MoPHES, a system described in IEEE in October 2025, detects anxiety levels during a live conversation — with 80.5% accuracy. For depression the figure is lower at 63%, yet even that outperforms models seven times its size. For the first time, mental state assessment is woven into the dialogue itself rather than relegated to a separate test.

Why couldn't chatbots do this before?

Most AI systems for mental health follow one of two patterns: either they administer standardized questionnaires (PHQ-9, GAD-7), or they hold a supportive conversation with no clinical assessment whatsoever. The first approach is tedious and disrupts natural communication. The second talks but doesn't truly listen.

A professional psychologist doesn't work that way. They continuously evaluate the client's state — through word choice, tone, and topic selection. Mild anxiety calls for emotional support and self-regulation techniques. Pronounced symptoms demand a different strategy, potentially including a referral to a psychiatrist. Without this feedback loop, a dialogue remains just a conversation.

A systematic review by Abd-Alrazaq and colleagues (2020), published in the Journal of Medical Internet Research, analyzed 12 studies of mental health chatbots. The conclusion: bots genuinely help reduce symptoms of depression and stress, but most cannot adapt their responses to the user's current state. This is precisely the problem MoPHES solves.

How does MoPHES work?

A research team from China (Wei, Zhou, Wang) proposed an architecture built on two compact language models, each with 0.5 billion parameters. The first is an assessment model: it analyzes user messages and determines levels of anxiety (4 grades) and depression (4 grades). The second is a dialogue model: it conducts the conversation informed by the assessment results.

The key mechanism: assessment occurs every 5 turns. The model doesn't wait for the person to complain — it proactively tracks changes. Results are stored locally on the device, never sent to a server.

The assessment model was trained on a dataset of 6,046 labeled samples. Roughly 30% corresponded to moderate levels of both anxiety and depression simultaneously — meaning the model was trained not just on extreme cases but also on the most common states.

How accurate is it?

MoPHES based on MiniCPM4-0.5B achieved 80.5% accuracy for anxiety detection and 63% for depression. For comparison: DeepSeek-R1-7B (a model 14 times larger) reached only 59% and 51.5% respectively. Qwen2.5-7B managed 33% and 51.5%.

The normalized score for anxiety in MoPHES was 0.927 out of 1 — near-perfect severity ranking. DeepSeek-R1-7B scored 0.853.

Depression proved more challenging. This is expected: depressive states manifest less overtly in speech than anxiety. Anxious individuals more often talk about fears, tension, and the future. Depression expresses itself through apathy, slowing down, and avoidance — signals that are harder to detect in text-based dialogue.

Why does this matter right now?

According to the WHO World Mental Health Report (2022), roughly one billion people worldwide live with mental disorders — 13% of the global population. Yet more than 70% of them never receive effective help. The problem isn't just a shortage of specialists — many people simply don't realize they need help or can't gauge how serious their condition is.

Technology that assesses mental state during an ordinary conversation changes the very point of entry. A person doesn't need to fill out a questionnaire, schedule a doctor's appointment, or admit that something is wrong. They just need to talk.

What does this mean for AI therapy?

MoPHES demonstrates that computational models of mental states can be more than a research tool — they can be part of a real product. Built-in assessment enables an AI system to do what digital tools for depression detection already do: notice a problem before the person themselves is aware of it.

Of course, 63% accuracy for depression is not clinical-grade. But MoPHES runs on the user's device, requires no internet connection, and keeps data local. For screening — a first approximation, not a diagnosis — this is a significant step forward.

It's also noteworthy that smaller models proved more accurate than larger ones. This means mental state assessment can work on a smartphone, without cloud servers and without data breach risks — provided the model is properly fine-tuned for the specific task.

These results align with the direction seen in clinical trials of AI therapists: the future belongs not to generic chatbots, but to systems that understand who they're talking to. The Nearby app is developing exactly this approach — adaptive dialogue that accounts for the user's emotional state.

Frequently asked questions

Can AI diagnose depression or anxiety disorder?

No. MoPHES and similar systems perform screening — a preliminary assessment of symptom levels. Only a psychiatrist or clinical psychologist can make a diagnosis based on a comprehensive evaluation. AI helps spot problems earlier but does not replace a specialist.

Is assessment data stored securely?

In the MoPHES architecture, all data is processed and stored on the user's device. Nothing is sent to external servers. This is a fundamental difference from cloud-based solutions and one of the key advantages of compact models.

Why is accuracy lower for depression than for anxiety?

Anxiety manifests more clearly in speech: people more often mention fear, worry, and tension. Depression expresses itself through reduced activity, apathy, and avoidance — features that are harder to extract from text. As models evolve and datasets grow, accuracy will likely improve.

When will such systems appear in real applications?

Certain elements — adaptive responses based on emotional state — are already used in some mental health apps. Full integration of assessment into dialogue, as in MoPHES, remains at the research stage, but the gap between laboratory and product is closing fast.