Personalization strategies in digital mental health interventions: a systematic review and conceptual framework for depressive symptoms

📝 Paper Summary

Digital Mental Health Interventions (DMHIs) Personalization Frameworks

This systematic review proposes a four-dimensional framework for personalization in depression interventions and finds that while two-thirds of apps use personalization, most rely on simple rules rather than machine learning.

Core Problem

Personalization is widely touted to improve adherence and outcomes in digital mental health, but the field lacks a unified definition, resulting in fragmented vocabulary and unclear implementation standards.

Why it matters:

Standard 'one-size-fits-all' digital interventions suffer from high dropout rates (up to 50%) and unsatisfactory response rates (<50%)
The lack of a shared framework conflates distinct concepts like 'customization' (aesthetic changes) with therapeutic 'personalization', hindering systematic research
Despite the technical potential of smartphones, it is unclear how much advanced personalization (like Machine Learning) is actually being used in practice

Concrete Example: A standard app might show every user the same CBT module order regardless of their specific symptoms. A personalized approach would use a decision rule to 'step up' guidance if a user's PHQ-9 score worsens, but current literature often fails to distinguish this therapeutic adaptation from trivial changes like picking an avatar color.

Key Novelty

The COGC Framework (Content, Order, Guidance, Communication)

Conceptualizes personalization as purposefully designed variation in four specific dimensions: Content (therapeutic material), Order (sequence of modules), Guidance (human support level), and Communication (prompts/timing)
Distinguishes 'Personalization' (therapeutic variation) from 'Customization' (user preference without therapeutic impact), 'Interactivity' (replying to input), and 'Group adaptation' (cultural localization)
Maps mechanisms to four types: User choice, Provider choice, Rule-based systems (static logic), and Machine Learning (adaptive algorithms)

Evaluation Highlights

66% of reviewed interventions (62/94) implemented at least one personalization strategy
Rule-based mechanisms were the most common (48% of mechanisms), while Machine Learning was rarely used (3% of mechanisms)
Only 2 of 138 included studies directly compared personalized vs. non-personalized versions, with inconclusive results

Breakthrough Assessment

7/10

Strong conceptual contribution providing a much-needed taxonomy for the field. However, empirical findings reveal the field is still immature, with very little advanced ML implementation or direct evidence of efficacy.

⚙️ Technical Details

Problem Definition

Setting: Systematic review of empirical studies on DMHIs for depressive symptoms in adults (2015–2022)

Inputs: 138 papers describing 94 distinct digital mental health interventions

Outputs: Classification of personalization strategies across dimensions (Content, Order, Guidance, Communication) and mechanisms (User choice, Provider choice, Rules, ML)

Pipeline Flow

Search Databases (PubMed, SCOPUS, PsycInfo)
Screen 3,143 abstracts
Select 138 full texts
Extract Data & Classify via Framework

System Modules

Taxonomy Generation

Define personalization dimensions

Model or implementation: Conceptual Framework

Novel Architectural Elements

Differentiation of personalization from customization/usage/interactivity
Matrix classification of Dimension (What is personalized) vs. Mechanism (How it is personalized)

Comparison to Prior Work

vs. Teepe et al.: Broader scope (all personalization, not just JIT) and finding that ML is still rare (3%) compared to rule-based methods
vs. Lustria et al.: Focus specifically on depressive symptoms and modern DMHIs (post-2015), creating a depression-specific taxonomy
vs. General personalization literature: Explicitly excludes 'customization' (aesthetic choices) to focus on therapeutic ingredients

Limitations

Review restricted to depressive symptoms; anxiety and other comorbidities were excluded unless depression was the primary target
Potential publication bias; commercial apps without published studies are underrepresented
Framework validity was not tested via expert interviews or external consensus methods
Only found 2 studies directly comparing personalized vs. non-personalized arms, limiting conclusions on efficacy

Reproducibility

The review protocol was registered on PROSPERO (CRD42022357408). Search strings and inclusion criteria are fully documented in the paper appendices.

📊 Experiments & Results

Evaluation Setup

Systematic review and classification of 94 DMHIs extracted from 138 papers

Benchmarks:

Prevalence Analysis (Frequency count of personalization strategies)

Metrics:

Percentage of interventions using personalization
Distribution of personalization mechanisms (Rule-based vs ML vs Choice)
Distribution of personalized dimensions (Content vs Communication etc.)
Statistical methodology: Descriptive statistics (counts and percentages)

Key Results

Benchmark	Metric	This Paper	Δ
Review Corpus (94 interventions)	Percentage Personalized	66	+66
Identified Mechanisms (107 total)	Percentage Rule-based	48	+48
Identified Mechanisms (107 total)	Percentage Machine Learning	3	+3
Identified Mechanisms (107 total)	Percentage User Choice	36	+36
Review Corpus (94 interventions)	Percentage Personalized Content	32	+32
Review Corpus (94 interventions)	Percentage Personalized Communication	30	+30

Main Takeaways

Personalization is common (66%) but unsophisticated; most apps use static 'if-then' rules (48%) or user choice (36%) rather than AI/ML (3%)
Most personalized interventions (69%) only adapt a single dimension (e.g., only changing reminders, or only changing content), missing the potential for holistic adaptation
Personalization based on passive sensing (e.g., sleep patterns, activity data) is technically feasible but largely absent in current interventions
Direct empirical evidence is scarce: only two studies directly compared personalized vs. standard versions, with mixed/inconclusive results regarding adherence and efficacy

📚 Prerequisite Knowledge

Prerequisites

Basic understanding of Digital Mental Health Interventions (DMHIs)
Familiarity with Cognitive Behavioral Therapy (CBT) structure (modules, guidance)
Distinction between rule-based logic and machine learning

Key Terms

DMHI: Digital Mental Health Intervention—therapeutic programs delivered via websites or mobile apps, often based on CBT

iCBT: Internet-based Cognitive Behavioral Therapy—a common form of DMHI delivering standardized psychotherapy content online

JIT: Just-In-Time adaptations—mechanisms that provide support at the exact moment of need, often triggered by context or sensors

customization: Minor adjustments based on user preference that do not change therapeutic elements (e.g., changing background color)

personalization: Purposefully designed variation in an intervention's therapeutic elements or structure to fit individual needs

rule-based: Personalization driven by static 'if-then' logic (e.g., 'if score > 10, unlock module B')

blended care: Treatment combining digital interventions with face-to-face therapy sessions