Quantitative risk estimation based on clinical data, typically collected through patient questionnaires, is based on linear regression models that are readily interpretable by clinicians.
However they have limited capacity in capturing complex interaction patterns between the various attributes.
To overcome this limitation, deep learning models based on non-negative neural networks are proposed that are inspired by causal-inference graphs, combining the power of neural network in capturing complex patterns in the data with easier interpretability by construction with respect to traditional neural networks.