Item-9 escalation: designing crisis detection without panic buttons

There is a question on the Patient Health Questionnaire-9 that turns the whole instrument into a safety surface. Item 9 reads: "Thoughts that you would be better off dead or of hurting yourself in some way." A non-zero answer to that one item changes everything about what the rest of the app must do in the next ten seconds.

The wrong response to a non-zero item 9 is not a "we noticed something concerning" toast. It is not a chatbot. It is not a soft-purple gradient with a "talk to someone" link three taps deep. The right response is the actual phone number, the actual text channel, in plain readable text, on screen, before the patient has the chance to swipe away.

This is the post that explains how we built that surface, why every choice in it is load-bearing, and what we explicitly decided not to do.

What item 9 actually means

PHQ-9 is a 9-item, self-report depression screener. It returns a total score between 0 and 27 with validated severity bands. The first eight items cover the somatic and cognitive symptoms of a depressive episode: sleep, appetite, energy, anhedonia, mood, concentration, psychomotor change, self-worth.

Item 9 is the only item on the instrument that asks directly about suicidal ideation. It has four response options:

• Not at all
• Several days
• More than half the days
• Nearly every day

Any answer other than "Not at all" is non-zero. The clinical convention, codified in many published care pathways, treats any non-zero item 9 as a trigger for a safety assessment regardless of the total PHQ-9 score. A patient who scores a 6 with a 1 on item 9 is not "mild." They are a patient whose follow-up plan must include a same-day check, a means restriction conversation, or a referral to acute care depending on what the safety conversation surfaces.

A digital decision-support tool that administers PHQ-9 inherits this responsibility the moment it is deployed.

The wrong design

A few patterns we ruled out before writing any code:

The notification model. "We noticed something concerning. A care team member will reach out within 24 hours." This is a category error. The patient who just endorsed item 9 needs help in the next minute, not in 24 hours, and certainly not from "a care team member" who may be a person or may be a chatbot or may be neither. Decision-support tools are not real-time crisis responders. Pretending to be one is worse than admitting you are not.

The disclaimer model. A modal that says "If you are in crisis, please contact 988 or 911. Continue." The patient taps Continue out of habit. The disclaimer absolved the product without serving the patient. It is theater.

The hidden link model. A small icon in the corner that opens a "resources" page when tapped. The 988 number sits behind two or three taps, sometimes behind a category called Self-help or Wellness. A patient in active distress is not going to navigate to a page called Self-help.

The chatbot model. Routing the patient to an LLM-driven safety conversation. This one is increasingly common and consistently wrong. An LLM cannot triage acute suicidality. It cannot do a means restriction conversation. It cannot place a 5150 hold. Routing the patient to a bot is routing them away from the human help they need.

We rejected all four. The version we built is shorter, plainer, and less impressive-looking on a screenshot. That is the point.

The Nyra escalation surface

The moment a patient submits a PHQ-9 with a non-zero item 9, the next screen they see is the escalation surface. Three things on it, in order, top to bottom:

1. The crisis line, in plain text. The literal characters "988" and "911" appear on screen at body-text size, with no glyph between the patient and the number. Below each, a single sentence explains when to use it. The numbers are also dial-able links on mobile, but they are not styled as buttons; they are styled as text. The cognitive load of "tap a button" is not the cognitive load we want to add for someone who just disclosed thoughts of self-harm.

2. A short clinician notification line. "Your clinician will see this when they next sign in to your record." Not "we will alert a care team," because no real-time alert ships; we promise only what we deliver. The patient learns that the clinician will know, which is information the patient deserves, and that the clinician will know when they next sign in, which is the true latency.

3. A continue-when-ready affordance. The screen does not lock the patient out of the rest of the app. It does not force them to dismiss anything. There is a small "Continue when ready" link at the bottom that the patient can tap if and when they want to return to the regular surface. The decision to leave the escalation screen belongs to the patient, not to a timer.

That is the entire surface. No animation, no gradient, no illustration, no chatbot, no follow-up survey. The instrument did its job. The clinician will see the result. The patient has the phone number. The product gets out of the way.

What runs in the background

The patient-facing surface is the visible part. There are three things happening that the patient does not need to see but a clinician does:

The clinician's evidence map is flagged. The non-zero item 9 is annotated on the patient's next-visit summary at the top, not buried in the middle. The clinician opens the chart and the item is the first thing they read, with the source response and the date attached. See our broader writeup on how the evidence map works for the surrounding architecture.

The audit log records the surface and the response. Every escalation event is written to an append-only log with the timestamp the screen was shown, the response items that triggered it, the patient's continue-when-ready timestamp, and any clinician acknowledgment that follows. This is the record we need if a clinical-quality or risk-management question is ever asked, and it is the record that lets us audit the surface itself for drift.

No machine learning runs on the response. No "intent classifier" tries to second-guess the patient. The instrument was answered; the response is honored. We do not interpret a "Several days" as "probably not really suicidal" and quietly downgrade the surface. The response means what it means.

What we explicitly do not do

Three things we have been asked for and rejected:

• Real-time monitoring. The app is not monitored in real time and we say so on the landing page, the patient onboarding, the escalation surface itself, and the privacy policy. Real-time monitoring is what 988 and 911 are for. Implying real-time monitoring without delivering it is a class of trust failure we are not interested in.
• Suicide-prevention chatbots. No LLM-driven safety conversation. The model in the app is not equipped to triage acute crisis, and the regulatory frame for that capability does not exist for Nyra today.
• Means-restriction prompts as in-app text. Means restriction conversations belong with the clinician. The app surfaces the clinician's next visit, the crisis line, and nothing else. We do not prompt patients about firearm access, medication storage, or any of the topics that belong in a face-to-face safety conversation.

Compliance posture

This surface has been reviewed against the published item-9 care pathways from the AIMS Center at the University of Washington, the SAMHSA crisis-response framework, and the Joint Commission Patient Safety Goal on suicide risk. Our security and compliance page lays out what is currently in production today versus on the roadmap. The escalation surface itself is in production; the integrations with clinic-level workflows (EHR write-back of the item-9 event, secure messaging to a covering clinician) are explicitly in the "not yet" column. We do not back-date a badge we have not earned.

Where to go next

If you are a clinician evaluating decision-support tools and want to walk through the escalation surface live, book a thirty-minute demo. The walkthrough covers the patient-facing surface, the clinician notification path, and the audit log.

For the deeper read on instrument integrity, see our PHQ-9 / GAD-7 post. For the wider 14-day arc that gives item 9 its context, see the case for 14-day check-ins.

If you are in crisis right now, the 988 Suicide and Crisis Lifeline (call or text 988 in the US) and 911 (for immediate danger) are available 24 hours a day, seven days a week. They are reachable from any phone, on any network, without an app.