Tackle HealthTech AI data challenges early on

Artificial intelligence (AI) is forecast to save hundreds of thousands of lives and billions of dollars by detecting diseases earlier and identifying at-risk individuals. To bring AI solutions to market, however, HealthTech companies must overcome data acquisition, interoperability, data cleansing and privacy challenges.

AI is only as good as the data used for its modelling. In medical applications, AI must go beyond training models to aggregate data from external clinical and non-clinical sources. That could include data from multiple electronic medical record (EMR) systems, sequencing labs, patient questionnaires and other sources.

This data must also be made interoperable through conformity with a standard such as HL7^® FHIR^®, so senders and receivers understand the information in the same way. The data must also be protected in line with security and privacy laws and regulations.

While an AI application can sift through masses of research or clinical data, it remains constrained by its model, which typically will not provide the full context. This means that AI can only assess a patient for risk of heart attack, for example, based on entered patient data.

AI and machine learning need access to siloed data

Consider diagnostic imaging, for example. A study in the Journal of Digital Imaging suggested that nearly 60% of radiology orders did not mention important chronic conditions despite their increasing prevalence. The data would be more helpful if diagnostic imaging workflows included contextual medical record information.

Imaging AI and machine learning applications can ease radiologists’ workloads and cognitive burdens by analysing curated clinical data. However, first, they need access to siloed imaging data and proprietary picture archiving and communication systems (PACS).

This requires knowledge of healthcare data and interoperability with systems from Epic, GE Healthcare, 3M Health Care, INFINITT, Guerbet, Ricoh, Canon Medical, Roche Diagnostics and others. Compliance with healthcare protocols and standards is also necessary to facilitate information flows across different sources.

Once achieved, the right information can be put in front of radiologists at the right moment, eliminating the need for complex EMR searches to retrieve specific patient data. Clinicians have an accurate presentation of current diagnoses, doctor’s notes, wearables data and even genomic information before reading a study.

AI data preparation challenges are often underestimated

Multi-source datasets depend on interoperability and compliance with multiple health data standards. An AI application may be able to exploit the full potential of HL7 FHIR but may also need to work with legacy standards such as HL7 V2 and non-standard or even non-clinical data sources. Relying on a single standard is unlikely to ensure widespread adoption of a new application or device because, in the real world, older, long-standing systems are always in use.

The requirement for comprehensive, aggregated data means that HealthTech companies face real challenges in training on very disparate types of data for whatever purpose. Data preparation challenges for AI applications, in particular, are often underestimated.

Data pulled in from various sources, including devices or systems with legacy messaging standards, previously had to be cleaned and checked for errors by data scientists because it is rarely in structured tables. This preprocessing and labelling transforms data into a suitable format for AI applications.

Robust machine learning algorithms commonly used in AI, such as neural networks, can automate much of the work. They can take care of some of the necessary preprocessing and cleaning through interpreting patterns in the training data. This is especially helpful when the data includes natural language text or other data types that are challenging to deal with.

Consider a single data platform approach early in development

However, the full range of preprocessing requirements can only be fully resolved through a single platform approach that normalises all data sources and provides the connective tissue with other devices and systems.

It is an approach designed for the challenges of the HealthTech industry, providing a patient-centric model that is ready for analysis. A data platform can include numerous trusted pipelines that aggregate data from across sources’ formats, complemented with AI-based techniques. It will take care of labelling, which is critical for training supervised machine learning models. A platform should also track data lineage, allowing developers to use subsets to train predictive models and keeping the link back to the entire dataset to ensure context is retained.

Combined with encryption at rest and in flight, a single platform approach will help HealthTech companies succeed when incorporating AI into their solution. New applications benefit from easier and more streamlined deployments, particularly since commercial success often depends on the ability to develop the solution rapidly to achieve genuine scalability.

Data preparation challenges for AI applications are often underestimated and addressed too late. A single platform approach is the best way to overcome these challenges and should be considered as early in the development process as possible.

About the author

Dr Nick Tayler is a Clinical Safety Specialist and the global AI Safety Lead at InterSystems, a creative data technology provider which delivers a unified foundation for next-generation applications for healthcare, finance, manufacturing and supply chain customers in more than 80 countries. Based in Melbourne, Australia, Dr Tayler provides clinical expertise to assist systems architects in the development of integrated, user-focused implementations of the InterSystems IntelliCare™ unified electronic medical record (EMR) system across the Asia Pacific region.

Top image credit: iStock.com/Jacob Wackerhausen