Abstract
We propose a novel deep learning framework to identify clonal hematopoiesis of indeterminate potential (CHIP), a somatic mutation condition associated with adverse cardiovascular outcomes, using routine cardiac magnetic resonance (CMR) imaging. Utilizing 152 multi-view late gadolinium enhancement (LGE) scans from 136 cardio-oncology patients, we developed a convolutional neural network to (1) detect CHIP status and (2) stratify the risk of future cardiomyopathy specifically within the CHIP-positive cohort. To ensure robustness, we performed rigorous feature importance analysis to rule out reliance on demographic confounders such as age and immune checkpoint inhibitor usage. The model achieved an AUC of 0.71 for CHIP detection and, notably, an AUC of 0.87 for predicting future cardiomyopathy in CHIP-positive patients, significantly outperforming demographic-only baselines. These results demonstrate the feasibility of using LGE-CMR signatures as a non-invasive "radiogenomic" screening tool, potentially enabling accessible risk stratification and precision medicine for high-risk cardiovascular populations.
Introduction
Clonal Hematopoiesis of Indeterminate Potential (CHIP) is an age-related condition characterized by the expansion of hematopoietic stem cells carrying somatic mutations, most commonly in genes like DNMT3A and TET2. While CHIP is not a malignancy, it has emerged as a potent "hidden" risk factor for cardiovascular disease, doubling the risk of heart failure and atherosclerosis.
In the field of cardio-oncology, identifying CHIP is particularly vital. Cancer patients are often predisposed to cardiac injury due to the cardiotoxic effects of chemotherapy and radiation. However, the current gold standard for CHIP detection—next-generation sequencing (NGS)—is expensive and not routinely integrated into cardiovascular care. This study explores a "radiogenomic" approach: using deep learning to identify the imaging signatures of CHIP within standard Late Gadolinium Enhancement (LGE) Cardiac Magnetic Resonance (CMR) scans. By identifying these signatures, we aim to provide a non-invasive method to screen for genetic risk and predict future heart failure without requiring primary genetic testing.
Methodology
Model Architecture
The researchers developed a unified deep learning framework based on a modified ResNet-18 architecture. To capture the full spatial context of the heart, the model was adapted to accept a multi-view input. Instead of analyzing a single slice, the network processes a 4-channel input representing standard CMR views:
- Short Axis (SAS)
- 4-Chamber View
- Vertical Long Axis (VLA)
- Left Ventricular Outflow Tract (LVOT)
Clinical Tasks
The framework was designed to address two distinct clinical objectives:
- CHIP Screening: Binary classification to detect the presence of CHIP mutations.
- Risk Stratification: Predicting which patients within the CHIP-positive cohort would eventually develop cardiomyopathy, defined as a significant clinical drop in Left Ventricular Ejection Fraction (LVEF).
Interpretability and Confounder Control
To ensure the model was learning relevant biological features rather than "shortcut" patterns, the team utilized Grad-CAM and saliency heatmaps to visualize the regions of interest. Furthermore, they implemented a "late-fusion" hybrid model to compare imaging features against tabular data (age and medication history), ensuring the model's predictions were not merely proxies for patient age.
Dataset and Experimental Setup
The study utilized a prospective cohort from Yale-New Haven Hospital (2019–2023), consisting of 136 cardio-oncology patients and 152 total scanning series. Ground truth for CHIP status was established via targeted NGS of 74 genes with a median depth of 1000x. Within this cohort, CHIP prevalence was 41.5%, with DNMT3A and TET2 being the most frequent mutations. Longitudinal data was collected to track the subsequent development of cardiomyopathy across the patient population.
Results and Discussion
Performance Metrics
The deep learning model demonstrated strong predictive capabilities across both primary tasks:
- CHIP Detection: The model achieved an AUC of 0.71 (73% accuracy). While NGS remains the definitive diagnostic tool, this performance suggests that LGE-CMR can serve as an effective non-invasive screening mechanism.
- Cardiomyopathy Prediction (CHIP+): The model's most striking performance occurred within the CHIP-positive sub-cohort, where it predicted future heart failure with an AUC of 0.87 (81% accuracy).
- Global Prediction: For the entire cohort, the model predicted cardiomyopathy with an AUC of 0.70 (85% accuracy).
The Fibrosis Link
A key insight from the study is the validation of the link between CHIP and myocardial fibrosis. Saliency maps consistently highlighted areas of LGE (scarring), suggesting that CHIP-driven inflammation leaves a quantifiable "fingerprint" on the myocardium. This supports the hypothesis that CHIP-positive patients experience distinct pathological remodeling that can be captured by AI.
Feature Importance
In hybrid models combining imaging and tabular data, imaging features accounted for over 99% of the gradient sensitivity. In contrast, age and medication usage contributed less than 0.2%. This confirms that the model relies on genuine cardiac pathology visible in the LGE scans rather than demographic confounders.
Conclusion and Future Work
This study demonstrates that deep learning can extract "radiogenomic" markers from routine CMR scans to identify high-risk genetic conditions and predict future cardiac decline. This approach could significantly lower the barrier to precision medicine in cardio-oncology by utilizing existing imaging data to risk-stratify patients.
Future research will focus on expanding the dataset to include larger, more diverse populations, such as the UK Biobank, to validate the generalizability of these CHIP signatures. Additionally, the team aims to investigate whether specific gene mutations (e.g., DNMT3A vs. TET2) produce distinct fibrotic patterns, potentially allowing for even more granular diagnostic capabilities.
Summary updated