Semantic-based secure infrastructure for interoperable, in-depth biomedical data and analytics

The use of electronic health care records (EHRs) is paramount in today's modern health care systems. The reasons are manifold and range from enabling physicians to have remote access to data, alerting consultants of critical laboratory values, or indicating harmful multi-drug interactions. These records are also employed to some extend to analyse the underlying diseases. However, EHRs are based upon historical models of clinical care containing mostly unstructured and poorly defined information related to the patients' experience and clinicians' interpretation of their illnesses, diagnoses as well as reports from clinical imaging. Moreover, the rapid advent of modern diagnostic medical imaging methods ranging from MRI, PET, and CT scans to ultrasonography and ECGs, have resulted in rapidly increasing unstructured and unlabelled data in the form of 2D and 3D images. Finally, whilst most clinical settings are progressing to complete paperless environments thereby constantly increasing the amount of clinical information that is computationally accessible, most of the settings are employing different and diverse electronic health care records, rendering the integration and interoperability of health care data extremely difficult and laborious.

What is required is a unifying interoperability and analysis framework that would act as an integrated data environment allowing EHRs to be used for computer-based exploration using harmonized methods, i.e. using common interfaces with same standards and semantics. This would ensure that electronic health data can be accessed and interpreted across multiple disciplines in a consistent and readily usable fashion, allowing to easily explore novel avenues in analytics. Such an approach would allow the separation between health care data and analytical tasks. The advantages would be that predictive models developed by one health care provider could easily be applied be validated by others. Moreover, and perhaps more importantly, such an approach will cater for the development of accurate in-depth phenotypic characterisation of diseases at an individual patient level, which would facilitate the improvement of diagnosis, risk prediction, and patient management.

Furthermore, what is required to unlock the full potential of EHRs, is the possibility to extract information from unstructured data stored in form of free texts or as images. Modern AI approaches for such data allow the semantic extraction of information to a very high degree and are to some extend capable of harvesting information on human expert level. The automatic extraction of information from such unstructured data would contribute greatly to the proposed harmonized environment and would not only allow the rapid exploration of, i.e. the search for, patient records expressing specific symptoms, but would furthermore allow future, more symbolic AI approaches for the analysis. This would allow to close the loop from data collection over predictive modelling back to clinician able to constantly validating predictive models in their daily work.

One aspect of using modern AI approaches is the large number of examples required for inducing models for text and images. Simply sharing EHRs is often not feasible due to legal and ethical constraints. What is needed is a novel approach of sharing information without sharing the actual data. One avenue might be using anonymization-based techniques, sampling examples from learned distributions of the data. Such data can subsequently be used with less constraints by others, strengthening the models and enabling the sharing of medical data on a broader, yet secure, scale.

The main goal of this fellowship is the development of such a semantic-based "information commons for research" framework that will form a fully integrated health research information platform, with access to both deep clinical data and extensive large and unstructured data on individual patients and populations.

WP1: Harmonized data framework

WP1 will research the development of an integrated health-data framework based on syntactic and semantic interoperability. We will investigate how current data held in various resources (EHR, LIMS) can be represented for different content models based on existing ontologies and standards. We will then employ the semantically characterised data to develop and apply symbolic AI techniques that will facilitate the creation and manipulation of comprehensible conceptualisations of information, allowing us not only to make decisions, but to truly learn from data.

WP2: Computer-usable information extraction from text

Current electronic healthcare records commonly contain poorly defined information in textual form. Based on our proposed framework, we will develop methodologies using advances AI techniques (Word2Vec, LSTMs) and apply them to characterise textual data and transmute them into structured formats. Initially we intend to apply our approach to a few chosen domains and then scale it up.

WP3: Analysis of complex multi-dimensional images

Similarly, multi-dimensional data cannot easily be used for computer-based analytical systems. To harvest the potential information hidden in unstructured data, we will develop approaches to employ modern AI techniques (CNNs, GANs) for knowledge extraction. We plan to further combine this feature learning with defined vocabularies to directly extract meaningful annotations.

WP4: Anonymization approaches for distributed analysis

AI techniques rely on large datasets with millions of examples. However, this creates challenges with respect to sharing data whilst respecting the wishes of patients and meeting the GDPR requirements. To address this, we will develop anonymization approaches able to learn distributions from data. These distributions can be employed to sample novel, artificial data, which can be shared without disclosing any real data and be used to boost model induction.