JMIR Publications

ABSTRACT

Background:

Hospitalization is a stressful experience for children, yet standard well-being assessments often rely on resource-intensive proxy reports, completed by parents, or repetitive burdensome questionnaires, which can lead to reduced data quality over time.

Objective:

This study aimed to develop and evaluate MediPals, a serious game incorporating Large Language Model (LLM)-driven non-player characters (NPCs) to interactively assess the well-being of hospitalized children through gameplay.

Methods:

MediPals was designed using a privacy-preserving hybrid architecture in which an external LLM dynamically generates dialogue trees, while a self-hosted LLM summarizes clinically relevant insights for nurses. A pilot feasibility study was conducted at the University Hospital Ghent with hospitalized children and nurses to assess usability, acceptability, and clinical potential.

Results:

Five patients and four nurses participated in the pilot study, with participation durations ranging from 2 to 10 days. Nurses reported a satisfaction rate of 71% while children rated their experience at 62%. On average, patients engaged in 2.7 daily conversations with NPCs. These findings suggest moderate acceptance and engagement from both children and healthcare providers.

Conclusions:

While the MediPals pilot study demonstrates that hybrid, safety-validated LLM architectures can securely lower communication barriers for hospitalized children, consistent well-being assessment remains heavily dependent on long-term player engagement. By pairing structural privacy safeguards with a multi-genre gamified ecosystem, future interventions have the potential to transform longitudinal pediatric well-being monitoring in hospitals.