CMR-Analysis (including machine learning)

1355246 - Discriminating hypertrophic cardiomyopathy (HCM) patients from healthy, athletic, and hypertensive patients using pure shape analysis

Friday, January 27, 2023

2:18 PM – 2:30 PM

Location: Silver Pearl 3

Presenting Author(s)

HS

Hunain Shiwani, MD

Clinical Research Fellow
University College London and Barts Heart Centre, United Kingdom

Primary Author(s)

HS

Hunain Shiwani, MD

Clinical Research Fellow
University College London and Barts Heart Centre, United Kingdom

Co-Author(s)

YJ

Yue Jiang, MSc

PhD student
University College London, United Kingdom
Joao B. Augusto, MD

Consultant Cardiologist
Hospital Prof Doutor Fernando Fonseca
Lisbon, Portugal
RH

Rebecca Hughes, MD

Clinical Research Fellow
University College London and Barts Heart Centre
London, United Kingdom
LL

Luis R. Lopes, PhD

Associate Professor
University College London and Barts Heart Centre, United Kingdom
JA

Jessica Artico, MD

Clinical Research Fellow
St Bartholomew's Hospital, England, United Kingdom
SR

Stefania Rosmini, MD

Consultant Cardiologist
King's College Hospital NHS Foundation Trust, England, United Kingdom
SC

Silvia Castelletti, MD

Cardiologist
Istituto Auxologico Italiano, Italy
GJ

George Joy, MBBS

Research Fellow
University College London, United Kingdom
IP

Iain Pierce, PhD

Scientist
Barts Heart Centre at St Bartholomew's Hospital, United Kingdom
LB

Liam Burke

PhD Student
University College London, United Kingdom
HX

Hui Xue, PhD

Director, Imaging AI Program
National Institutes of Health
Bethesda, Maryland, United States
PK

Peter Kellman, PhD

Senior Scientist
National Institutes of Health, Maryland, United States
AH

Alun Hughes, MD, PhD

Professor of Cardiovascular Physiology and Pharmacology
University College London, England, United Kingdom
Thomas A. Treibel, MD, PhD

Consultant Cardiologist
University College London, England, United Kingdom
Charlotte Manisty

Consultant Cardiologist
University College London and Barts Heart Centre
London, England, United Kingdom
SM

Saidi A. Mohiddin, MD, PhD

Consultant Cardiologist in Inherited and Acquired Heart Muscle Disease and in CMR
Barts Heart Centre at St Bartholomew's Hospital, England, United Kingdom
Gabriella Captur, n/a

Consultant Cardiologist, Senior Clinical Lecturer
UCL
London, England, United Kingdom
James C. Moon, MD

Clinical Director, Imaging
Barts Heart Centre and UCL
London, England, United Kingdom
Rhodri Davies, MD, PhD

Associate clinical professor
University College London
London, Wales, United Kingdom

Disclosure(s):

Thomas A. Treibel, MD, PhD: No relevant disclosure to display

Charlotte Manisty: No relevant disclosure to display

Background: Hypertrophic cardiomyopathy (HCM) is a common inherited cardiac disease defined by left ventricular hypertrophy (LVH) and increased wall thickness. However, distinguishing mild or early expressions of HCM from adaptive changes (exercise) or afterload (hypertension) can pose a diagnostic challenge. We sought to improve diagnostic accuracy of HCM by analysing myocardial thickness distribution.

Methods:

Cardiac magnetic resonance (CMR) cines from 491 healthy volunteers and 149 patients that fulfilled HCM criteria recruited locally were automatically analysed using a clinically validated AI algorithm (1, 2) to measure the wall thickness and map each to a common coordinate frame.

A dimensionally reduced model of wall thickness distribution was built using principal component analysis (PCA) to derive eigenvectors (or wall thickness "shape components”) for all subjects. The shape components were then combined with age, sex, height, weight, end-diastolic volume (EDV), end-systolic volume (ESV) and left ventricular mass (LVM). All values were used to train a gradient boosted tree algorithm (XGboost) to discriminate HCM from health.

A model explanation method was used to find the most discriminative patient demographic details, volumetrics, and shape components (SHAP (SHapley Additive exPlanations) analysis) [Figure 2]. To prevent overfitting, parameters that provided low discriminative value were excluded from the final model.

Validation was performed on four independent datasets, acquired at different institutions. They consisted of 4056 healthy volunteers from the UK Biobank, 376 hypertensives (on 3+ anti-hypertensive agents) from the UK Biobank, 152 veteran athletes, and a separate 489 cohort of patients with a clinical diagnosis of HCM.

Results:

The 4 parameters most predictive of HCM were shape components that favoured increased global wall thickness, asymmetric septal hypertrophy, increased apical thickness and a spiral distribution of hypertrophy. [Figure 1]. Age, sex, height, weight, EDV, ESV and LVM had no incremental discriminative value above shape components to diagnose HCM and were therefore excluded from the final model [Figure 2].

Overall model accuracy using 4 shape components was 91% with a sensitivity of 88%, specificity of 92%, +LR of 10 [95%CI: 9.42 - 12] and -LR of 0.13 [95%CI: 0.10 - 0.17]. The model classified 311/4056 healthy volunteers, 0/152 veteran athletes, 77/376 hypertensives and 431/489 HCM patients as having HCM [Figure 3].

Conclusion: We describe a machine learning method capable of discriminating HCM from healthy patients, athletes and hypertensive patients using the distribution of myocardial thickness alone. Beyond descriptive features, we have also now been able to quantify the value of shape components most predictive of HCM.