A MULTI-MODAL DEEP LEARNING FRAMEWORK FOR EARLY DETECTION OF INFECTIOUS DISEASE OUTBREAKS USING CLINICAL, ENVIRONMENTAL, AND MOBILITY DATA

Abstract

The growing cases of outbreaks of infectious diseases warrant a paradigm shift in how the heterogeneous data streams in real time are analysed to proactive early warning systems that have the potential to analyse the heterogeneous data streams in real time. The article introduces a novel multi-modal deep learning system, which integrates clinical electronic health records, environmental indicators such as temperature, humidity and air quality, and human mobility patterns in order to predict the outbreaks before they are officially announced. The model was tested on a large dataset to compare it against nine baseline models such as Transformer, LSTM-Attention, and XGBoost. The proposed multi-modal deep learning model had superior results on all metrics with an accuracy of over ninety-eight percent, an AUC-ROC of over ninety-nine percent and a Matthews correlation coefficient of more than ninety-five percent. Most importantly, the system has been demonstrated to provide a lead time of about a hundred and twelve hours or nearly five days of early warning before World Health Organization declaration, with a false alarm rate only about a fifth of an alert per month, more than sevenfold better than the next best operating system. Robustness Analysis The robustness analysis revealed that the model can tolerate moderate level of Gaussian noise, and the model is data-efficient to a large extent. The confusion matrix showed that there was equal per-pathogen accuracy in more than ninety-four percent in the case of novel or unknown pathogens and almost ninety-eight percent in the case of established pathogens such as COVID-19. The computational needs were low, with inference latency of less than fifty milliseconds and memory footprint of less than three gigabytes, making it feasible to run it on the current public health infrastructure. These findings can conclusively demonstrate that the multi-modal deep learning and cross-modal attention can fundamentally change the epidemic disease surveillance system to become proactive and data-driven in preventing outbreaks, which has a significant implication on global preparedness in terms of public health.