Knowledge-infused Learning in Healthcare
Statistical AI has been successfully applied in healthcare services (e.g. electronic health records) and practices (e.g. counseling conversation in mental health) both in online and clinical settings. However, several concerns remain that prevent the broader adoption of AI techniques. These include interpretability, traceability, and explainability of recognized patterns and their inter-relationships, which improves clinical decision making. While learning the attribute-level characteristics of the data, a data-driven model is affected by inherent sparsity, lack of variation, and ambiguity in the data. This limits the generalizability of the model (e.g. exposure bias, vanishing gradient). Further, in the healthcare domain, most of the decisions are determined through domain knowledge, experience (including precedence), and reasoning. This can be incorporated by adding a semantics-driven approach that incorporates logical constraints together with domain knowledge to identify latent relationships in data. This allows the model to remove ambiguity, augment relevant information to resolve sparsity, and perform stratified learning. Thus, our task is to efficiently integrate top-down and bottom-up learning paradigms to facilitate interpretable and explainable intelligent systems. Infusion of relevant semantic information to input, latent, and output layers of deep nets through weight-customization of knowledge graph (e.g DBpedia, ConceptNet, SNOMED-CT, DataMed) could provide answers to the following four questions in healthcare infrastructure: (a) Do we really need a powerful and complex learning mechanism in developing early intervention system in healthcare? (b) How to understand and visualize the learning from a model at a conceptual level, (c) Did the model give appropriate weights to concepts/relation that interest end-user clinicians?, and (d) Why should I learn from data when I already have structural knowledge that captures it. An inquiry into these questions, allowed us to investigate three different knowledge-infusion paradigms: (a) Shallow Infusion, (b) Semi-Deep Infusion, and (c) Deep Infusion of knowledge from KG. The talk will explain Knowledge-Infused learning through four capabilities needed for healthcare applications: Medical Entity Normalization, Knowledge-Aware Hybrid Patient-Clinician Matchmaking, and Abstractive Summarization.