Current mental health diagnoses consist of qualitative behavioral descriptions. An emergent phenomenon from deep neural network solutions can help inform and improve these diagnostic labels.
In practice, the penultimate layer of a deep model often dictates an arbitrarily high dimensional space, and by training’s end contains valuable information on how a model learned to sub-classify input subjects. Though it may not be feasible for humans to understand, say, 200 depression sub-classifications, techniques like t-SNE and k-means clustering may help group subjects in a way that is ultimately digestible by humans.
Insights from these learned groupings have the potential to usher in discoveries of deeper, distinct underlying mental conditions that ultimately advance our understanding of the root causes behind complex human behavior.