Design Patterns That Work | For Developers

1. Hybrid Models: Human + Digital

Key Finding

Digital tools combined with human support consistently outperform either alone. Meta-analyses show 2-3x better outcomes with hybrid approaches compared to pure digital interventions.

Why Hybrid Works

Human involvement provides elements that current technology cannot replicate:

Accountability: Knowing a human will review progress increases engagement
Nuance: Humans catch what algorithms miss
Crisis response: Reliable escalation when needed
Relationship: Therapeutic alliance predicts outcomes
Adaptation: Humans adjust to individual needs in real-time

Implementation Patterns

Therapist + App

App supports therapy (homework, skill practice, mood tracking) with regular therapist sessions. Therapist reviews app data.

Evidence: Best outcomes in most studies. Examples: SilverCloud + clinician, Beating the Blues with coaching.

Coach-Guided Digital

Non-clinical coaches provide support and accountability via messaging. App delivers structured content.

Evidence: Effective for mild-moderate symptoms. Examples: Ginger, Lyra models.

Peer Support + Digital

Trained peer specialists augmented with AI tools. AI provides resources and prompts; human provides connection.

Evidence: Promising for community mental health. AI can provide "just-in-time" support to peer counselors.

Stepped Care Triage

AI assesses severity and routes to appropriate care level: self-help, coach, therapist, or crisis resources.

Evidence: UK IAPT model demonstrates stepped care effectiveness at scale.

Design Considerations

Key Questions

What data should flow from digital to human? How is it summarized?
What are the handoff protocols between AI and human?
How do you staff the human component sustainably?
What happens when humans are unavailable?

2. Physiological Approaches

Key Finding

Interventions targeting physiological mechanisms (HRV biofeedback, breathing, etc.) show larger, more consistent effects than purely cognitive digital interventions. Effect sizes of d = 0.8+ are common.

Why Physiological Works

Mechanism is clear: Unlike "mood tracking," we understand how resonance breathing affects HRV and vagal tone
Non-verbal: Works across languages, literacy levels, cultural contexts
Immediate feedback: Users feel effects in real-time
Lower cognitive demand: Doesn't require insight or verbal processing
Measurable: HRV provides objective outcome measurement

Evidence Base

Intervention	Conditions	Effect Size	Source
HRV Biofeedback	Anxiety	d = 0.81	Goessl et al., 2017
Resonance Breathing	Stress, anxiety	d = 0.7-0.9	Multiple RCTs
Vagal Nerve Stimulation	Depression	Varies	FDA-approved
Body Scan Meditation	Anxiety, pain	d = 0.5-0.7	Multiple meta-analyses

Implementation Patterns

Resonance Breathing Guide

Visual pacer for ~0.1Hz breathing (6 breaths/min). Maximizes HRV coherence. Simple, no special hardware required.

Our implementation: Resonance Breathing Tool

HRV Biofeedback with Sensor

Real-time HRV measurement via phone camera or wearable. User sees their coherence score improve with practice.

Note: Requires hardware/camera access; adds complexity.

Progressive Muscle Relaxation

Guided tension-release sequences. Strong evidence for anxiety, insomnia. Easy to implement as audio/visual guidance.

Biophilic Visual Immersion

Fractal patterns, nature imagery reduce physiological stress markers. Can be passive (screensaver-like) or active (interactive).

Our implementation: Biophilic Tools

Design Considerations

Technical Notes

Breathing pace matters: ~6 breaths/min (5-7 range) for resonance
Visual pacers should avoid cognitive load (no reading required)
Session length: 5-20 minutes shows effects in research
Accessibility: provide multiple modalities (visual, audio, haptic)

3. Measurement-Based Care

Key Finding

Regular symptom tracking with feedback to users and clinicians improves outcomes. The UK IAPT program demonstrates this at scale with 50%+ recovery rates.

Why Measurement Works

Early warning: Detect deterioration before crisis
Progress visibility: Users see improvement (or lack thereof)
Treatment adjustment: Clinicians can adapt based on data
Accountability: Regular check-ins maintain engagement
Evidence generation: Supports research and quality improvement

Validated Measures

Measure	Constructs	Items	Frequency
PHQ-9	Depression severity	9	Weekly-biweekly
GAD-7	Anxiety severity	7	Weekly-biweekly
PHQ-2	Depression screening	2	Daily possible
PSS-4	Perceived stress	4	Daily-weekly
WHO-5	Wellbeing	5	Weekly
Session-by-session	Treatment response	Varies	Each session

Implementation Patterns

EMA (Ecological Momentary Assessment)

Brief check-ins multiple times daily. Captures mood, behavior, context in real-world settings. Reduces recall bias.

Platform: mindLAMP excels at this.

Weekly Outcome Tracking

PHQ-9 + GAD-7 weekly with trend visualization. Alerts for significant deterioration or improvement.

Passive Sensing

Sleep patterns, activity levels, social interaction inferred from phone sensors. Reduces survey burden.

Caution: Privacy implications; correlation ≠ causation.

Clinician Dashboard

Aggregated patient data with flags for deterioration. Supports treatment decisions without data overload.

4. Engagement Without Manipulation

The Challenge

80%+ of mental health app users drop out within 2 weeks. But engagement optimization techniques from consumer apps (streaks, notifications, variable rewards) can be manipulative or harmful in mental health contexts.

Ethical Engagement Patterns

Intrinsic Motivation Design

Focus on autonomy (choice), competence (skill-building), relatedness (connection). Avoid extrinsic rewards that undermine intrinsic motivation.

Progress Visibility

Show meaningful progress (symptom trends, skills learned) not arbitrary points or streaks.

Respectful Notifications

Notifications should serve user goals, not app metrics. Easy to customize or disable. Never guilt-trip for missed sessions.

Human Connection

Even minimal human contact (weekly coach check-in, peer support) dramatically improves retention.

Engagement Anti-Patterns

Avoid These

Streaks that pressure: Missing a day shouldn't feel like failure
Variable reward schedules: Slot machine mechanics are manipulation
Social comparison: "You're behind others!" is harmful
Artificial urgency: "Limited time!" creates anxiety
Dark patterns: Making it hard to pause or leave

5. Privacy-Preserving Personalization

The Tension

Personalization can improve relevance and outcomes. But mental health data is among the most sensitive data possible. How do you personalize while protecting privacy?

Principles

Minimize data collection: Only collect what's necessary for function
On-device processing: When possible, keep data local
Clear consent: Users understand what data is used and how
Data portability: Users can export and delete their data
Transparency: Explain how personalization works

Implementation Patterns

User-Controlled Preferences

Let users explicitly choose content, timing, modalities. Respect choices without "nudging" toward more data sharing.

On-Device Personalization

Models run locally on device. Raw data never leaves phone. Federated learning for improvements.

Temporal Personalization

Learn optimal times for interventions from user patterns. E.g., "evenings seem to be harder for you - would a check-in then help?"

Content Adaptation

Surface relevant content based on stated goals and progress. Avoid "echo chambers" that reinforce rumination.

Privacy Checklist

🔒 Before Collecting Any Data

Is this data necessary for the stated function?
Can we achieve the same goal with less data?
Can processing happen on-device?
How long do we need to retain this data?
What happens if this data is breached?
Can users delete this data easily?
Have we explained data use clearly to users?

6. Cultural Adaptation Frameworks

The Challenge

Mental health is culturally constructed. Interventions developed in WEIRD (Western, Educated, Industrialized, Rich, Democratic) contexts often fail elsewhere. Digital tools can bridge or widen these gaps.

Adaptation Dimensions

Dimension	Considerations	Approach
Language	Beyond translation: idioms, metaphors, symptom expression	Cultural translation, not just linguistic
Concepts	"Depression" doesn't translate universally	Use local idioms of distress where appropriate
Imagery	Visual metaphors are culturally specific	Test with target populations; avoid assumptions
Collectivism	Individual vs. family/community focus	Consider including family features
Stigma	Varies enormously across cultures	Privacy features; consider framing (wellness vs. mental health)
Technology access	Device type, data costs, connectivity	Low-bandwidth options; SMS fallbacks

Design Process

Community involvement: Include target community members in design
Expert review: Cultural consultants assess materials
Pilot testing: Test with target population before scale
Iteration: Adapt based on feedback
Avoid assumptions: Don't assume what works for one group works for another

Summary: What the Evidence Shows

Strong Evidence

Hybrid human + digital models
HRV biofeedback and breathing
Measurement-based care
CBT delivered digitally with support

Promising But Needs More Evidence

Passive sensing / digital phenotyping
AI-augmented peer support
Just-in-time adaptive interventions
VR exposure therapy

The Common Thread

Successful patterns share a common feature: they augment human capacity rather than trying to replace it. Technology handles scale, consistency, and measurement. Humans provide connection, nuance, and crisis response.

← AI Safety Framework

Requirements for safe mental health AI.

Anti-Patterns to Avoid →

What consistently fails and why.