ECG2Gene: Deep Learning-Based Germline Tumor Susceptibility Gene Risk Assessment Using Electrocardiograms

Shanwei Zhang, Qinghao Zhao, Yize Zhao, Jun Li, Yongchao Long, Jinshuai Gu, Ling Yang, Huaidong Du, Yiping Chen, Zhengming Chen, Dianjianyi Sun, Jun Lv, Liming Li, Iona Y. Millwood, Robin G. Walters, Canqing Yu, Yuxi Zhou, Shenyong Chen, Shenda HongarXiv 2026 · arXiv

TL;DR

ECG2Gene is a novel two-stage deep learning framework that extracts "digital phenotypes" of cancer-susceptibility genes from routine 12-lead electrocardiograms (ECGs) to predict future cancer risk. By training models to recognize the subtle physiological signatures of germline mutations (like BRCA1/2), the researchers demonstrate that a standard heart test can serve as a low-cost, non-invasive tool for opportunistic cancer risk assessment.

What Problem Does This Paper Address?

Identifying individuals with hereditary cancer susceptibility—such as carriers of pathogenic variants in BRCA1 or BRCA2—is critical for early intervention. However, population-level genetic screening is hampered by high costs, limited access, and the complexity of interpreting variants of uncertain significance.

The authors propose a "phenotype-to-gene" approach. They hypothesize that because germline mutations exist in every nucleated cell (including cardiomyocytes), they might leave subtle, systemic physiological traces. If these traces are encoded in routine, inexpensive tests like the ECG, we could identify high-risk individuals during standard medical checkups, long before cancer develops.

Methods

The researchers developed a two-stage framework called ECG2Gene:

  1. Stage 1: Multi-task Gene Prediction. Using a pretrained deep learning backbone (Net1D), the model was trained to predict the carrier status of six high-risk genes (ATM, BRIP1, CHEK2, PALB2, BRCA1, BRCA2) from 12-lead ECG waveforms. This stage produces six continuous "ECG-derived gene scores."
  2. Stage 2: Cancer Risk Prediction. These gene scores are fed into an XGBoost classifier along with clinical variables (age, sex, BMI, and smoking status) to predict the future incidence of breast, ovarian, or prostate cancer.

To make the model interpretable, the authors introduced ECG-Wide Association Analysis (EWAS). Similar to Genome-Wide Association Studies (GWAS), EWAS identifies specific time points and leads in the ECG waveform that are significantly associated with particular gene scores.

The two-stage ECG2Gene framework: from germline sequencing labels to ECG-based risk prediction.

Datasets

The study utilized three major cohorts:

  • UK Biobank (UKB): The primary dataset for training and internal validation, containing matched genotype and ECG data for over 79,000 participants.
  • MIMIC-IV-ECG: Used for external validation of prostate and female cancer models.
  • China Kadoorie Biobank (CKB): An East Asian cohort used to test the model's generalizability across different ethnicities.
Flowchart of the datasets used for training, internal validation, and external testing.

Key Insights and Ideas

  • The Heart as a Systemic Sensor: The core "aha moment" is that the heart's electrical activity isn't just about rhythm; it's a high-resolution physiological signal that reflects systemic genetic health.
  • Genetically Supervised Phenotyping: Instead of training a model to predict cancer directly from ECGs (which is a very "noisy" task), the authors used genetic carrier status as an intermediate supervisory signal. This anchors the model in known biological risk factors.
  • Waveform Specificity: The EWAS revealed that different genes affect different parts of the heartbeat. For example, BRCA1 associations were prominent in the precordial leads (V1-V4) during the T-wave and other cardiac phases, while PALB2 associations clustered in the QRS complex.
EWAS Manhattan plot showing the association between the BRCA1 gene score and specific ECG time points across 12 leads.

Key Results

The model demonstrated modest but significant predictive power for future cancer risk:

  • Breast Cancer: Achieved an AUROC of 0.683 in the UK Biobank when combining gene scores with age, BMI, and smoking status.
  • Prostate Cancer: Achieved an AUROC of 0.660 in UKB and 0.709 in the CKB cohort (when combined with age).
  • Ovarian Cancer: Achieved a peak AUROC of 0.633.
  • Risk Stratification: Kaplan-Meier analysis showed clear separation between high-risk and low-risk groups. In the UK Biobank, the high-risk group for prostate cancer showed a significantly more rapid decline in cancer-free survival (log-rank $P = 0.00114$).
ROC curves for breast and prostate cancer prediction in the UK Biobank cohort.

Limitations and Future Work

  • Class Imbalance: Pathogenic variants are rare, meaning the model was trained on a relatively small number of "positive" carriers, which may limit its ability to learn complex features.
  • Modest Performance: While the results are statistically significant, the AUROCs (0.6-0.7) suggest that ECG2Gene is a tool for risk enrichment and triage rather than a definitive diagnostic replacement for genetic testing.
  • Scope: The study only looked at six genes and three cancer types. Future work aims to expand this to more genes and a wider variety of cancers.
  • External Heterogeneity: Performance varied across external cohorts (e.g., lower AUROCs in MIMIC), indicating a need for better cross-population calibration and larger paired datasets.

Summary updated

Abstract

Electrocardiograms (ECGs) may contain physiological information related to hereditary cancer susceptibility, but its relationship with future cancer risk remains unclear. Here, we present ECG2Gene, a two-stage deep-learning framework that derives digital phenotypes related to germline tumor susceptibility from routine 12-lead ECGs and predicts future cancer risk. In Stage 1, carrier status for pathogenic or likely pathogenic variants in ATM, BRIP1, CHEK2, PALB2, BRCA1, and BRCA2 supervises a multitask model that generates six continuous ECG-derived gene scores. In Stage 2, XGBoost integrates these scores with clinical variables to predict subsequent breast and ovarian cancer in women and prostate cancer in men. We also introduce ECG-Wide Association Analysis (EWAS) to characterize associations between ECG-derived gene scores and ECG waveforms. Across the clinical feature combinations evaluated in UK Biobank, the best-performing sex-specific models achieve AUROCs of 0.683, 0.633, and 0.660 for breast, ovarian, and prostate cancer, respectively. In external evaluations, models combining ECG-derived gene scores with age achieve prostate cancer AUROCs of 0.709 and 0.579 among men in the overall China Kadoorie Biobank and MIMIC-IV-ECG cohorts, respectively. ECG2Gene links genetically supervised ECG phenotyping and waveform association analysis to future cancer outcomes, demonstrating the potential of routine ECGs combined with clinical information for non-invasive, opportunistic cancer risk assessment.
BibTeX
@misc{zhang2026ecg,
  author = {Shanwei Zhang and Qinghao Zhao and Yize Zhao and Jun Li and Yongchao Long and Jinshuai Gu and Ling Yang and Huaidong Du and Yiping Chen and Zhengming Chen and Dianjianyi Sun and Jun Lv and Liming Li and Iona Y. Millwood and Robin G. Walters and Canqing Yu and Yuxi Zhou and Shenyong Chen and Shenda Hong},
  title = {{ECG2Gene: Deep Learning-Based Germline Tumor Susceptibility Gene Risk Assessment Using Electrocardiograms}},
  year = {2026},
  eprint = {2611.00614},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.00614},
}
Figure