Negative selection based anomaly detector for multimodal health data

Early detection of emerging disease outbreaks is crucial to effective containment and response, yet initial outbreak signatures can be difficult to detect with automated methods. Outbreaks may be masked by noisy data, and signs of an outbreak may be hidden across multiple data feeds. Current biosurveillance methods often perform unimodal statistical analyses that are unable to intelligently leverage multiple correlated data of different types while still retaining quantitative sensitivity. In this paper, we propose and implement an anomaly detection system for health data based upon the human immune system. The adaptive immune system operates over a high-dimensional antigen space in a distributed manner, allowing it to efficiently scale without relying on a centralized controller. Our negative selection algorithm based on the immune system provides effective and scalable distributed anomaly detection for biosurveillance. It detects anomalies in the large, complex data from modern health monitoring data feeds with low false positive rates. Our bootstrap aggregation method improves performance on high-dimensional data sets, and we implement a parallelized version of the algorithm to demonstrate the potential to implement it on a scalable distributed architecture. Our negative selection algorithm is able to detect 90% of all outbreaks with a false positive rate of 11.8% in a publicly available multimodal synthetic health record data set. The scalability and performance of the negative selection algorithm demonstrate that immune computation can provide effective approaches for national and global scale biosurveillence.