Demo: Tracing a DSM-5 Depression Assessment from Score to Formula on Edge Hardware

Explainability claims in mobile health are usually checked by reading a paper. This demo lets attendees check one by hand. The Contextual Multimodal Linkage (CML) Framework maps speech and network biomarkers onto the nine DSM-5 indicators of a depressive episode through five closed-form layers, and on a commodity mini-PC with no cloud and no uplink, a visitor runs an assessment, unfolds any indicator down to the exact formula and weights that produced it, then swaps the context profile and watches the same evidence re-weighted live. Nothing is learned end to end, so the explanation is not reconstructed after the fact, it is the computation, and every run is deterministic and re-runnable. The demonstration uses synthetic data and claims no detection accuracy: what is on display is the decision procedure itself, which is the part a black-box model does not expose. The framework comes from Tibor Haller’s Master’s thesis and unifies the group’s earlier acoustic (Länzlinger et al.) and home-router network (CareNet) linkages to the same DSM-5 indicators.