Andrew Johnson is CTO for AIMed with responsibility for database management, web development along with client management. A highly experienced publishing executive with a passion for technology.
Some believe if healthcare providers could accurately predict when and why patients are likely to access their services, they will be able to save a substantial amount through better allocation of resources. A group of researchers from the Finnish Centre for Artificial Intelligence (AI), Aalto University, the University of Helsinki, and the Finnish Institute for Health and Welfare tested the concept recently, with the help of machine learning (ML).
A new risk adjustment model
Electronic Health Records (EHRs) of 1.4 million Finnish citizens aged 65 and above were used to develop a new “risk adjustment model” powered by deep neural network to forecast the frequency of elderly seeking treatments at local hospitals or healthcare facilities in the following year at an individual level.
Traditionally, risk adjustment models are created based on simple regression using patients’ demographics, socio-economic status, and counts of previous years’ diagnoses. While the method is being deployed in many countries like the US, Germany, and the Netherland, its accuracy is often subject to question. Using a recurrent neural network (RNN) regression, researchers now rely on data coming from outpatient visits, diagnoses, and medical procedures from previous years.
This will not only remove the need to use data that contain personal identifiers or sensitive features (i.e., income and race), but also account for the possibilities that heavy users of healthcare services may be compensated by separate budgets and the time difference between the creation of the model and its actual use.
The potential of ML
Researchers said the study showed the potential of deep learning in deriving at a more precise risk adjustment prediction by handling more variables and codes. Furthermore, the research team had successfully integrated features of deep learning into the existing regression models, combining the best sides of both, to better support clinical decision making and allocations of funding in the coming year. At the same time, patients who become ill more often can stand a chance to offer more attention and care.