Early Prediction of Long COVID using Explainable Multimodal Machine Learning

The long-term effects of COVID-19 popularly called Long COVID have been plaguing a sizable global population, with symptoms persisting at times for weeks after initial recovery. Although a number of predictive models have been proposed using structured clinical data, very little has been done about using data from multiple domains such as imaging, behavioral patterns, and wearable sensors-to augment early detection and prediction. This work introduces an explainable multimodal machine learning approach that takes clinical variables, chest imaging, and behavioral signals (fatigue trends and activity level) into the risk prediction of Long COVID development. A hybrid architecture is proposed combining XGBoost for tabular clinical features with Convolution Neural Networks (CNNs) for visual data, thereby bridging gaps in current research wherein existing approaches lack ensemble methods or exclude patient- reported/wearable data. The model is then trained and tested with publicly available clinical datasets with simulated behavioral inputs. Initial findings show that with multiple data sources combined, there is better performance in contrast to dealing with a single source.