eHealth 3.0: Personalized digital twins to capture and use different kinds of clinical knowledge

Abstract

The first generation of eHealth is already a fact: usage of telecommunication to e.g. diagnose patients remotely is now an integrated part of healthcare. We are therefore now in the middle of the second generation of eHealth: artificial intelligence (AI) and machine learning (ML). There are important showcases illustrating that such ML can perform diagnosis and image analysis at the level of a trained radiologist, for specific applications REF. However, ML has critical and inherent shortcomings, which severely limit the impact that ML in itself can have on healthcare. For instance, ML models need the right type of data from thousands of patients to train the models, they do not make use of or add to the physiological or biochemical understanding of the patient or drug, and they are always developed for a single purpose: they have a hard time generalizing to usage of new variables, data, or for prediction of new variables. ML is therefore often referred to as narrow AI. In contrast, mechanistic models can overcome all of these shortcomings, but they take a very long time to develop, and there are no mechanistic models available for all relevant processes in patients. We therefore present a new hybrid approach, which combines strengths of ML and mechanistic modelling; and illustrate how this combination can be used to develop digital twins of a patient.