How Frida tracks cognitive stability: a look under the hood

Frida is a self-management app for people living with schizophrenia and related conditions. Its core job is to help a person see, day by day and week by week, how stable they are — and to help the people who care about them notice when stability is slipping. This article is a look under the hood: what we measure, how we combine it, what we deliberately don't do, and the principles that shaped the design.

It is also written in the same spirit as the rest of this blog — honest about the limits, clear about what is evidence and what is product judgement.

The starting principles

Before any feature, we wrote down what we wanted Frida to be — and not be.

• Calm tech. The app should never demand attention it does not deserve. No streak counters that punish you for one bad day. No dark patterns that trade your wellbeing for engagement metrics.
• Honesty over engagement. If a feature is not making you better, the right thing is to say so. We are not selling daily active users to advertisers.
• Patient as customer, not product. Frida is paid software. We do not sell data. We do not run ads.
• Clinical credibility. Every claim is grounded in published research or established clinical practice, and clearly labelled when it is not.
• Personalisation over generic dashboards. The signs of relapse are personal. The system has to learn yours, not impose someone else's.

What Frida actually measures

Daily check-ins

Once a day, ideally as part of a morning or evening routine, the user logs four things on simple sliders: sleep (hours and quality), mood, energy, and stress. The whole thing takes under thirty seconds. We deliberately resisted the urge to ask twelve questions; the data is more useful if it is actually entered every day.

Personal early warning signs

During onboarding, the user picks three to five personal early warning signs from a curated list — "voices louder," "sleep falling off," "thoughts racing," "feeling watched," and so on — or writes their own. These become daily yes/no toggles, separate from the general mood data. The list is deliberately the user's, not ours.

Medication tracking

Each medication the user takes is tracked with simple "took it / missed it / skipped it deliberately" flags, with optional notes. We do not score adherence percentages. The point is to see the pattern, not earn a grade.

Routine markers

Lightweight markers of daily routine — did you go outside, did you eat at least two meals, did you have a conversation with someone — provide cheap signal on functioning that pure mood data misses.

Optional integrations

If the user wears a smartwatch or uses Apple Health / Google Fit, Frida can pull in sleep duration and step count to supplement the manual data. We do not require it, and we are intentional about not over-relying on passive data alone (see wearables for schizophrenia).

How those become a "stability picture"

Frida does not try to predict relapse. We were tempted, like everyone else in this space, to build a fancy ML model with a "relapse risk" score. We chose not to, for two reasons. First, the published evidence does not yet support a consumer-grade prediction that does more than basic trend lines. Second, a falsely confident risk score can do real harm, both by alarming people who are fine and by reassuring people who are not.

Instead, Frida shows the user three things:

1. A baseline — what their normal range looks like, established over the first few weeks.
2. A recent trend — how the last seven days compare to that baseline.
3. An early warning summary — how many of their personal signs have been flagged in the last week.

If recent values drift outside the baseline range for several consecutive days, or if early warning signs accumulate, the app surfaces a gentle prompt — "It looks like sleep has been unusually short for five days. Do you want to share this with your support person?" — not a red alert. The decisions stay with the user.

The support person

One of the most consistently useful design choices we made is the optional support person — a parent, partner, sibling, or care coordinator that the user explicitly invites. The support person sees a chosen subset of the user's data (say, sleep trend and early warning summary, but not journal entries) and gets a low-friction notification if the user opts to share an emerging pattern.

This mirrors a finding repeated across the digital mental health literature: tools that include a real human in the loop sustain engagement and produce real benefit at much higher rates than tools that rely on the user alone.

The AI companion

Frida includes a chat-based AI companion. It is a help feature, not a therapist. It exists to provide grounding exercises, gently externalise distressing thoughts, and offer a non-judgemental space to write things out at 2am when no one else is awake. The guardrails are explicit:

• It validates emotions but never validates delusional content.
• It uses a grounding protocol when the user appears distressed or psychotic.
• It detects crisis language and routes immediately to safety resources, including the 988 line in the US.
• It does not give medication advice, diagnoses, or medical opinions.
• It says "I don't know" and "this is something to discuss with your prescriber" frequently.

We are deliberate about the limits. AI companions can do real harm if they pretend to be more than they are, and we have studied the failure cases of others in the space carefully. See our piece on building Frida with ColdAI for more on this.

Privacy and data

Sensitive mental health data deserves the strongest privacy posture we can offer. Concretely:

• The user owns their data. Export and delete are first-class features.
• Sharing with a support person requires explicit consent and is granular — what they see is up to the user.
• We do not sell data. We do not run ads. We do not use individual data to train external models.
• The chat with the AI companion is treated as the most sensitive surface and is encrypted in transit and at rest.

What Frida is not

It is not a substitute for medication, therapy, or a clinician. It is not a digital therapeutic — we hold no FDA clearance, and we do not claim to treat schizophrenia. It is not a relapse predictor. It is not a crisis service. For crisis, the right answer is 988 in the US, your local equivalent elsewhere, or an emergency room (see when to call 911).

The technical bones, briefly

The mobile app is built with Expo and React Native, with a TypeScript Express backend and PostgreSQL as the source of truth. Local AsyncStorage holds enough data for offline use; the server is a record-keeper, not a smart application layer. Subscription management runs through RevenueCat. The AI companion uses streaming completions over Server-Sent Events with carefully written system prompts derived from CBTp principles.

None of this is exotic. We chose proven technology so we could spend our time on the product decisions that actually matter for our users.

Where this is going

The features we are working on now follow the same principles. Better personalisation of warning signs, drawn from the user's own data over time. Tighter integration with care teams who want to see summary trends between visits. Better support for caregivers, who often shoulder a lot of the load. And, slowly, more careful evaluation — including independent research — to test whether Frida actually moves the metrics we care about. If we find it does not, we will say so.

If you are using Frida and have something to tell us — what works, what does not, what we got wrong — we read every email at hello@coldai.org. The product gets better because of those notes.

This article is for educational purposes only and is not medical advice, diagnosis, or treatment. Always consult a qualified mental health professional. If you or someone you know is in crisis, call or text 988 in the US, or your local emergency number.