Figure 1
From: Individualizing deep dynamic models for psychological resilience data
The proposed new method has two essential parts, dimensionality reduction and individualized trajectory estimation. Both tasks are performed using neural networks (upper row). We train two VAEs—one for mental health (blue) and one for stressor load (red)—to estimate the distribution in the latent space for each observation. The variance of these distributions is expressed as size of the dots and reflects uncertainty. Summary statistics of the temporal pattern of latent values are used as inputs to a feed-forward neural network (the “ODEnet”) which is trained to provide ODE parameters that minimize the squared distance of the ODE solution and the latent values, where latent stressor load values are updated at each measurement time point.
