Learning Framework for Patient Digital Twins
Digital twins can be used to model physical subjects over prolonged periods of time given certain intervention scenarios. We are developing a holistic graphical multi-modal digital twin framework for functioning in elderly, that allows to monitor humans over years, but also on smaller timescales, for example during rehabilitation processes.
We want our digital twin to combine the properties explainability, predictiveness, and integration of experts to be beneficial in actual clinical settings. Our goal is to enhance long-term patient outcomes with this new approach to modelling human functioning.
Our framework aims to offer a continuous, multi-faceted view on human body functioning. It promotes a paradigm shift to tackle the obstacles of deep learning integration into clinical environments, predominantly arising from the difficulty physicians face in interpreting complex models. Instead of requiring doctors to adapt to often not or hard to interpret AI models, we integrate conventional clinical evaluation criteria into single modality-based disentangled autoencoder latent spaces, simplifying the patient state space for clinical personnel.
The model comprises a beta-variational autoencoder that distills data into a disentangled latent space, which can be transformed into a doctor’s space, that consists of concepts the doctors are already familiar with. Each modality-based disentangled latent space delineates attributes into vectors for individual organ or mental properties, which together are used to learn an attribute graph. This latent space based graph enables the prediction of future latent space states, based on the patient’s historical data, allowing the exploration of possible intervention scenarios and offering a dynamic tool to physicians for treatment outcome prediction and hence optimized intervention planning.
In our preliminary pilot studies, we are using multimodal wearable data and health records for the detection of secondary health conditions in spinal cord injury individuals. Our digital twin framework offers flexibility in multi-modal data type integration (autoencoder), fosters trust through expert-guided interventions (interpretable doctor’s space) and predicts future patient health trajectories (GNN). With this data-based digital twin, we hope to make a contribution to patient care and the long-term modelling of individuals.
