Data-driven generation of parameters within a hierarchical model of mental health expertise

Mr Sasha Hegazy
Computer Science, Aston University

This talk presents an incremental approach to solving node weightings in a tree structure. The tree represents expertise used to quantify risks associated with mental­health problems and it is incorporated within a web­based decision support system called GRiST. The aim of the algorithm is to find the set of relative node weightings in the tree that helps GRiST simulate the clinical risk judgements given by mental-health experts. This approach extends the solution presented in our earlier ARRIVE algorithm, to incorporate a larger pool of data and previous cases in the solution, hence producing better elicitation results. It is also useful in incorporating new cases into the GRiST tree parameters estimation process, one by one as they are encountered.

The original ARRIVE algorithm showed that a very large number of nodes (several thousand for GRiST) can have their weights calculated from the clinical judgements associated with a few hundred cases (about 200 for GRiST). The new algorithm, iARRiVE, allows GRiST to learn by updating the node weightings to account for new cases. The results show that it can provide the best fit to an unlimited set of cases and thus ensure GRiST parameters provide the optimal solution for all the cases in its memory. Its solution can be applied to similar knowledge­engineering domains where relative weightings of node siblings are part of the parameter space.