JMIR Publications

ABSTRACT

Background:

AI systems can use wearable and self-reported data to generate personalized health recommendations, but broader access to personal data may also raise privacy concerns and reduce acceptance.

Objective:

This formative study examined how evaluations of an AI-generated health recommendation varied across stated data-provenance conditions when the recommendation itself was held constant. The aim was to identify data-use patterns that could inform the design of future AI-based personal health systems.

Methods:

In a within-participant online study, 194 Prolific participants evaluated the same sleep recommendation across four cumulative data conditions: sleep only; sleep plus physical activity; sleep, activity, and resting heart rate; and these sensor data plus personal journal entries. Participants rated trust, usefulness, personalization, comfort with sharing, privacy concern, data appropriateness, and behavioral intention. Scenario differences were tested using linear mixed-effects models with Holm-corrected contrasts.

Results:

Trust and usefulness changed little across sensor conditions. The journal condition showed the largest differences from the preceding condition: comfort was 0.74 points lower (dz=−0.47), data appropriateness 0.59 lower (dz=−0.45), trust 0.26 lower (dz=−0.32), behavioral intention 0.32 lower (dz=−0.36), and privacy concern 0.59 higher (dz=0.37; all P<.001). Personalization did not differ significantly.

Conclusions:

Broader stated data access did not consistently correspond to greater perceived value, and the condition including journal data showed lower acceptance. As a formative evaluation, these findings identify data type as an important consideration for subsequent system design and evaluation, particularly for data minimization, per-data-type consent, and provenance communication. Clinical Trial: N/A