artificial intelligence, and geography to medicine. Dr. Cusimano founded the St. Michael’s Hospital Injury Prevention Research Office, was National Director of Research and then the Vice President of the Think First National Injury Prevention Foundation for over a decade, led the Canadian CIHR Team in Traumatic Brain Injury and Violence, is a scientific advisor in concussion for the Brain Trauma Foundation, and, a Fellow of the Canadian Academy of Health Sciences acknowledging his contributions to surgery and impact on public policy nationally and internationally. With a Ph.D in Education, he has promoted the development of medical-surgical education and evaluation models, and has been dedicated to educating the general public and a generation of doctors and neurosurgeons who contribute to the field today. He is an outspoken advocate for brain health and brain injury prevention.
