Automating mHealth Development: A Prompt Engineering Approach to Translate Clinical Patient Descriptions into Nursing Basic Components

Supplementary Data: Automating mHealth Development

This repository contains the supplementary materials for the paper "Automating mHealth Development: A Prompt Engineering Approach to Translate Clinical Patient Descriptions into Nursing Basic Components". These resources are provided to ensure the reproducibility of our prompt engineering methodology and to share the underlying component schemas used in our evaluation.

📂 Repository Structure

original_nbcs.zip: Contains JSON files defining the Nursing Basic Components (NBCs) used in this study. These schemas provide the structural requirements (e.g., Name, Type, Input, Output) that the Large Language Model (LLM) must follow during the translation process.

Experiments.pdf: A detailed document containing the complete history of our prompt engineering steps. It includes the specific constraints, context injections, and output format requirements used across all eight versions of the system prompt.

🛠️ Prompt Evolution Summary

The development of the final system prompt was an iterative process aimed at improving structural validity, clinical relevance, and the generation of dynamic care flows.

Ver.

Main Change

Motivation

Observed Result

1

Generic initial prompt.

Initial validation.

Vague responses; incompatible structure.

2

Added "Healthcare Professional" component.

Distinguish app actions from professional actions.

Better identification of feasible app tasks.

3

Split into Physical/Virtual interventions.

Precision in remote vs. in-person care.

Reduced errors in activity identification.

4

Automatic summarization.

Reduce complexity and output size.

Loss of critical information; too brief.

5

Hybrid approach (Gamification + Summary).

Increase engagement and balance detail.

Partial improvement; context still lacking.

6

Injection of NBC structure.

Guide parameter filling explicitly.

Significant improvement in structural accuracy.

7

Integration of Conditional Logic.

Enable dynamic user flows.

Dynamic care flows.

8

Full Clinical Context + Logic Refinement.

Increase clinical relevance.

High precision; output ready for direct integration.