Multimodal Prediction of COVID-19 ICU Admissions and Demand Using Clinical, Governmental, and Social Media Data with GBM and LSTM Models

During pandemics such as COVID-19, it is very important to accurately and quickly predict how bad the disease will get so that resources and patient care can be used most effectively. This study suggests a dual-model machine learning framework that uses data from three sources—Electronic Health Records (EHRs), government-reported case statistics, and social media sentiment trends—to predict ICU admissions at both the patient and population levels. A Gradient Boosting Machine (GBM) classifier was trained on structured clinical features, such as age, comorbidities (such as diabetes and high blood pressure), pneumonia status, and the need for intubation to predict the likelihood of a patient being admitted to the ICU. This model was 93% accurate and had an AUC of 94.5%. SHAP-based feature importance showed that age, hypertension, and pneumonia were the best predictors. Using trends in hospitalization rates, changes in public policy, and social media sentiment over time, a Long Short-Term Memory (LSTM) model was created to predict how many people will need an ICU over time. This model was 95% accurate and had a ±10% error margin for predicting ICU admissions over the next 14 days. All data were aligned in time and combined using region-level tags. To improve model performance, data preprocessing, hyperparameter tuning (using grid search and Bayesian optimization), and comparisons with baseline models, such as ARIMA and linear regression, were performed. This method shows how multimodal, easy-to-understand AI models can be used in healthcare decision support systems for real-time patient triage and hospital capacity planning.