The thesis work presents novel methods for the analysis and classification of medical images and, more generally, complex data. First, an unsupervised machine learning method is proposed to order anterior chamber OCT (Optical Coherence Tomography) images according to a patient’s risk of developing angle-closure glaucoma. In a second study, two outlier finding techniques are proposed to improve the results of above mentioned machine learning algorithm, we also show that they are applicable to a wide variety of data, including fraud detection in credit card transactions. In a third study, the topology of the vascular network of the retina, considering it a complex tree-like network is analysed and we show that structural differences reveal the presence of glaucoma and diabetic retinopathy. In a fourth study we use a model of a laser with optical injection that presents extreme events in its intensity time-series to evaluate machine learning methods to forecast such extreme events.