🤖 AI Mental Health Screening Conversational Simulator

Conversation Begins

A screening bot starts like a person would, not a form.

• Open-ended: First message type (no checkbox in sight)
• Hidden: Clinical framing (questionnaire stays invisible)
• Warm: Tone target (reduces screening dropout)
• PHQ-9: Underlying instrument (9-item depression scale)

Why not just show the form

Rigid forms feel clinical and get abandoned early.

Greeting design

One open question invites a natural, unscripted reply.

Setting expectations

The bot signals safety before asking anything sensitive.

Questionnaire Items Woven In

Each PHQ-9 item is rephrased as a casual follow-up question.

• 9: PHQ-9 items (depression screening core)
• 7: GAD-7 items (anxiety screening variant)
• Conversational: Rephrasing style (not verbatim clinical wording)
• Adaptive: Item order (follows conversation flow)

Rewriting clinical language

"Little interest in things" becomes a plain, human question.

Pacing the items

One item per turn keeps the exchange feeling unhurried.

Staying on the standardized scale

Wording changes, but the measured construct never does.

Natural Language Responses

Patients type real sentences, not multiple-choice answers.

• Free text: Response format (no fixed answer options)
• 5–15 words: Typical reply length (short, conversational)
• Yes: Ambiguity handled (via gentle clarifying follow-up)
• Preserved: Language nuance kept (tone, hedging, intensity)

Why free text matters

People disclose more naturally than in dropdown menus.

Clarifying vague answers

The AI asks a soft follow-up when intensity is unclear.

Capturing nuance

Hedges like "kind of" still carry scoring signal.

Response Mapping

Free text is translated into the standardized 0–3 PHQ-9 scale.

• 0–3: Score range per item (not-at-all to nearly-every-day)
• NLP classifier: Mapping method (frequency + sentiment cues)
• Re-ask: Low confidence handling (clarifying question triggered)
• Logged: Mapping transparency (clinician-reviewable trail)

From words to numbers

"Most days" maps toward the higher frequency scores.

Confidence scoring

Ambiguous phrasing lowers mapping confidence automatically.

Human oversight

Every mapped answer stays traceable back to its source text.

Score Computed & Shared

All nine mapped scores sum into one standardized severity result.

• 0–27: Score total range (full PHQ-9 scale)
• 4: Severity bands (minimal / mild / moderate / severe)
• Conversational: Result delivery (no raw number dumped)
• Care routing: Next step (handed to clinical workflow)

Summing the scale

Nine item scores add into one total severity number.

Sharing results gently

The category is explained in plain, supportive language.

Where scoring hands off

Severity output feeds a separate triage or care step.