Neura

Scaling Trust in Passive AI Mental Health Screening

How might AI detect mental health risks without making people feel monitored? Neura explores how AI systems can assess mental wellness while preserving dignity, autonomy, and trust.

⏱️ Duration: 8 Weeks (2024)

👥 Team: 2 Designers · 1 Engineer

My role: Concept, Strategy & Lead UX

📱 Platforms: Smartwatch · Mobile App · Web Dashboard

🎯 Core Focus:
• Distributed System Thinking
• Human-AI Interaction & Ethics
• Privacy-first Escalation Framework

The Really Gap

Ubiquitous Sensing vs. Deeply Resisted Monitoring

Bridging the chasm between advanced hardware capabilities and older adults' emotional resistance to clinical surveillance.

The Technical Paradox: Modern smart homes and wearables capture exhaustive biometric data, yet current health tracking paradigms trigger immediate psychological rejection.

The Neura Intervention: A non-intrusive wellness strategy that dissolves rigid clinical metrics into the intuitive, comforting vocabulary of everyday routines.

The Core Tension: The Balancing Act

Normal

Mild

Moderate

Severe

Supportive Responses

Self-guided Activities

Caregiver Notified

"Depression risk accumulates quietly over time. As a user moves along this continuous clinical spectrum, design must manage the extreme stakes of timing. Triggering the wrong intervention tier at the wrong time introduces critical user risks."

🟥 Over-Reactive

Trigger: Sending alerts based on isolated emotional dips.

Consequence: Explodes user anxiety, triggers false alarms, and erodes long-term platform trust.

🟦 Under-Reactive

Trigger: Only responding when clinical thresholds are crossed.

Consequence: Commits the system to severe delays and fails to prevent preventable harm.

The Design Dilemma: How might we accurately screen for escalating risks without making users feel constantly surveilled—and still enforce safety protocols when lives are at stake?

Failure Analytics

01

Triggering Anxiety

Real-time alerts on minor emotional fluctuations turn daily life into a continuous state of high-stress alarm fatigue.

02

Clinical Friction

Formal, periodic surveys force older adults into an "exam-taking" mindset, causing emotional burnout and data falsification.

03

Identity Stigmatization

Universal, one-size-fits-all dashboards impose an omnipresent "patient identity," stripping users of their autonomy and dignity.

The Stigmatization of Traditional Health Interfaces

Deconstructing three structural flaws in unified, patient-centric monitoring design.

"Detection should not feel like detection."

To protect long-term trust, we shifted the paradigm from active clinical testing to passive, invisible interaction embedding, distributing the cognitive load across an ecosystem.

Distributed Trust Architecture

Designing where sensing, interpretation, and intervention should live

Zone 1

Zone 2

Zone 3

Transforms raw telemetry into emotional themes and self-guided interventions.

Signal

Reflection

Interpretation

Collects behavioral and biometric signals without introducing clinical friction.

Aggregates long-term behavioral patterns into interpretable risk trajectories.

Smart Home · Voice Assistant · Smartwatch

Mobile App

Caregiver Dashboard

Zone 1

Zone 2

Zone 3

The challenge was not only detecting risk.

It was deciding when intervention becomes justified.

Neura keeps mild and moderate signals private by default.

Escalation happens only when risk is severe, sustained, and requires human review.

The system notifies. Humans decide.

Three Critical Design Decisions

Decision 1: Trust Over Speed

Prioritizing time-sustained emotional trends over isolated, erratic micro-fluctuations.

Filtering Emotional Noise: Isolating transient mood swings from authentic clinical decline to mitigate disruptive alarm fatigue.

Macro-Trend Observation: Aggregating multi-day, multi-modal telemetry to deliver a deliberate, stable assessment of mental resilience.

Decision 2: Make Assessment Invisible

Embedding standard clinical screening protocols into ordinary ambient conversations.

Freeform Conversation

Embedded Screening

Behavioral Activation Prompt

Zero-Friction Screening: Eliminating intimidating medical questionnaires by continuously mapping daily speech semantics to validated scales (e.g., PHQ-9).

Contextual Semantic Processing: Decoupling the diagnostic process from formal clinical settings to reveal authentic, unfiltered mental states.

Decision 3: Separate Responsibility

Different Interfaces for Different Stakes

Life-centered Interface

(Preserving Dignity & Agency)

Clinical Insight Dashboard

（Actionable Visibility）

Role-Based Visibility: The older adult owns their clinical telemetry (7 Mild), but the interface wraps this metric in empowering, conversational feedback (You're feeling sad, I'm here to listen) rather than sterile, medical-grade emergency alerts.

Actionable Activation: Repurposing high-risk indicators into soft, inviting behavioral triggers that stimulate connection without causing alarm.

Governance & Ethics

The Escalation Protocol: Defining the Safety Boundaries

Tier 1

Local Privacy Shield

Raw voice interactions and fine-grained daily analytics are computed exclusively via an On-Device Local LLM. Data never touches the cloud; Sensitive personal signals remain private by default.

Tier 2

Contextual Consent Unlocking

If a decline persists within the mild-to-moderate range for 14 consecutive days, the system triggers a Granular Consent Prompt to the user, seeking explicit permission to unlock trend summaries for family caregivers.

Tier 3

Human-in-the-Loop Emergency

Upon detecting severe crisis indicators, Neura automatically compiles an EMR-Ready Clinical Abstract. The data bypasses automation and routes directly to the primary care physician for human verification—preventing automated panic.

Architecting a multi-tiered privacy shield and data unlocking protocol for high-risk scenarios.

Governance is not about detecting risk. It is about deciding when private data becomes shared responsibility.

Interfaces As Outcomes

Each interface exists to support a different responsibility

Passive Biometric Tracking

(Including mood, activity, sleep quality, etc.)

The Mobile Interaction Ecosystem

(Watch + Mobile)

Seamless multi-device continuity translating continuous biometric telemetry into high-affinity daily micro-interventions.

The watch is intentionally minimal.

It does not diagnose. It does not alert aggressively. It reassures.

Emotion appears before metrics.
Trends are separated from interventions.
Language is distilled into themes, not labels.

Mobile becomes the reflection space.

Users see how they feel — not how they are scored.

Action is suggested only when patterns persist.

The Clinical Insights Workbench

For interpretation — not monitoring.

A high-density web interface empowering caregivers and physicians with actionable, long-term mental health trajectories.

The web layer centralizes interpretation.

It does not surface emotional fluctuation.

It surfaces sustained patterns.

Behavioral signals are synthesized into interpretable risk indicators.

Caregivers see trends — not raw emotion.

Reports remain shareable — not automatic.

Escalation is triggered only at severe, sustained thresholds.

The system notifies.
Humans decide.

Autonomy is preserved at every layer.

What I've learnt from Neura

Defining the clinical verification pipeline and ethical countermeasures before market deployment.

This project shifted my view of AI-assisted health products.

The hardest part was not improving detection accuracy.

It was deciding when the system should stay quiet, ask for consent, or escalate to a human.

Future iterations would focus on:

Validating escalation thresholds with clinicians
Measuring long-term trust and user comfort
Testing privacy-by-default data sharing models

As AI becomes better at sensing human behavior, designing intervention boundaries becomes as important as designing intelligence itself.