Estrategias de Ensemble Learning para el diagnóstico asistido en neurociencias
As part of Digital Healthcare, an ever-growing interest is on the development of innovative Artificial Intelligence (AI) techniques to support specialists' decisions. More specifically, in the Neuroscience field, recent studies have highlighted the great potential of Machine Learning (ML) with Magnetic resonance for the automatic and early diagnosis of neurodegenerative diseases such as Alzheimer's. Within this context, the use of ensemble learning methods is proving to be a very promising strategy since it allows for the information content from different types of variables to be exploited and at the same time to optimise the computational effort necessary for automatic learning.
Methods at the Frontier of Artificial Intelligence at the Service of Assisted Diagnosis
As already illustrated in a recent article entitled Data Science for the study of Alzheimer's: hidden patterns in the connectome, Exprivia offers innovative methods at the frontier of research in Data Science and AI to identify biomarkers and to diagnose neurodegenerative diseases with the goal of obtaining reliable predictive systems that ensure high performance. Of these, l’Ensemble Learning is an automatic learning approach that tends to improve classification by combining different algorithms, whose accuracy is “mediated” based on the performance of each single classifier. This approach is conceived to simulate human behaviour in decision making and, for this reason, is seen to be particularly suited to medical diagnosis, in which people ask for the "second opinion" of doctors to enhance the reliability of a diagnosis.
Nel In the downloadable Paper, published by the international scientific magazine MDPI Electronics, Exprivia has proposed an innovative framework based on Ensemble Learning techniques, to tell the difference between healthy check-ups and people with Alzheimer's, applied to different groups of physical measurements associated with certain tracts of the brain, known as “white matter fibres”.
One of the most popular methods used to analyse the state of white matter fibres, and therefore neurological conditions, is the measurement of the diffusivity of water molecules along the fibres using diffusion-weighted magnetic resonance imaging (DWI). In fact, in healthy people, the movement of the molecules of water within one region of the brain should follow certain directions (anisotropy), and so the loss of this directionality is a clear sign of the presence of a pathology. To measure the anisotropy, different groups of indicators are used (e.g. fractional anisotropy, mean diffusivity, radial diffusivity, longitudinal diffusivity), extracted using processed images of the DWI.
Since in a typical experiment each of these indicators, for the entire cerebral area, can exceed hundreds of thousands of variables, training a classifier based on the total anisotropy measurements could be computationally challenging on the one hand, and not very efficient in terms of predictive performance on the other. In this particular case, the Ensemble Learning strategy meant this problem could be overcome, by basing the system not on an algorithm trained in all of the variables simultaneously but on the best combination of algorithms trained in specific groups of measurements, so that the final decision emerges as the opinion most shared between those "asked" of each single classifier. The results published show how this technique has allowed for better prediction performance to be achieved compared to traditional training approaches that exploit the concatenation of the variables in one single high-dimensionality matrix.
A Promising Approach to Multimodal Data Processing in Neuroscience
Exprivia's constant commitment to adopting innovative AI methodologies increasingly favours the introduction on the market of digital healthcare solutions for automatic diagnosis, with an approach that focuses on the use of state-of-the-art technologies. More specifically, the framework proposed in the article promotes the training procedure for an ML model, potentially incorporating an increasingly diversified plethora of clinical and biological information generated by different methods of diagnosis, thus providing a holistic vision of the disease.