# Interpretable clinical decision support systems in high-risk pregnancy: a scoping review of models, methods, and implementation

## TL;DR

This paper reviews interpretable machine learning tools for maternal healthcare, focusing on high-risk pregnancy prevention and clinical adoption challenges.

## Contribution

A scoping review of interpretable clinical decision support systems for high-risk pregnancy, highlighting model types and implementation gaps.

## Key findings

- Most ML studies used Random Forests or Support Vector Machines for high-risk pregnancy prediction.
- Post hoc interpretability methods like SHAP and LIME were frequently applied in reviewed systems.
- Retrospective evaluations and limited reproducibility due to poor code/data reporting were common issues.

## Abstract

Clinical Decision Support Systems (CDSS) powered by machine learning (ML) are increasingly recognized as valuable tools for improving maternal healthcare, particularly in the prevention of high-risk pregnancies. However, their adoption in real-world settings remains limited due to concerns about transparency, reproducibility, and integration into clinical workflows. Interpretable ML methods offer a promising solution by enhancing the usability and trustworthiness of these systems. This scoping review maps interpretable CDSS for maternal high-risk pregnancy prevention and intrapartum management, including both ML and rule-based systems. We examined model characteristics, implementation and validation approaches, and interpretability methods. We searched PubMed and supplemented results with targeted screening in Google Scholar. Included studies reported interpretable outputs and clinical performance. Key data extracted encompassed study design, CDSS type, validation strategies, interpretability techniques, and clinical outcomes. Nineteen studies met the inclusion criteria. Most ML studies used Random Forests or Support Vector Machines. non-ML systems commonly implemented standardized rules, scoring systems with early-warning alerts. Post hoc methods such as SHAP and LIME were frequently used. Reporting of code/data availability was variably documented, which may limited reproducibility. Most evaluations were retrospective, constraining generalizability. Future work should prioritize transparent, prospective, and open science practices, with interpretable outputs aligned to clinical reasoning for successful integration.