Human Activity Recognition (HAR) has been studied extensively, yet current approaches are not capable of generalizing across different domains (i.e., subjects, devices, or datasets) with acceptable performance. This lack of generalization hinders the applicability of these models in real-world environments. As deep neural networks are becoming increasingly popular in recent work, there is a need for an explicit comparison between handcrafted and deep representations in Out-of-Distribution (OOD) settings. This paper compares both approaches in multiple domains using homogenized public datasets. First, we compare several metrics to validate three different OOD settings. In our main experiments, we then verify that even though deep learning initially outperforms models with handcrafted features, the situation is reversed as the distance from the training distribution increases. These findings support the hypothesis that handcrafted features may generalize better across specific domains.

Glaucoma is a heterogeneous group of diseases characterised by cupping of the optic nerve head and visual field damage, starting with a progressive loss of vision that leads to permanent blindness. When diagnosed in time, Glaucoma can be delayed by adequate treatment. More efficient processes for diagnosis are being proposed, and the role of artificial intelligence in the field is growing. This work presents a pipeline to evaluate the impact of generative modelling in Computer-Aided Diagnosis (CADx) of Glaucoma based on Deep Learning, particularly focused on the optic disc region. The methodology relies on transforming retinal fundus images to improve and degrade their quality to augment the training data and assess the diagnostic performance. The objective evaluation of the proposed model based on Generative Adversarial Networks revealed quantitative and qualitative improvements in image quality. To support this, we propose a new model to evaluate the quality of fundus images, which can also be used within the pipeline to reject samples with lower image quality for diagnosis. Its performance surpassed related work, achieving a balanced accuracy of 0.929. Concerning Glaucoma CADx, the results obtained in public datasets point to a considerable gain in Sensitivity, Specificity, and Accuracy, achieving scores of 0.883 (+0.054), 0.957 (+0.019), and 0.931 (+0.031), respectively, after image data augmentation when compared with previous work targeted at offline inference in mobile devices. Considering the restriction of choosing simpler backbone networks that can run on edge devices, our findings support the importance of image quality diversity and realistic augmentation.

Searching for characteristic patterns in time series is a topic addressed for decades by the research community. Conventional subsequence matching techniques usually rely on the definition of a target template pattern and a searching method for detecting similar patterns. However, the intrinsic variability of time series introduces changes in patterns, either morphologically and temporally, making such techniques not as accurate as desired. Intending to improve segmentation performances, in this paper, we proposed a Mask-based Neural Network (NN) which is capable of extracting desired patterns of interest from long time series, without using any predefined template. The proposed NN has been validated, alongside a subsequence matching algorithm, in two datasets: clinical (electrocardiogram) and human activity (inertial sensors). Moreover, the reduced dimension of the data in the latter dataset led to the application of transfer learning and data augmentation techniques to reach model convergence. The results have shown the proposed model achieved better segmentation performances than the baseline one, in both domains, reaching average Precision and Recall scores of 99.0% and 97.5% (clinical domain), along with 77.0% and 71.4% (human activity domain), introducing Neural Networks and Transfer Learning as promising alternatives for pattern searching in time series.

The prevalence of type 2 diabetes mellitus is increasing worldwide. Current methods of treating diabetes remain inadequate, and therefore, prevention with screening methods is the most appropriate process to reduce the burden of diabetes and its complications. We propose a new prognostic approach for type 2 diabetes mellitus based on electronic health records without using the current invasive techniques that are related to the disease (e.g. glucose level or glycated hemoglobin (HbA1c)). Our methodology is based on machine learning frameworks with data enrichment using temporal features. As as result our predictive model achieved an area under the receiver operating characteristics curve with a random forest classifier of 84.22 percent when including data information from 2009 to 2011 to predict diabetic patients in 2012, 83.19 percent when including temporal features, and 83.72 percent after applying temporal features and feature selection. We conclude that he pathology prediction is possible and efficient using the patient’s progression information over the years and without using the invasive techniques that are currently used for type 2 diabetes mellitus classification.