Evolving complex yet interpretable representations: application to Alzheimer’s diagnosis and prognosis Proceedings

With increasing accuracy and availability of moredata, the potential of using machine learning (ML) methods inmedical and clinical applications has gained considerableinterest. However, the main hurdle in translational use of MLmethods is the lack of explainability, especially when non-linearmethods are used. Explainable (i.e. human-interpretable)methods can provide insights into disease mechanisms but canequally importantly promote clinician-patient trust, in turnhelping wider social acceptance of ML methods. Here, weempirically test a method to engineer complex, yet interpretable,representations of base features via evolution of context-freegrammar (CFG). We show that together with a simple MLalgorithm evolved features provide higher accuracy on severalbenchmark datasets and then apply it to a real word problem ofdiagnosing Alzheimer’s disease (AD) based on magneticresonance imaging (MRI) data. We further demonstrate highperformance on a hold-out dataset for the prognosis of AD.Keywords — grammar evolution, feature representation,interpretability, Alzheimer’s disease, machine learning