Technologies

The IPQ Knowledge graph offers unique capabilities over and above current conventional approaches to creating graph databases.

Hybrid (Human/Machine) Knowledge Model

In most current practice knowledge graphs (KG) are developed through the integration of existing ontologies as well as the extraction -- using natural language processing (NLP) -- of information from various corpora of literature, e.g. research publications, clinical notes, patient records and others. Limiting the content of the graph in this way presents a significant challenge. Any gaps in knowledge, whether in specific data or at a concept level, as well as any biases and/or data conflicts will be incorporated into the resulting knowledge graph. IPQ’s approach is a hybrid that uses both human knowledge and experience as well as data/information acquired using advanced technologies. We expand the concepts and relationships contained in the KG based on deconvolution of critical knowledge from Subject Matter Experts (SMEs) that allows us to capture both an upper level ontology that specifies how our customers work and the processes that are meaningful for them to perform their work that are not explicit in the literature. This enables the IPQ KG to identify gaps in our current knowledge and conflicts within the information much the way human researchers do, enhancing the quality of the graph which in turn significantly improves results of any analytics applied against the graph.

Temporal Model

The IPQ KG also addresses critical temporal issues. For example, clinicians and researchers are confronted daily with the reality that disease is a process that evolves over time and is not a state. Understanding this is critical to the quality and value of the Knowledge Graph. The data that is collected, from many sources under a multitude of circumstances, tells its own story about the course of a disease, about the entire patient journey, and about the changing environment in which data is collected. These temporal considerations are therefore vital for a clear, unbiased understanding of the science and are key to producing accurate and impactful results. This level of understanding is also applicable to any complex biological condition, and necessitates the development of a supporting structure in the knowledge graph that accounts for that condition in the graph’s ability to link objects and analyze them in an appropriate temporal manner. This capability will enhance clinical decision support for diagnosis and treatment, patient outcome, more directed development of drugs and diagnostics, and it will reduce unnecessary testing and treatment.

Causal Reasoning

Current AI/ML analytics and technology development focus primarily on solutions that apply a correlative analytics approach. Correlative analytics, however, does not address the deeper understanding of root cause. Reliance on correlative analytics can lead to incorrect or misleading results and are not always reliable because the “reasoning” used by these algorithms is not visible and not easily validated. For example, Machine Learning approaches are designed to optimize the association, i.e. combination, of input features to enable the accurate prediction and appropriate validation of output characteristics. In clinical diagnosis this typically would use combinations of patient descriptors, such as clinical observations, to establish disease state or potential response to a specific therapeutic intervention. Such approaches may introduce bias through the selection of descriptors and the individual weights assigned to these descriptors. Clinical descriptors are commonly correlated with specific diagnoses or observed responses to specific drugs and are termed “phenotype”. However, correlations are not always correct, for the reasons stated above, and there is a need for causal analytics that can leverage human-like deduction over a broad set of facts. Capturing these facts and representing them in a way that allows for their use in causal analytics is not being addressed in current knowledge graph technologies; however, it is addressed in the IPQ KG. The ability to perform causal analytics over a comprehensive knowledge base, will provide clinicians and researchers with an in silico partner to complement their own knowledge in pursuit of improved patient outcomes and effective drugs.

IPQ’s goal is to stratify a patient’s clinical journey from health through disease because, fundamentally…

...DISEASE IS A PROCESS, NOT A STATE

With an understanding of the breadth and complexity of data available, IPQ works with clients to identify and quantify their most significant challenges, bridging across the healthcare ecosystem. This is the core conduit which allows clients to access the knowledge needed to execute algorithms on all appropriate data. The application goes beyond current standard clinical and financial decision support systems by providing knowledge that is more accurate for a particular patient. IPQ’s platform produces knowledge that allows for a thorough understanding of how disease progresses at an individual level over time.

Accurate Phenotyping

Phenotypes are commonly defined as the result of interaction between one’s genomics and environment, with phenotypic features focused on an individuals’ observable traits, e.g. hair color, height, blood type, level (or absence/presence) of specific biomarkers. As currently used, the elements that comprise a phenotype reflect an attempt to identify similar characteristics of a group of patients who may exhibit the same diagnosis, response to treatment or outcome for use in diagnosis and prediction of risk. Associations made in this manner tend not to focus on the central theme, that disease is a process that evolves over time. Such an approach may limit the potential to develop further disease stratification, impacting diagnosis, treatment decisions and outcomes. The use of these phenotypes to infer endotypes may further negatively impact target selection in drug development.

IPQ extends the definition of phenotype toward a next generation of phenotype that is based on observations of disease progression in the patient. This will build on methods that have shown success in disease stratification of patients to identify distinguishable pathways of progression and to establish measurable parameters, both clinical and non-clinical, that can be used to assign patients early in the course of their disease.

Industrial collaboration

IPQ partners with the Healthcare and Life Sciences communities to use Next Generation Phenotyping in clinical trial design and drug discovery to enable endotype target selection which will further demonstrate the value of this approach.

The IPQ Analytics platform is built using a novel approach toward the construction of patient-centric and provider-centric models that form the basis for deep, accurate analytics. Unique algorithms developed for quantum computing enable the integration and analysis of very large, heterogeneous data sources which allows IPQ to expose real-world complexities using novel semantic models, commercial-grade knowledge graph tools and methods, NLP, and other technologies.

Researchers have made impressive progress in the past decade employing machine learning, especially deep learning techniques, on large publicly available datasets in a wide variety of tasks and scenarios, such as document classification, machine translation, and natural language understanding. However, the latest advances have highlighted weaknesses of state-of-the-art data-driven deep learning models which cannot be overlooked as they severely limit their applicability in real-world contexts, in particular those within a medical context. Most successful ML applications have involved supervised learning where the potential for bias in defining the outcome may limit the ability to enable discovery of novel critical factors and relationships. Graph analytics have provided additional insights using data-driven analysis. This also points to the common problem of missing (or conflicting) data where Bayesian and simulation methods have become important in bridging these gaps. A lack of grounding to the real world, and absence of interpretability beg the question of whether or not the combination of these approaches can produce analysis and prediction at a level necessary to evolve medical decision making. AI seeks to reconcile the stochastic nature of learning in neural networks with reasoning and explainability via semantic representations. Explainability, however, is not causality.

IPQ moves beyond current limitations by identifying them and addressing them through the combination of a hybrid modeling approach, novel enhancements to current knowledge graph capabilities, and deep data harmonization.