Skip to main content

Healthcare Robotics and Therapy

A woman on a dementia ward becomes restless in the late afternoon. A device on her bedside table plays something familiar — regional songs from her youth. She starts to sing along, fragmentary. The device follows: adjusting tempo, holding the harmonic thread when she loses it, softening when she goes quiet. It is not playing a recording. It is accompanying her, in real time, through a moment that will not repeat the same way tomorrow.

The clinical case is established. The German S3 Demenz guideline (DGPPN/DGN, 2023) names music therapy explicitly as a first-line non-pharmacological intervention for agitation in dementia; the UK's NICE NG97 recommends psychosocial and environmental interventions for distress before antipsychotic medication. The constraint is not whether the approach works; it is staffing. Germany has roughly 1.8 million people living with dementia, with a projected nursing shortfall of around 500,000 positions by 2030, and only 12.3% of music therapists globally work in geriatric settings. What helps through the difficult hours is often simple. Time is what a short-staffed ward does not have.

The broader argument for why this market — adaptive sound, not song generation — is the consequential one is in the journal piece The Wrong Debate. This page covers the clinical specifics.

Why cultural specificity is clinical, not aesthetic​

The right music for a patient is shaped by the era, region, language, and household repertoire they grew up in. Generic Western-commercial playlists cover a vanishing fraction of that surface for anyone outside a narrow demographic. A device that has to find familiar material for a 1950s Bavarian patient, a Tamil grandmother in Düsseldorf, or a retired Mexican farmworker on a Texas ward is querying a corpus along dimensions commercial catalogues were never indexed for.

CORPUS's diversity bonus rewards underrepresented contributions; the semantic pipeline makes the resulting library queryable by region, era, tradition, instrumentation, ceremonial context. The patient is already agitated. There is no re-roll. The model has to land on the first try — which a generalist song generator on a prompt-discard-reprompt loop cannot.

Alarm fatigue​

Uniform alarm tones across devices, units, and urgency levels train clinical staff to filter the entire band out. A CORPUS-trained model can produce alarm cues that differentiate urgency through timbral and harmonic shape, coherent across a hospital's device estate, distinct enough per tier to stay legible after a long shift.

Why hospital procurement requires CORPUS-grade licensing​

  • Provenance under audit, not vendor attestation. EU AI Act Article 53(1)(d) makes a training-content summary an obligation for GPAI providers; clinical integrators inherit the burden.
  • Personality rights on vocals. Any patient-facing voiced or sung material rests on a corpus where the named singers consented; see Personality rights and vocal performance.
  • EU data residency. CORPUS infrastructure runs on self-administered servers in Germany under EU data protection law; see Data residency.

A model that cannot be defended under audit cannot be deployed where audit is a precondition for procurement. See Why CORPUS.