Title : Meta-data Harmonisation and Data pre-processing for Machine Learning Classifiers and Decision-Support on Regional Cohorts with Infective Endocarditis
Abstract:
Artificial intelligence (AI) and machine learning (ML) are increasingly being used for infective endocarditis (IE) conditions, particularly for their diagnosis and prognosis with predictive modelling. This will lead to medical reporting greater improvements in diagnostic accuracy, risk stratification and mortality prediction tasks.
However, the development of robust and generalised models greatly depends on the availability of quality, standardised and interoperable healthcare data sources. International metadata standards, including those of SNOMED CT, ICD-10, HL7 FHIR as well as the Dictionary of Medicines and Devices (DM+D), have been developed to enable interoperability with secondary data sources across healthcare systems.
Despite the increased adoption of these standards, there is limited evidence describing or quantifying the practical challenges for establishing a benchmark approach of AI-trained and tested models, using disease-specific datasets with standardised semantic structures across fused healthcare data sources across the South of England.
The dataset of interest for harmonisation and fusion included patient demographics, area-level socio-economic deprivation (Index of Multiple Deprivation), antimicrobial regimens, microbiology, haematology and biochemistry, affected valve or cardiac implantable electronic device (CIED), cardiac surgery referrals, valve surgery, multimodality cardiac imaging (including echocardiography and positron emission tomography), follow-up clinics, and death verification.
The common metadata elements with their standardised definitions were successfully established in order to facilitate the mapping to a pre-set regional Infective Endocarditis variables library with feasible contexts and local constraints. This led us to a successfully benchmarked harmonisation process with systematic characterisation data variations in clinical coding, terminology, completeness, definitions and interoperability across the region’s healthcare hospitals. Indeed, our work focused on the development and evaluation of a regional IE research data library of the south of England, which meets the challenges of metadata harmonisation. It is then automatically followed by data pre-processing for training, validating and performance-testing of our deployed supervised machine learning algorithms. These are based on decision-tree classifications which lead to improving decision-support for diverse IE cases for care. Also, our future work aims at incorporating more data sources across the regional healthcare institutions, for larger metadata harmonisation campaigns which will provide us with much more generalised and specialised IE decision-trees classifiers for operations.
