Banca de QUALIFICAÇÃO: LEMYSON OLIVEIRA LEMOS

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : LEMYSON OLIVEIRA LEMOS
DATE: 08/10/2024
TIME: 08:30
LOCAL: Meets: meet.google.com/joc-rqeo-wvt
TITLE:

Prediction of vaccine dose application with N-BEATS: A digital health solution for the management of immunobiologicals in the SUS

KEY WORDS:

Time-series analysis, deep learning, N-Beats.

PAGES: 35
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUMMARY:

With advancements in time series forecasting, the N-BEATS (Neural Basis Expansion Analysis Time Series) algorithm has emerged as a standout for modeling complex data. Introduced by Oreshkin et al. (2020), N-BEATS surpasses traditional models, such as autoregressive and Multi-Layer Perceptron (MLP) models, by handling nonlinear patterns without requiring detailed prior knowledge about the data. Its architecture, based on stacked blocks that iteratively refine predictions, proves particularly useful for critical scenarios such as vaccine dose forecasting.

This study proposes applying N-BEATS to predict the number of vaccines to be administered weekly in vaccination systems, creating a model capable of capturing complex patterns without extensive preprocessing. The work involves collecting historical vaccination data, specifically from the RN+VACINA vaccination system in the State of Rio Grande do Norte, with weekly aggregation to simplify analysis and enhance pattern identification.

N-BEATS is a deep neural network model that, during training, utilizes an architecture composed of expansion blocks that transform input data into forecasts through a series of stacked layers. Each block applies transformations to the data, and the model adjusts its parameters to minimize the difference between predictions and actual values. Training involves feeding the model with time series data, divided into input and forecast windows. The model’s performance is evaluated using metrics like MAPE and R², and hyperparameters are adjusted to improve forecasts. This approach allows N-BEATS to capture the complexity of time series data and provide accurate predictions.

Preliminary results showed that N-BEATS achieved an R² of 0.81 and a MAPE of 16.59%, outperforming established models in the literature, such as XGBoost, which had an R² of 0.73. Both models were trained on the same dataset, but differed in preprocessing: XGBoost involved more extensive refinement by removing nulls, outliers, and normalizing the entire dataset, whereas N-BEATS only removed nulls and aggregated data weekly. Predictions were also made for the municipality of Natal, resulting in metrics of MAPE 21.59% and R² 0.77. This demonstrates that the model can be effectively applied to vaccine forecasting at both state and municipal levels.

Next steps include comparing N-BEATS with other forecasting methods and presenting the results at the Congresso Brasileiro de Automatica (CBA) in October 2024, with the dissertation defense scheduled for January 2024. The study aims to optimize vaccine distribution planning, offering a more proactive and efficient approach to public health resource management.

COMMITTEE MEMBERS:
Presidente - 1153006 - LUIZ AFFONSO HENDERSON GUEDES DE OLIVEIRA
Interno - 2885532 - IVANOVITCH MEDEIROS DANTAS DA SILVA
Interno - 2488270 - RICARDO ALEXSANDRO DE MEDEIROS VALENTIM