A recent Nature review, describes the latest developments in the use of machine learning approaches to interrogate datasets in neurodegenerative diseases. The authors discuss the potential uses of ML in early diagnosis, prognosis and the development of new therapeutics as well as the challenges in the integration of ML in data analysis.
ML is commonly used in healthcare settings to enable robust interrogation of multiple datasets to identify undiscovered patterns and relationships between different elements in the data. In recent years, the application of ML in research has been widely discussed, showing some promise in areas of diagnosis, prognosis and drug development.
In this review, researchers at Sheffield Institute of Translational Neuroscience and BenevolentAI discuss the potential use of ML to combat the current challenges in analysing complex data from neurodegenerative diseases.
The rapid accumulation of data in the past decade due to advancements in technology has resulted in an increasing amount of health datasets that are high-dimensional, meaning the number of features can exceed the number of observations. In other words, the vast amounts of data make it very difficult to gain any relevant biological insight just from standard analytical approaches. As a result, many ML approaches have been developed to try and overcome these challenges by reducing the number of variables analysed.
It is estimated that by the year 2050, 22% of the global population will be over the age of 60. This poses a huge unprecedented economic challenge for countries across the world due to age being the main risk factor for neurodegenerative disorders such as Alzheimer’s and Parkinson’s disease. This estimate emphasises the urgency for the development of new diagnostic approaches and therapeutic interventions. The researchers suggest that ML could provide a solution in this context as it could not only provide insight into disease mechanisms, it could also help with early diagnosis, prognosis, patient stratification and drug development.
The review discusses the first area of neurology that benefited from AI – neuroimaging. As machine learning is based on evidence, it provided an objective diagnostic approach from medical images. More recently, ML methods have been applied to key indicators, including motor function and language feature analysis, where it showed some promise in reducing the time taken to perform clinical assessment.
Nonetheless, the use of ML has several challenges including the lack of large datasets, particularly multidimensional patient data, that ML models rely on to be powerful. Despite data limitations, new ML approaches are being developed to address the problem of small datasets. Another challenge of machine learning is that they are considered ‘black boxes’ – meaning that we cannot fully understand how the algorithm got from A (input) to B (output). This lack of transparency can often impact the willingness of researchers to adopt these approaches. Nevertheless, explainable AI is a growing field which attempts to build models that can be interpreted and explained.
The researchers in this review stressed the importance of national and international collaborations amongst experts within biomedicine and machine learning in order to resolve these challenges and enable integration of machine learning into neurology practice. With growing health challenges in society due to our ageing population, these multidisciplinary groups will be essential in delivering the benefits of new tools for disease diagnosis and drug development.
Image credit: Freepix