Human-centered AI to promote youth mental health: a serendipitous natural experiment enabled by a digital health platform.

Compliance with mobile ecological prospective assessments (mEPAs) varied significantly across the four phases of the natural experiment.

• Data were analyzed using one-way ANOVA with Tukey post hoc tests in R 4.4.2.
• Compliance was measured by completed mobile ecological prospective assessments (mEPAs).
• The study involved Indigenous youth in rural communities over the final year of a 5-year initiative.
• Compliance varied significantly across 'most phases,' suggesting not all pairwise phase comparisons reached statistical significance.

Comprehensive nudging in Phase 1 (all three nudges active) yielded the highest completion rates and fastest response times.

• Phase 1 included all three nudges: daily system-triggered reminders, weekly non-personalized messages, and weekly personalized 'Best Picture' messages.
• Completion rates and response times were compared across four phases using ANOVA with Tukey post hoc tests.
• The three nudge types were: (1) daily system-triggered reminders, (2) weekly non-personalized messages such as land-based activity reminders, and (3) weekly personalized 'Best Picture' messages showcasing youth-submitted images.

Removal of personalized scientist-triggered nudges had the most substantial impact on compliance.

• Phase 2 removed both non-personalized and personalized nudges, leading to a decline in completion rates and response times from Phase 1 levels.
• Phase 4 removed only personalized nudges, and compliance again declined relative to phases with personalized nudging.
• The 'Best Picture' nudge, which showcased youth-submitted images, was identified as particularly effective in sustaining engagement.

Reintroduction of non-personalized and personalized nudges in Phase 3 improved compliance following the Phase 2 decline.

• Phase 3 reintroduced both non-personalized and personalized nudges after they had been removed in Phase 2.
• This phase comparison was part of the natural experiment created by an unexpected system disruption.
• The four-phase structure allowed within-study comparison of nudging conditions without a planned experimental design.

Consistent system-triggered reminders and personalized 'Best Picture' nudges were identified as the most effective nudge types for sustaining compliance.

• Daily system-triggered reminders were continuously active across all phases.
• Personalized 'Best Picture' messages featured youth-submitted images and represented a form of two-way, personalized communication.
• Non-personalized weekly messages (e.g., land-based activity reminders) were less impactful than personalized nudges when considered in isolation.

The study emerged from a serendipitous natural experiment created by an unexpected system disruption within a 5-year digital health initiative in rural Indigenous communities.

• The broader initiative embedded a culturally appropriate digital health intervention into school curricula in rural Indigenous communities.
• The unexpected system disruption created four distinct nudging phases, enabling assessment of varying levels of platform nudging on compliance.
• The platform featured two interfaces: a citizen-facing mobile app for ecological assessments and nudges, and a scientist dashboard for monitoring engagement and triggering nudges.
• The study focused on the final year of the 5-year initiative.

The digital health platform's human-controlled backend and customizable citizen-facing interface reflect principles of human-centered AI, emphasizing trust and autonomy.

• The platform enabled real-time interaction between youth and scientists, supporting integration across health, education, and research sectors.
• Human-centered AI principles were operationalized through a scientist dashboard allowing manual triggering of personalized nudges.
• The authors describe this approach as 'a scalable model for ethical, effective digital interventions that balance technological precision and participant agency.'