ESTiG - Resumos em Proceedings Não Indexados à WoS/Scopus
URI permanente para esta coleção:
Navegar
Entradas recentes
- Assessing the reliability of AI-based angle detection for shoulder and elbow rehabilitationPublication . Klein, Luan; Chellal, Arezki Abderrahim; Grilo, Vinicius; Gonçalves, José; Pacheco, Maria F.; Fernandes, Florbela P.; Monteiro, Fernando C.; Lima, JoséAngle assessment is crucial in rehabilitation and significantly influences physiotherapistsŠ decisionmaking. Although visual inspection is commonly used, it is known to be approximate. This preliminary study aims to integrate and evaluate AI image-based approaches for assessing upper-limb angles. The study involved 28 participants performing four different rotational joints movement in the shoulder and elbow complex. Two AI algorithms, utilizing MediaPipe Holistic and Yolo v7, were employed for angle estimation. The accuracy of the estimations was evaluated against a wall-mounted compass, considering the ground truth. The results showed that the AI image-based algorithms displayed promising capabilities in assessing the exercises. Yolo v7 achieved the highest quality of estimations, with MAE equal to or less than 5ž, while MediaPipe, despite producing poorer results, where the MAE reaches values of 17ž, offered more features and required lower computational power than Yolo v7. However, it is worth noting that Yolo v7 was limited to exercises in 2D and did not estimate the position of key body points in 3D. Nevertheless, Yolo v7 would provide a cost-effective and easily implementable solution for measuring angles in rehabilitation activities for 1 Degree of Freedom (DOF) exercises. Overall, this study demonstrates the great promise of angle estimation for rehabilitation purposes of the AI approach.
- An evaluation of image preprocessing in skin lesions detectionPublication . Silva, Giuliana; Lazzaretti, André; Monteiro, Fernando C.This study aims to evaluate the impact of image preprocessing techniques on the performance of Convolutional Neural Network (CNNs) in the task of skin lesion classification. The study is made on the ISIC 2017 dataset, a widely used resource in skin cancer diagnosis research. Thirteen popular CNN models were trained using transfer learning. An ensemble strategy was also employed to generate a final diagnosis based on the classifications of different models. The results indicate that image preprocessing can significantly enhance the performance of CNN models in skin lesion classification tasks. Our best model obtained a balanced accuracy of 0.7879.
- A comparative analysis of MATLAB and Python neural networks for diabetes predictionPublication . Pimentel, G.O.; Dessanti, A.L.; Teixeira, João Paulo
- Automatic speech recognition for Portuguese: a comparative studyPublication . Borghi, P.H.; Teixeira, João Paulo; Freitas, D.
- Deep learning and machine learning techniques applied to speaker identification on small datasetsPublication . Manfron, E.; Teixeira, João Paulo; Minetto, R.
- Comparative analysis of windows for speech emotion recognition using CNNPublication . Teixeira, Felipe; Soares, S.; Abreu, J.L.; Oliveira, P.; Teixeira, João Paulo
- ECG and sEMG conditioning and wireless transmission with a biosignal acquisition boardPublication . Luiz, Luiz; Teixeira, João Paulo; Coutinho, F.
- Forecasting COVID-19 in european countries using long short-term memoryPublication . Carvalho, Kathleen; Teixeira, Rita; Reis, Luis Paulo; Teixeira, João PauloEffective time series forecasts are increasingly important in supporting judgment in various decisions. Various prediction models are available to support these projections based on how each area provides a diverse set of data with variable behavior. Artificial neural networks (ANNs) significantly contribute to medical research since using predictive ideas allows for the study of disease progression in the future, as well as the behavior of other variables. This study implemented the proposed model based on Long Short-Term Memory (LSTM) to forecast COVID-19 daily new cases, deaths, and ICU patients. The methodology uses quantitative and qualitative data from six European countries: Austria, France, Germany, Italy, Portugal, and Spain to predict the last 242 days of the COVID-19 pandemic. The dataset uses the healthcare parameters of the number of daily new cases, deaths, ICU patients, and mitigation procedures, such as the percentage of the population fully vaccinated, the mandatory use of masks, and the lockdown. Two approaches were used to evaluate the model’s performance: the mean absolute error (MAE) and the mean square error (MSE). The results demonstrate that the LSTM model efficiently captures general trends in COVID-19 metrics but shows limitations when predicting data with low values or high variability, such as daily deaths. The model reported the lowest errors for Spain and Portugal, while France and Germany exhibited higher error rates due to differences in data reporting and pandemic dynamics. These findings highlight the importance of contextualizing predictive models based on specific regional characteristics.
- Schizophrenia diagnosis support with spectral and cepstral features of speechPublication . Teixeira, Felipe; Mendes, João; Salviano F.P.; Abreu, J.L. Pio; Teixeira, João PauloSchizophrenia is a severe mental illness affecting over 20 million people worldwide, significantly impairing quality of life and daily functioning. Current diagnostic methods rely heavily on subjective assessments and interactions between doctors and patients, leaving room for potential misdiagnoses. Recent advancements in technology have introduced non-invasive, fast, and userfriendly approaches, such as machine learning, to support psychiatric diagnosis. In this study, spectral features extracted from speech samples of individuals with and without Schizophrenia were analyzed. Using an ensemble bagged tree model, we achieved an accuracy of 96.3%, a sensitivity of 94.6%, and an F1-score of 95.4%. These results highlight the potential of speech-based machine learning models as effective tools for aiding Schizophrenia diagnosis.
- Prediction of average power produced by wind turbines using MLP neural networksPublication . Campos, Letícia; Luiz, Luiz; Poubel, Raphael; Teixeira, João PauloThis paper explores wind turbine power output prediction using Multi-Layer Perceptron (MLP) neural networks. Accurate forecasting of wind energy production is critical for grid stability and optimizing energy systems. The study compares various prediction techniques, including physical, statistical, and hybrid methods. The methodology employs real-world data sourced and uses records from 2016-2017. Data preprocessing includes filtering, seasonal decomposition, time series analysis, and dividing the dataset into training, validation, and testing sets. The model’s structure and hyperparameters were carefully tuned, employing 144 samples from the produced power as input, representing 24-hour cycles, to forecast the next hour. The study evaluated multiple MLP configurations, varying in hidden layer sizes and training strategies, to identify the optimal architecture for short-term wind power forecasting. The evaluation uses statistical metrics to assess prediction accuracy, including RMSE, NRMSE, and R2. Early stopping and randomized dataset splits were evaluated to enhance model performance and robustness. The main goal of this paper is to demonstrate the utility of MLP in forecasting for wind power generation systems. The models analysed obtained results between 94-95% for the coefficient of determination (R2). To improve performance, we should add environmental variables to the forecasting models or use deep-learning models.