International Journal of Advanced Research in Computer and Communication Engineering

Abstract: This study investigates machine learning approaches for predicting COVID-19 hospitalization rates in San Francisco, utilizing public datasets from DataSF encompassing testing metrics, deaths, and demographics from March 2020 to April 2024. The primary objective is to accurately predict daily patient counts in both Intensive Care Units and Medical-Surgical units through two distinct modeling tasks: point regression and long-horizon forecasting. For the point regression task, features were engineered from aggregated daily statistics, including lagged death counts and race-disaggregated testing data. A comparative analysis of five regressors models- K-Nearest Neighbors (KNN), Decision Tree, Linear Support Vector Machine, Non-linear Support Vector Machine, and Multi-layered Perceptron- was conducted using k-fold cross validation.

Preliminary results indicate that the K-Nearest Neighbors regressor significantly outperformed other models, achieving high R² scores of 0.97 for ICU and 0.98 for Med/Surg patient predictions, demonstrating its effectiveness in capturing complex, non-linear relationships within the temporal data. For multi-horizon forecasting, Long Short-Term Memory and Gated Recurrent Unit models were trained on 120 days of data to predict 120 days in the future. Though with some deviation from the true noise of the output, these models successfully capture broader trends, indicating that COVID-19 hospitalization rates are predictable, to a degree. Overall, this research demonstrates the high efficacy of KNN for point-in-time predictions and establishes a promising baseline for deep learning-based long-term forecasting of COVID-19.

Downloads: | DOI: 10.17148/IJARCCE.2025.1412125