Is your data landscape ready for AI?
The ever-changing complex healthcare ecosystem creates enormous amounts of data every day. Data that if used in AI workflows can transform services, give back to clinicians the gift of time and provide maximum benefit to patients. However, are our data landscapes ready to accommodate AI?
In this section, we propose a new model, called the LOAD model, that assists in identifying factors that can cause problems when we introduce new functionality (like Machine Learning (ML) workflows) to a complex ecosystem.
The LOAD model
The LOAD model stands for
The model describes the four dimensions that can affect the adoption of new digitalisation innovations, such as AI tools. The aim of the model is to provide some structure on identifying socio-technical factors when implementing new AI tools to be embedded in complex environments. By investigating the four dimensions of the LOAD model, domain experts are covering factors stemming not only from technical, but also from organisational, and human sources.
In the diagram below, we can see the four dimensions of the LOAD model and how they all focus on the data dimension. This represents having the data as our primary point of the investigation while analysing the other 3 dimensions from the data perspective.
The LOAD model was conceptualised through researching the factors contributing towards the failure to realise the planned benefits of new IT systems in the NHS. In this research, a collection of 18 case studies, written by staff working for the National Health Service (NHS) in the UK, was analysed. The case studies describe factors that contributed to the failure or success of recent digital innovation developments in a variety of settings. The results of the analysis showed that the developments failed due to a mixture of technical and human factors, with the human factors being by far the most dominant. The most common causes of IT failure in these case studies are related to people and their interactions.
This suggests that in order to investigate whether our current data landscape is ready to adopt new ML techniques and tools, we have to investigate both the technical, but also the human and organisational factors that can pose any threat to the new project.
The 4 LOAD dimensions
The Data dimension of the model describes factors introduced by the data needed to be used in the new system or functionality. For example, data in healthcare are often captured on pieces of paper, stored in physical folders and moved around through pigeon holes. Often, a change of medium (i.e. from physical to electronic) is required when data is moved between a producer and a consumer. This is straightforward in the case of electronic data, which can be easily converted into report form, for document generation and printing. But the situation is more complicated when data on paper must be entered into a destination software system. Data entry is a time-consuming process typically done by clerical staff, who may not have a strong understanding of the meaning of the data they are entering. Errors can easily be injected that may significantly reduce the quality of the information. Also, if the data to be moved are in physical form (e.g. letter, cassette, X-ray film, blood sample etc.), then there are transportation costs.
Additionally, the Actors dimension describes factors stemming from both people and systems interacting together with the data to accomplish some value. For example, data entered into a system by secretarial staff can contain errors if the information requires medical knowledge/vocabulary that the staff lack. We refer to this as the risk of a clash of grammars (the meaning of the data being altered by the change of context because of a cultural, experience or other types of reasons) which can result to lower data quality entered into the system. Also, integrating data from data silos sources (sources that haven’t previously been shared up to this point) can introduce high costs; such sources typically have limited external connectivity and are tailored for use by one type of users bringing a risk of data quality problems.
Moving on to the Organisation dimension, sharing data outside the immediate organisational unit can result in a number of administrative costs, such as reaching and complying with data-sharing agreements, as well as complying with wider information governance requirements. Also, a risk of staff reluctance to share ownership of data, may exist on both sides of the movement.
Finally, the Landscape dimension describes the environment in which the other three dimensions are interacting within. For example, data captured by a GP for the purpose of providing patient care are retrieved to be used in research projects to analyse the reasons behind the symptoms the patients are having. Or, cancer data captured by a particular foundation trust that is needed for research purposes by another agency.
© The University of Manchester