Search

As long as the COVID-19 pandemic is still active in most countries worldwide, rapid diagnostic continues to be crucial to mitigate the impact of seasonal infection waves. Commercialized rapid antigen self-tests proved they cannot handle the most demanding periods, lacking availability and leading to cost rises. Thus, developing a non-invasive, costless, and more decentralized technology capable of giving people feedback about the COVID-19 infection probability would fill these gaps. This paper explores a sound-based analysis of vocal and respiratory audio data to achieve that objective. This work presents a modular data-centric Machine Learning pipeline for COVID-19 identification from voice and respiratory audio samples. Signals are processed to extract and classify relevant segments that contain informative events, such as coughing or breathing. Temporal, amplitude, spectral, cepstral, and phonetic features are extracted from audio along with available metadata for COVID-19 identification. Audio augmentation and data balancing techniques are used to mitigate class disproportionality. The open-access Coswara and COVID-19 Sounds datasets were used to test the performance of the proposed architecture. Obtained sensitivity scores ranged from 60.00% to 80.00% in Coswara and from 51.43% to 77.14% in COVID-19 Sounds. Although previous works report higher accuracy on COVID-19 detection, this research focused on a data-centric approach by validating the quality of the samples, segmenting the speech events, and exploring interpretable features with physiological meaning. As the pandemic evolves, its lessons must endure, and pipelines such as the proposed one will help prepare new stages where quick and easy disease identification is essential.

Musculoskeletal disorders (MSD) are a highly prevalent work-related health problem. Biomechanical exposure to hazardous postures during work is a risk factor for the development of MSD. This study focused on developing an inertial sensor-based approach to evaluate posture in industrial contexts, particularly in automotive assembly lines. The analysis was divided into two stages: 1) a comparative study of joint angles calculated during movements of the upper body segments using the proposed motion tracking framework and the ones provided by a state-of-the-art inertial motion capture system and 2) a work-related posture risk evaluation of operators working in an automative assembly line. For the comparative study, we selected data collected in laboratory (N = 8 participants) and assembly line settings (N = 9 participants), while for the work-related posture risk evaluation, we only considered data acquired within the automotive assembly line. The results revealed that the proposed framework could be applied to track industrial tasks movements performed on the sagittal plane, and the posture evaluation uncovered posture risk differences among different operators that are not considered in traditional posture risk assessment instruments.

The study of Domain Generalization (DG) has gained considerable momentum in the Machine Learning (ML) field. Human Activity Recognition (HAR) inherently encompasses diverse domains (e.g., users, devices, or datasets), rendering it an ideal testbed for exploring Domain Generalization. Building upon recent work, this paper investigates the application of regularization methods to bridge the generalization gap between traditional models based on handcrafted features and deep neural networks. We apply various regularizers, including sparse training, Mixup, Distributionally Robust Optimization (DRO), and Sharpness-Aware Minimization (SAM), to deep learning models and assess their performance in Out-of-Distribution (OOD) settings across multiple domains using homogenized public datasets. Our results show that Mixup and SAM are the best-performing regularizers. However, they are unable to match the performance of models based on handcrafted features. This suggests that while regularization techniques can improve OOD robustness to some extent, handcrafted features remain superior for domain generalization in HAR tasks.

Drop-on-demand printing using colloidal or pigmented inks is prone to the clogging of printing nozzles, which can lead to positional deviations and inconsistently printed patterns (e.g., data matrix codes, DMCs). However, if such deviations are detected early, they can be useful for determining the state of the print head and planning maintenance operations prior to reaching a printing state where the printed DMCs are unreadable. To realize this predictive maintenance approach, it is necessary to accurately quantify the positional deviation of individually printed dots from the actual target position. Here, we present a comparison of different methods based on affinity transformations and clustering algorithms for calculating the target position from the printed positions and, subsequently, the deviation of both for complete DMCs. Hence, our method focuses on the evaluation of the print quality, not on the decoding of DMCs. We compare our results to a state-of-the-art decoding algorithm, adopted to return the target grid positions, and find that we can determine the occurring deviations with significantly higher accuracy, especially when the printed DMCs are of low quality. The results enable the development of decision systems for predictive maintenance and subsequently the optimization of printing systems.

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.