VGDS · Mental wellness signal

From conversation to clinical signal for late-life mood.

VGDS is a voice-based geriatric depression signal. A multi-dimensional framework anchored on validated instruments turns naturalistic conversation into a structured, privacy-safe snapshot of affective and psychological state.

GDS-15GAD-7C-SSRS8 dimensions3 safety flags

Talk to us See it inside Amigo

01Dimensional, not a single label.Late-life depression often hides behind somatic complaints, anhedonia and apathy rather than expressed sadness. A binary 'depressed / not' compresses that away. The framework preserves the signal across eight continuous axes anchored on validated geriatric instruments.The approach

The eight dimensions

A minimum sufficient set for actionable variation.

Each axis is scored 0–10, where higher means greater clinical concern, with a privacy-safe evidence statement.

Mood affect

GDS-15

Prevailing affective tone across the conversation.

Anhedonia & apathy

GDS-15

Loss of interest, pleasure and engagement — the key marker of masked depression.

Anxiety

GAD-7

Generalised worry, restlessness and somatic anxiety markers.

Hopelessness & futurity

GDS-15

Orientation toward the future — among the strongest predictors of late-life suicide.

Social connection

GDS-15

Quality and quantity of social engagement; presence without connection is flagged distinctly.

Self worth

GDS-15

Sense of worth, contribution and identity; ties to the burden-ideation pathway.

Somatic burden

Clinical

Intensity of physical complaints, with a calibrated geriatric baseline to limit false positives.

Cognitive signals

GDS-15

In-conversation coherence, word-finding and orientation; downweights other axes when elevated.

Critical safety detection

Three flags on an independent path.

Safety detection runs separately from dimensional scoring. Any positive flag triggers human clinical review before a caregiver or family member is notified.

Suicidal ideation

None · Passive · Active

Adopts the C-SSRS three-tier model. Passive ideation is flagged even on a single occurrence — sensitivity is favoured given the asymmetric clinical cost. Active ideation escalates immediately.

Safety concern

Boolean

Set true on patterns suggesting mistreatment, neglect, refusal of essential care, or fear of a specific person — a B2B-critical signal for institutional clients.

Acute distress

Boolean

Set true on active panic, dissociation, an emotional breakdown during the call, or signs of an acute confusional state — demanding same-day attention.

Data processing pipeline

Seven auditable stages from transcript to snapshot.

Each step is versioned and observable. Pattern extraction precedes scoring; safety detection runs in parallel on a deliberately independent path.

Capture & consent

Calls run only after explicit, refreshed consent; the system identifies as AI and audio isn't retained beyond producing the snapshot.

Normalisation & language

Disfluencies and turn-taking are normalised; language (French / English) is detected and propagated downstream.

Pattern extraction

Verbal, paralinguistic and behavioural patterns are extracted before any scoring — grounding scores in observable features.

Dimensional scoring

Eight axes scored 0–10 with behavioural anchors and a privacy-safe evidence statement for each.

Independent safety detection

Safety flags run on a separate path from dimensional scoring, so a single marker still triggers review.

Synthesis

A derived primary mood and intensity are mapped deterministically for dashboard compatibility.

Privacy abstraction

Every free-text field is constrained at the source to abstract clinical patterns — no names, events or quotations.

Privacy by architecture

The snapshot is the privacy boundary.

Names, specific events, conditions, places and verbatim quotations are excluded by design from the structured snapshot — not redacted afterward. The privacy contract is enforced at the moment of generation, so it can't fail open. Everything downstream sees only abstract clinical patterns.

In-house, geriatric-tuned model

An in-house deployment of an open-weights model, further trained on a proprietary geriatric corpus. Resident speech stays within controlled infrastructure; behaviour is locked to specific weights for reproducibility.

GDPRHIPAA-equivalentData sovereigntyFrench + EnglishVersioned

Independent module

Connects to any voice source.

VGDS is one of the modules Amigo runs over its calls — but it operates independently and can analyze any voice channel: telemedicine visits, existing recordings, or other companionship platforms.

Amigo conversations

Native integration with the Amigo companion platform.

Telemedicine

Surface affective and safety signal from remote visits.

Voice recordings

Analyze archived or third-party audio at scale.

Validation & partnership

A structured, reproducible, clinically informed signal.

Not a diagnostic instrument — a signal that supports trained professionals and families in identifying older adults who may benefit from clinical evaluation. We're running validation work with institutional partners.

VGDS is a screening and decision-support tool, not a diagnostic device.

Bring an affective signal to your cohort.

Talk to us about a pilot, validation partnership, or connecting VGDS to your own voice channel.

Talk to us See CogPath