# A pipeline for developing deep learning prognostic prediction models in cardiac magnetic resonance image analysis

## TL;DR

This paper introduces a four-step pipeline for building deep learning models to predict outcomes from cardiac MRI data, aiming to improve arrhythmic risk prediction in cardiology.

## Contribution

A structured pipeline for developing explainable deep learning models in cardiac MRI analysis, focusing on arrhythmic risk prediction.

## Key findings

- A four-step framework is proposed to guide DL model development for cardiac magnetic resonance image analysis.
- The pipeline emphasizes data selection, segmentation, feature extraction, and model explainability.
- Initial results suggest potential for improved arrhythmic risk prediction in dilated cardiomyopathy.

## Abstract

Patients and healthcare professionals require clinical prediction models to accurately guide healthcare decisions, although an awareness of the limitations of regression-based models has recently increased. Deep learning (DL) has emerged as a promising alternative to traditional regression-based models, due to its ability to effectively analyse heterogeneous types of data, ranging from numerical variables to medical images.

Building a DL model presents various challenges, including conceptualizing the clinical problem, selecting appropriate variables and model architecture, and providing explainability. We propose a four-step pipeline for developing DL-based prediction models for cardiac magnetic resonance image analysis. This framework aims to support researchers in exploring DL application across the broad spectrum of cardiology, with a specific focus on advancement in arrhythmic risk prediction.

The field of cardiomyopathy faces challenges when assessing arrhythmic risk due to the low accuracy of the current prediction models. Research efforts have focused on developing DL models able to predict major arrhythmic events in dilated cardiomyopathy. While the initial results are promising, further tests are needed before translating these models into clinical practice.

Graphical AbstractThe figure summarizes the four steps presented in the manuscript. The first step involves selecting the data according to the ‘five V’ principle. The second step involves deciding whether and how to apply a segmentation model during the pre-processing phase, in order to isolate only the relevant regions of interest and reduce noise and potential errors. The third step is feature extraction, where the region of interest is processed to predict a pre-specified outcome. The final step is to ensure the model's explainability, avoiding the ‘black box’ effect and making the model more trustworthy.

The figure summarizes the four steps presented in the manuscript. The first step involves selecting the data according to the ‘five V’ principle. The second step involves deciding whether and how to apply a segmentation model during the pre-processing phase, in order to isolate only the relevant regions of interest and reduce noise and potential errors. The third step is feature extraction, where the region of interest is processed to predict a pre-specified outcome. The final step is to ensure the model's explainability, avoiding the ‘black box’ effect and making the model more trustworthy.