Using Knowledge Graph Learning to Predict and Explain Patient Outcomes in Electronic Health Records

The aim of this project is to develop a system that can automatically predict and explain patient outcomes. The purpose of the research is to improve patient care by analysing anonymised electronic medical records at very large scale. For example, the methods developed could predict that a new drug will have a rare but serious side effect, or that there is a potentially preventable cause of a negative treatment outcome of a specific group of patients. This is possible because we can represent information as a network. Networks are a general way to represent the connections between things, such as friendships between people, links between websites, or molecular reactions in a cell. Networks contain "nodes" (the things) and "edges" (connections between the things). In the friendship network, people would be the nodes and there would be an edge between all the pairs of people who are friends. Graph Theory is a set of mathematical principles we can use to analyse any type of network to try to understand how the structure of the connections relates to the overall function.

In this fellowship, a large network will be created that combines publicly available data on medications, diseases and cell biology with anonymised data extracted from electronic medical records. One of the most powerful aspects of this network approach is that it allows these different types of information to be directly connected, and represents exactly how they relate to each other. This allows a computer to reason about patient outcomes with the extra context of existing medical knowledge. Algorithms can analyse this network to make predictions based on the known connections between things (for example, paracetamol is known to work as a painkiller, other drugs similar to paracetamol might also be effective painkillers). Whilst the meaning of these relationships is often intuitive to a person, it is challenging to develop algorithms that can apply this type of reasoning. The purpose of this fellowship is to develop such methods and apply them to make clinically useful predictions.

The first part of the work is to combine the publicly available data and create the large network of known facts that could be useful to explain patient outcomes. This network will then be used to develop and optimise the algorithms that will make the predictions, by training them to predict known associations such as drug side effects or disease risk factors. With the network and the algorithms ready, the work then proceeds in two directions. Firstly, we can look at the graph and predict "missing information", meaning that given everything we know about the drugs that can cause a serious side effect (e.g. Stevens-Johnson syndrome), it's very likely that drugs A, B and C also could cause it. These predictions are then validated by analysing anonymised electronic medical records. The second side to the project is to explain outcomes that are observed in medical records. The first step there is to identify a trend, such as identifying a population of patients who respond poorly to treatment or have an unusually high rate of a negative outcome. We can use the graph to predict why this pattern exists, given all of the medical information available to the predictive algorithm. These patterns, along with their predicted explanations, will be subject to medical review and used to inform policy and best practice decisions to improve patient care.

This fellowship will develop predictive analytics algorithms for knowledge graphs, and apply those methods to a biomedical graph to predict patient outcomes in electronic health records.

The first steps are to develop the knowledge graph and optimise the predictive algorithms, building on my recent work [Bean et al. 2017]. The expanded graph will additionally include proteins, genes, ontologies, diseases and guidelines. Two predictive algorithms will be developed and optimised. A) Improving the performance of my published KG edge prediction method, focused on efficiency (gradient descent) and accuracy through to the use of both additional features (public knowledge and network features) and improved weighting of those features (confidence, provenance). B) New methods to predict the mechanistic basis for an edge in the graph. For example, the basis for a drug indication would involve the protein target of the drug, the pathways those proteins are involved in, and how those pathways relate to the indicated condition. Performance will be evaluated using the known mechanism of action for drugs.

Both of these predictive algorithms will be validated using anonymised EHR data. Firstly, new predictions for outcomes will be generated from the graph and validated using anonymised EHR data extracted using CogStack (data surfacing and harmonisation) and the SemEHR (semantic annotation and search) projects at KCH.

Secondly, one or more cohorts of patients with differential outcomes will be identified - for example treatment response or an unexpected drug-drug interaction. For these patients, classifiers will be trained to predict outcome from routine EHR data. The predictive features are essentially a set of clinical variables that are risk factors for the outcome. The explanation prediction algorithm will use these relationships to propose a mechanistic basis from the knowledge graph. Proposed mechanisms will be manually reviewed for plausibility.