You’re sitting in a hospital waiting room. The furniture, the walls, and the air all feel sterile and sanitized. There’s a disproportionate number of tissue boxes for the cramped room, and the doctor informs you that they didn’t identify the disease fast enough. This is a reality for thousands of people in America. Researchers at Johns Hopkins estimate that every year, 795,000 Americans become permanently disabled or die due to medical misdiagnosis [1]. Neurological conditions are uniquely difficult to diagnose, as many neurodegenerative conditions can share symptoms and causes [2]. One study found that one in four Alzheimer’s patients had been previously misdiagnosed [3]. These types of diagnostic errors cause widespread but preventable harm.
One especially promising avenue to lower misdiagnosis rates is the use of artificial intelligence (AI), specifically machine learning (ML) models, to economically and accurately diagnose neurological conditions. It’s 2024; you’re probably tired of hearing about AI, but its potential applications are too numerous to be ignored. ML models can assist doctors, researchers, and patients by efficiently identifying diseases by analyzing data from brain scans, medical histories, and biological samples to find patterns and connections. As these models improve in accuracy and generalizability, they can broaden access to medical care and lower hospitals’ labor burden, increasing the quality of care for everyone.
In a new study by the Mayo Clinic, researchers created an ML model to explore the capabilities of AI in diagnosing neurodegenerative conditions [4]. Researchers used a deep learning algorithm, a branch of ML that uses artificial neural networks inspired by the structure of neurons in animal brains. Artificial neural networks are computational models made up of layers of interconnected nodes, or so-called neurons, that process information by mimicking the way biological neurons transmit signals. Each neuron receives inputs, applies a mathematical transformation, and passes the result to the next layer, allowing the network to learn complex patterns and make predictions or classifications based on the data. The researchers then utilized weakly supervised feature learning, a technique where a computer learns to find patterns, infer relationships, and make connections in data even when the labels or guidance it gets are incomplete or not very detailed. For example, a model could be given digital images of farm animals, some with vague labels such as 'has four legs' or 'makes wool.' Given enough time and resources, the model could learn to group each image by animal type despite not understanding the concept of an animal, let alone being able to differentiate between species. This is how machines can learn to perform complex tasks, and just as a model can recognize farm animals, a sufficiently advanced ML model could recognize the signs of neurodegenerative disease [4].
Researchers extracted the tauopathy-positive samples from patients diagnosed with tauopathies, a subset of neurodegenerative disorders including Alzheimer’s Disease (AD) or corticobasal degeneration (CBD) [4]. Tauopathies are associated with abnormal deposits of tau proteins, a protein that stabilizes a neuron’s internal skeleton [5]. An aberrant build-up of tau proteins can cause neurons to deteriorate, resulting in various cognitive impairments [5]. These samples were extracted post-mortem as part of the Mayo Clinic Brain Bank for Neurodegenerative Disorders. These samples were specifically derived from the peri-Rolandic primary motor and sensory cortices, the cingulate gyrus with adjacent regions, and the corpus striatum regions of the brain due to their shared vulnerability to tau pathology and distinct disease-specific patterns critical for differentiating tauopathies. Tau was visualized using immunohistochemistry with phospho-tau-specific antibodies. Researchers then digitized the tauopathy-positive samples in preparation for training an ML model on the images [4].
First, researchers needed to identify labels for the samples to train the model with. They identified white matter with abnormally high amounts of tau proteins, lesions associated with tau deposits, and other indicators of tauopathies to use as training labels. Then, researchers linked those labels to certain neurodegenerative disorders, such as Alzheimer’s disease or Pick disease, and trained the models to predict tauopathies based on a single sample. The model achieved a high level of accuracy, with all models, each representing a different neurodegenerative disease, having an area-under-the-curve (AUC) score — which is an aggregate measure of performance — greater than 0.90, meaning their models were extremely accurate [4].
Although the researchers’ models were very successful, they were limited to the six tauopathies they trained for. Given the diversity of symptoms and signs in neurological conditions, general application remains difficult. However, the study’s results suggest that ML models are capable of diagnosing a variety of neurodegenerative diseases. This is a rapidly developing field, with breakthroughs in the field of AI and an increased awareness of the technology’s possibilities. While research has been conducted on ML applications in other fields, such as oncology or epidemiology, this was one of the first studies that used ML models to identify tauopathies.
A study that measured the diagnostic accuracy of neurological conditions found that a doctor’s initial diagnoses matched the final diagnosis in only 60.4% of cases [6]. Compared to the Mayo Clinic researcher’s models, some of which showed diagnostic accuracies of >95%, it is clear that ML models may possess the ability to rival, if not exceed human doctors. While the possibilities of AI in neuroscience remain largely untapped, the advancement of ML models is an exciting development that has the potential to reduce price-related barriers and increase the quality of care for all. Particularly, this emerging research highlights the potential of machine learning models in achieving unprecedented levels of accuracy in identifying tauopathies, with these models not only demonstrating their ability to rival human diagnostic capabilities but also paving the way for earlier, more precise interventions. As AI technology continues to advance, its integration into clinical practice could significantly reduce misdiagnosis rates and catch these diseases in earlier stages, enabling earlier intervention and improving outcomes for countless patients worldwide.