Understanding Responsible Development in AI-Based Clinical Prediction Models for Mortality: Protocol for a Scoping Review

TL;DR

This study aims to understand how AI-based mortality prediction models are developed and used in healthcare, focusing on ethical and responsible practices to avoid worsening health inequities.

Contribution

The study introduces a scoping review protocol guided by the responsible research and innovation framework to assess AI-based mortality models in clinical settings.

Findings

A transdisciplinary search strategy was developed to identify relevant literature on AI-based mortality prediction models.

The review will analyze how responsible development practices are integrated into model creation and application.

Results will highlight elements of model development through the lens of ethical and inclusive research practices.

Abstract

Prognostic inequity has been identified as a barrier to accessing end-of-life care for underrepresented groups. Artificial intelligence–based clinical prediction models (AIPMs) for prognostication of mortality have the potential to offer rapid, accessible, and accurate predictions that could streamline care. However, they may also exacerbate preexisting inequities in the health care system rather than address accessibility and quality. This can be caused by erroneous outputs from biased training data, outcomes from out-of-scope operationalization, and inexplicability due to opacity.

The goal of this study is to synthesize peer-reviewed literature on the creation and application of AIPMs to prognosticate mortality in acute care settings for adult patients, offering new insights into responsible and ethical model development.

A transdisciplinary, structured search strategy was developed in consultation with librarians from both health sciences and engineering sciences. The academic databases queried were Medline, Embase, IEEE Xplore, ACM Digital Library, Compendex, and Scopus. The search was conducted in spring 2025, and the results were uploaded to Covidence. A team of reviewers will screen in 2 rounds: titles and abstracts, then full texts. Eligibility will be determined by publication in academic journals or as full-length conference proceedings, language, model output, and AI use. Data will be charted using adapted charting tools and then analyzed by descriptive, summary, and qualitative synthesis.

The search was completed on March 25, 2025, with screening starting in May 2025. Results are anticipated for January 2026.

This review will provide a comprehensive summary of AIPMs that predict mortality, highlighting the specific elements included in their development. Informed by the responsible research and innovation (RRI) framework, we will consider interest-holder engagement, interdisciplinary collaboration, and computational and clinical ethics will in the context of the four RRI dimensions: anticipation, reflexivity, inclusion, and responsiveness.