(345o) Data-Driven and Knowledge Hybrid Real-Time Decision Making for Exposed Chemical Hazard Based on Knowledge Graph Inference Expanded By Deep Learning

Chemical exposure accidents pose a risk of serious injury and property damage if initial decision makings such as detection, diagnosis, and response are not taken appropriately. In the event of a dangerous goods accident, emergency responders need to make smart decisions with extensive information on hazardous substances (substance identification assistance, physical properties, human health information, containment and containment advice, etc.). However, existing chemical detection technology is focused on the development of response technology for leaks of mostly-used chemicals at pre-analyzed sites, and research on the detection and identification at dynamically changing environment or knowledge service for decision makings on hazardous chemicals considering on-site symptoms as precursor information is insufficient.

In this study, we propose an AI-based analytics support system (SEARCH) for chemical identification from actively collected exposure symptoms, for real-time response and mitigation to hazardous material incidents. In order to interpret the symptoms (79 symptoms including hearing loss, impaired vision, bloody nose, etc.) and real-time sensor data of chemical exposure accidents or contact with substances, information related to chemical substances (symptoms, toxicity, structural information, etc.) is collected and extracted from WISER, PubChem, MSDS, etc. to systematize the domain knowledge base. HAZOP results are analyzed to build knowledge base on site-dependent accident contexts. A total of 1,001 key chemicals, including corresponding symptom knowledge, were collected and extracted. The knowledge collected is systematically stored in the knowledge base in a triple format using AllegroGraph (see the attached figure). Graph embedding is performed through embedding techniques such as TransE, ConvKB, ComplEx, etc. provided by the Python AmpliGraph package. The collected set of symptom knowledge are cross-validated through molecular fingerprint and symptom triple embedding. Confidence on each compound-symptom relation is assigned from the results of cross-validation.

In case of WISER developed by NIH, chemical substances are estimated by simply pattern matching manually-inputted symptom words. Data-driven categorization of the accident site is performed first, and the proposed system actively asks and collects symptoms and predicts candidate chemicals (including probabilities) through knowledge-based reasoning that reflects the confidence of the symptoms validated through embedding. The knowledge base reasoning uses knowledge graphs as knowledge base that is easy to link and visualize, and relies on reasoning methods such as the dynamic RDFS++ reasoner, which is easy to infer dynamic data, and OWL2RL materializer, which is easy to infer static data. A deep learning model for chemical-symptom prediction is also developed to expand the symptom knowledge base for less-known substances that have no prior knowledge on exposure symptoms.

In addition to substance identification analytics, SEARCH provides a wide range of information on hazardous substances, including physicochemical characteristics, human health information, and containment and suppression advices. The prototype system is serviced with AI speakers as the main interface and demonstrated at the Gyeonggido Disaster and Safety Headquarters, in charge of a quarter of Korea. The voice recognition interface using AI speakers enables effectively acquire information by saving time and effort rather than directly inputting and matching information manually at incident sites. SEARCH-enhanced analytics and decision making support offers the means to transform the on-site noisy data into insights and actions related to advanced emergency responses.