Технологія LLM-видобування ознак тестування пацієнтів з текстових звітів для удосконалення прогнозування кількості хворих на коронавірус

Abstract

The article focuses on the application of modern large language models (LLMs) to automate the extraction of essential features from analytical textual reports on the COVID-19 pandemic in Ukraine during 2020–2022. These reports encompass a broad spectrum of data, including regional morbidity indicators, testing dynamics, vaccination outcomes, and demographic characteristics of patients. The study explores the integration of these extracted features into time series models to improve the accuracy of epidemic forecasts. Central to the research is the use of the Prophet model, which was enhanced to account for seasonal changes and anomalies in the data. The study addressed challenges such as the multi-wave nature of the COVID-19 time series, incorporating sharp increases and decreases in cases. Adjustments were made for anomalies caused by changes in quarantine measures, testing policies, and vaccination campaigns, particularly during winter surges. Optimizing the Prophet model involved advanced parameter tuning using methods such as grid search and stochastic optimization, tailored to the specific epidemiological context in Ukraine. Additionally, the study evaluated the potential of neural network models, including LSTM (Long Short-Term Memory), to analyze time series data. LSTM’s ability to capture nonlinear relationships and process multiple input variables complements traditional methods, providing deeper insights into long-term trends and interdependencies in the data. The goal of this study is to develop an effective forecasting tool that integrates LLM-extracted features with advanced modeling techniques. By combining Prophet with enhancements and neural network approaches like LSTM, the research aims to significantly improve the accuracy of short- and long-term forecasts. This is particularly crucial for timely decisionmaking in public health during periods of epidemiological uncertainty