Development of an Explainable Artificial Intelligence System to identify predictive makersfrom small data sets.
Abstract
The recent emergence of deep learning and deep artificial neural network systems offers thepossibility to build predictive tools. Typically, these computational systems require large datasetsfor training and dramatically lack explainability, i.e. a logical reasoning in the state-of-the-artdomain knowledge that would allow a domain expert to understand the prediction. In this proposal,MyndBlue describes the development of an explainable deep learning system applied to datasets ofrestricted size using data augmentation techniques and the participation of experts in a knowledgegrounding step, based on causal inference theory.This domain-agnostic approach can be used to predict events in any domain withlimited data, including military strategy and conflict prediction. The goal of this project is to build apredictive Artificial Intelligence System that will identify biomarkers that are statistically robustand that can provide a rationale/explanation for the prediction accuracy.The architecture of the proposed system is designed to predict uncommon abnormal eventscorresponding to a perturbation of in common normal situation, anticipating multicyclic varyingevents. We will augment limited data sets by adapting data from more abundant sources. Forexample, we can model the individual pattern of an ill patient by adapting data from a healthypopulation.Among the many possible applications to this predictive artificial intelligence (AI) system,we decided to build our research in the context of Post-Traumatic Stress (PTS). Indeed, the medicalcontext often allows the acquisition of a very limited amount of data through clinical trial, althougha predictive system has to provide a medically significant rational / explanation. Additionally,PTS consists in a reaction to an abnormally intense and/or long stress exposure where theorganization of the healthy basic physiological parameters (such as motor activity, heart ratevariability, respiratory rate, electrodermal activity and circadian synchrony) is affected.To build this predictive AI system, we will first acquire continuously recorded medicalgrade physiological data from a cohort of 200 PTS patients and 2,600 healthy volunteers using awrist-worn device developed by MyndBlue. In parallel, we will construct a causal clinicalknowledge base using causal inference theory with input from clinical experts in order to groundthis knowledge onto the recorded data. The total cohort will allow us to train a generative model, soas to considerably extend the dataset by generating synthetic data, under expert guidance. Theaugmented dataset will be large enough to authorize the use of the full power of deep learningpredictivity, through recurrent neural networks. The clinically labeled data will ensure significantclinical reliability, while the augmented data will provide robustness to the final predictive system.The explainability of the completed system resides in the design of a causal knowledgebase resulting from close collaboration between data science researchers and clinical experts. Theresults of this project could justify the organization of a more precisely targeted clinical trial, aimingat identifying/validating the discovered mechanisms and at collecting more data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 19, 2020
- Source ID
- N629092012076
Entities
People
- Denis Fompeyrine
Organizations
- Office of Naval Research
- United States Navy