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

cs.LGOct 31, 2026

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

Authors: Shanwei Zhang, Qinghao Zhao, Yize Zhao, Jun Li, Yongchao Long, Jinshuai Gu, Ling Yang, Huaidong Du,

Organizations: Department of Computer Science, Tianjin University of Technology, Tianjin, China · National Institute of Health Data Science, Peking University, Beijing, China · Department of Cardiology, Peking University People’s Hospital, Beijing, China · Department of Epidemiology & Biostatistics, School of Public Health, Peking University, Beijing 100191, China · Institute for Artificial Intelligence, Peking University, Beijing, China · Clinical Trial Service Unit & Epidemiological Studies Unit (CTSU), Nuffield Department of Population Health, University of Oxford, United Kingdom · Peking University Center for Public Health and Epidemic Preparedness & Response, Beijing 100191, China · Key Laboratory of Epidemiology of Major Diseases (Peking University), Ministry of Education, Beijing 100191, China · State Key Laboratory of Vascular Homeostasis and Remodeling, Peking University, Beijing 100191, China · DCST, BNRist, RIIT, Institute of Internet Industry, Tsinghua University, Beijing, China

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.

Figures & tables

Figure 1: Framework Overview: A, Construction of Positive Gene Carrier Labels and Variant Pathogenicity Annotation. B, Two-Stage ECG2Gene Framework. C, AUROC for Breast Cancer in Female; Right: AUROC for Prostate Cancer in Male.

Baseline	Population / Model input	Breast cancer AUROC ( $\pm$ 95% CI)
BOADICEA FH + PRS + QRF + PV	White women	$0.66\pm 0.005$
	White women aged younger than 50 years	$0.66\pm 0.015$
	White women aged 50 years and older	$0.66\pm 0.010$
	White pathogenic variant carriers	$0.66\pm 0.025$
	Non-White women	$0.54\pm 0.035$
ECG2Gene	Women (Age + BMI + Smoke)	$0.641\pm 0.075$

Table 1: Comparison with baseline breast cancer risk prediction models in UK Biobank

Figure 2: (A) Breast cancer among female UK Biobank participants (B) Prostate cancer among male UK Biobank participants (C) Breast cancer among female China Kadoorie Biobank participants (D) Prostate cancer among male China Kadoorie Biobank participants

Cohort	Overall AUROC	CVD AUROC	No CVD AUROC
Female: Breast Cancer (ECG-Derived Gene Score + Age + BMI)
UKB	$0.669\pm 0.065$	$0.737\pm 0.148$	$0.654\pm 0.076$
CKB	$0.536\pm 0.058$	$0.426\pm 0.187$	$0.558\pm 0.053$
Male: Prostate Cancer (ECG-Derived Gene Score + Age + BMI)
UKB	$0.660\pm 0.058$	$0.607\pm 0.082$	$0.668\pm 0.076$
MIMIC	$0.558\pm 0.009$	$0.536\pm 0.016$	$0.618\pm 0.0136$

Table 2: AUROC Performance of ECG-Derived Gene Scores Across Cohorts and CVD Subgroups

Figure 3: Feature importance in the female breast and ovarian cancer XGBoost risk prediction models. The left panel shows the importance of individual gene scores and clinical variables, and the right panel summarizes the overall importance of Gene Score and clinical variables.

Figure 4: Feature importance in the male Prostate cancer XGBoost risk prediction models. The left panel shows the importance of individual gene scores and clinical variables, and the right panel summarizes the overall importance of Gene Score and clinical variables.

Figure 5: BRCA1 gene score-associated ECG time points across 12 leads.

Figure 6: Dataset

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Group	XGBoost Input	Breast Cancer AUROC (95% CI)	Ovarian Cancer AUROC (95% CI)	Prostate Cancer AUROC (95% CI)
All	Gene Score + Sex + Age + BMI + Smoke	$0.820\pm 0.035$	$0.719\pm 0.162$	$0.810\pm 0.028$
Female	Age + BMI + Smoke	$0.641\pm 0.075$	$0.405\pm 0.219$	–
Female	Gene Score	$0.516\pm 0.069$	$0.542\pm 0.278$	–
Female	Gene Score+ Age	$0.643\pm 0.065$	$0.598\pm 0.251$	–
Female	Gene Score+ BMI	$0.593\pm 0.069$	$0.495\pm 0.282$	–
Female	Gene Score+ Smoke	$0.539\pm 0.072$	$0.511\pm 0.268$	–

Appendix

Table S1 : Cancer Risk Prediction Performance Using Different XGBoost Inputs

Group	XGBoost Input	Breast Cancer AUROC	Ovarian Cancer AUROC	Prostate Cancer AUROC
All	Gene Score + Sex + Age + BMI + Smoke	$0.818\pm 0.035$	$0.719\pm 0.166$	$0.814\pm 0.027$
Female	Age + BMI + Smoke	$0.637\pm 0.076$	$0.404\pm 0.214$	–
Female	Gene Score	$0.509\pm 0.071$	$0.542\pm 0.283$	–
Female	Gene Score + Age	$0.639\pm 0.068$	$0.599\pm 0.257$	–
Female	Gene Score + BMI	$0.591\pm 0.073$	$0.495\pm 0.287$	–
Female	Gene Score + Smoke	$0.532\pm 0.075$	$0.511\pm 0.272$	–

Appendix

Table S2 : Cancer Risk Prediction Performance Using Different XGBoost Inputs Among White Participants

Group	XGBoost Input	Breast Cancer AUROC	Ovarian Cancer AUROC	Prostate Cancer AUROC
All	PV + Sex + Age + BMI + Smoke	$0.811\pm 0.037$	$0.867\pm 0.096$	$0.798\pm 0.026$
Female	Age + BMI + Smoke	$0.641\pm 0.075$	$0.405\pm 0.219$	–
Female	PV + Age	$0.624\pm 0.072$	$0.693\pm 0.238$	–
Female	PV + BMI	$0.596\pm 0.080$	$0.465\pm 0.286$	–
Female	PV + Smoke	$0.506\pm 0.062$	$0.451\pm 0.250$	–
Female	PV + Age + BMI	$0.662\pm 0.071$	$0.600\pm 0.262$	–

Appendix

Table S3 : Cancer Risk Prediction Performance Using Different Pathogenic Variants Inputs

Group	XGBoost Input	Breast Cancer AUROC	Ovarian Cancer AUROC	Prostate Cancer AUROC
All	PV + Sex + Age + BMI + Smoke	$0.811\pm 0.038$	$0.869\pm 0.097$	$0.798\pm 0.026$
Female	Age + BMI	$0.636\pm 0.077$	$0.443\pm 0.272$	–
Female	Age + BMI + Smoke	$0.637\pm 0.076$	$0.404\pm 0.214$	–
Female	PV	$0.500\pm 0.000$	$0.500\pm 0.000$	–
Female	PV + Age	$0.622\pm 0.072$	$0.695\pm 0.237$	–
Female	PV + BMI	$0.595\pm 0.081$	$0.465\pm 0.287$	–

Appendix

Table S4 : Cancer Risk Prediction Performance Using Different XGBoost Inputs Among White Participants

Group	XGBoost Input	Breast Cancer AUROC	Ovarian Cancer AUROC	Prostate Cancer AUROC
All	Gene Score + Sex + Age + BMI + Smoke	$0.899\pm 0.062$	NA	$0.728\pm 0.049$
Female	Age + BMI	$0.663\pm 0.179$	NA	–
Female	Age + BMI + Smoke	$0.669\pm 0.186$	NA	–
Female	Gene Score	$0.710\pm 0.163$	NA	–
Female	Gene Score + Age	$0.749\pm 0.137$	NA	–
Female	Gene Score + BMI	$0.502\pm 0.197$	NA	–

Appendix

Table S5 : Cancer Risk Prediction Performance Using Different XGBoost Inputs Among Participants with Cardiovascular Disease

Group	XGBoost Input	Breast Cancer AUROC	Ovarian Cancer AUROC	Prostate Cancer AUROC
All	Gene Score + Sex + Age + BMI + Smoke	$0.800\pm 0.040$	$0.706\pm 0.166$	$0.831\pm 0.036$
Female	Age + BMI	$0.634\pm 0.086$	$0.447\pm 0.264$	–
Female	Age + BMI + Smoke	$0.634\pm 0.079$	$0.410\pm 0.227$	–
Female	Gene Score	$0.676\pm 0.073$	$0.576\pm 0.247$	–
Female	Gene Score + Age	$0.619\pm 0.075$	$0.590\pm 0.250$	–
Female	Gene Score + BMI	$0.612\pm 0.075$	$0.495\pm 0.274$	–

Appendix

Table S6 : Cancer Risk Prediction Performance Using Different XGBoost Inputs Among Participants without Cardiovascular Disease

Figure S1 : KM curves for breast cancer and ovarian cancer outcomes stratified by six gene prediction scores in female participants from UK Biobank.

Figure S2 : KM curves for prostate cancer outcomes stratified by six gene prediction scores in male participants from UK Biobank.

Group	XGBoost Input	Overall AUROC	White AUROC	Non-white AUROC
Female	Age + BMI + Smoke	$0.488$	$0.493$	$0.489$
Female	Gene Score	$0.491$	$0.492$	$0.498$
Female	Gene Score + Age	$0.474$	$0.500$	$0.461$
Female	Gene Score + BMI	$0.513$	$0.501$	$0.529$
Female	Gene Score + Smoke	$0.489$	$0.493$	$0.489$
Female	Gene Score + Age + BMI	$0.489$	$0.493$	$0.491$

Appendix

Table S7 : MIMIC-IV-ECG Female Breast Cancer Prediction AUROC in Overall White and Non-white Groups

Group	XGBoost Input	Overall AUROC	White AUROC	Non-white AUROC
Female	Age + BMI + Smoke	$0.466$	$0.476$	$0.502$
Female	Gene Score	$0.528$	$0.530$	$0.541$
Female	Gene Score + Age	$0.517$	$0.540$	$0.553$
Female	Gene Score + BMI	$0.535$	$0.532$	$0.563$
Female	Gene Score + Smoke	$0.537$	$0.538$	$0.560$
Female	Gene Score + Age + BMI	$0.528$	$0.543$	$0.557$

Appendix

Table S8 : MIMIC-IV-ECG Female Ovarian Cancer Prediction AUROC in Overall White and Non-white Groups

Group	XGBoost Input	Overall AUROC	White AUROC	Non-white AUROC
Male	Age + BMI + Smoke	$0.550$	$0.514$	$0.605$
Male	Gene Score	$0.514$	$0.522$	$0.499$
Male	Gene Score + Age	$0.579$	$0.559$	$0.607$
Male	Gene Score + BMI	$0.514$	$0.520$	$0.514$
Male	Gene Score + Smoke	$0.491$	$0.480$	$0.499$
Male	Gene Score + Age + BMI	$0.558$	$0.545$	$0.589$

Appendix

Table S9 : MIMIC-IV-ECG Male Prostate Cancer Prediction AUROC in Overall White and Non-white Groups

Group	XGBoost Input	Breast Cancer AUROC (95% CI)	Ovarian Cancer AUROC (95% CI)	Prostate Cancer AUROC (95% CI)
All	Gene Score + Sex + Age + BMI + Smoke	$0.679\pm 0.010$	$0.670\pm 0.021$	$0.837\pm 0.006$
Female	Age + BMI	$0.457\pm 0.014$	$0.411\pm 0.029$	–
Female	Age + BMI + Smoke	$0.453\pm 0.015$	$0.408\pm 0.028$	–
Female	Gene Score	$0.514\pm 0.014$	$0.532\pm 0.027$	–
Female	Gene Score + Age	$0.458\pm 0.012$	$0.470\pm 0.027$	–
Female	Gene Score + BMI	$0.507\pm 0.016$	$0.527\pm 0.032$	–

Appendix

Table S10 : AUROC in MIMIC Participants Without CVD

Group	XGBoost Input	Breast Cancer AUROC (95% CI)	Ovarian Cancer AUROC (95% CI)	Prostate Cancer AUROC (95% CI)
All	Gene Score + Sex + Age + BMI + Smoke	$0.735\pm 0.008$	$0.739\pm 0.021$	$0.761\pm 0.006$
Female	Age + BMI	$0.490\pm 0.014$	$0.443\pm 0.038$	–
Female	Age + BMI + Smoke	$0.488\pm 0.013$	$0.478\pm 0.037$	–
Female	Gene Score	$0.460\pm 0.013$	$0.513\pm 0.032$	–
Female	Gene Score + Age	$0.457\pm 0.012$	$0.527\pm 0.033$	–
Female	Gene Score + BMI	$0.512\pm 0.014$	$0.535\pm 0.035$	–

Appendix

Table S11 : AUROC in MIMIC Participants With CVD

Group	XGBoost Input	Breast Cancer AUROC	Ovarian Cancer AUROC	Prostate Cancer AUROC
All	Gene Score + Sex + Age + BMI + Smoke	$0.546\pm 0.051$	$0.455\pm 0.143$	$0.682\pm 0.079$
Female	Age + BMI + Smoke	$0.536\pm 0.058$	$0.407\pm 0.119$	–
Female	Gene Score	$0.511\pm 0.055$	$0.483\pm 0.138$	–
Female	Gene Score + Age	$0.536\pm 0.053$	$0.443\pm 0.130$	–
Female	Gene Score + BMI	$0.492\pm 0.055$	$0.488\pm 0.144$	–
Female	Gene Score + Smoke	$0.507\pm 0.055$	$0.493\pm 0.137$	–

Appendix

Table S12 : Cancer Risk Prediction Performance Using Different XGBoost Inputs

Group	XGBoost Input	Breast Cancer AUROC (95% CI)	Ovarian Cancer AUROC (95% CI)	Prostate Cancer AUROC (95% CI)
All	Gene Score + Sex + Age + BMI + Smoke	$0.561\pm 0.051$	$0.448\pm 0.142$	$0.683\pm 0.087$
Female	Age + BMI + Smoke	$0.553\pm 0.060$	$0.395\pm 0.117$	–
Female	Gene Score	$0.533\pm 0.057$	$0.484\pm 0.137$	–
Female	Gene Score + Age	$0.558\pm 0.053$	$0.434\pm 0.129$	–
Female	Gene Score + BMI	$0.502\pm 0.057$	$0.489\pm 0.143$	–
Female	Gene Score + Smoke	$0.528\pm 0.058$	$0.494\pm 0.136$	–

Appendix

Table S13 : AUROC in CKB Participants Without CVD

Group	XGBoost Input	Breast Cancer AUROC (95% CI)	Ovarian Cancer AUROC (95% CI)	Prostate Cancer AUROC (95% CI)
All	Gene Score + Sex + Age + BMI + Smoke	$0.490\pm 0.176$	–	$0.631\pm 0.125$
Female	Age + BMI + Smoke	$0.469\pm 0.179$	–	–
Female	Gene Score	$0.380\pm 0.145$	–	–
Female	Gene Score + Age	$0.426\pm 0.187$	–	–
Female	Gene Score + BMI	$0.437\pm 0.177$	–	–
Female	Gene Score + Smoke	$0.379\pm 0.158$	–	–

Appendix

Table S14 : AUROC in CKB Participants With CVD

Gene	Positive N	Positive ECG	Total (Female:Male)
ATM	296	346	155:141 (52.4% F)
BRIP1	153	172	76:77 (49.7% F)
CHEK2	165	196	93:72 (56.4% F)
PALB2	124	146	53:71 (42.7% F)
BRCA1	71	80	21:50 (29.6% F)
BRCA2	226	266	93:133 (41.2% F)

Appendix

Table S15 : Summary of ECG-positive samples by gene

Gene LP/P	Case Patients $N=2750$ (%)	Controls $N=43790$ (%)	Odds Ratio (95% CI)	$P$ value
BRCA1	19 (0.6) 0.85	26 (0.06) 0.11	11.71 (6.47 – 21.18)	2.36E-12
BRCA2	38 (1.3) 1.29	109 (0.2) 0.24	5.61 (3.87 – 8.14)	6.39E-15
ATM	26 (0.9) 0.78	168 (0.4) 0.41	2.4783 (1.64 – 3.75)	9.08E-05
CHEK2	17 (0.6) 1.08	108 (0.2) 0.42	2.5159 (1.51 – 4.20)	0.0016
PTEN	1	41	NA	NA
TP53	3	5	NA	NA

Appendix

Table S16 : Associations between Pathogenic Variants in Established Breast Cancer–Predisposition Genes and Risk of Breast Cancer in the UK Biobank

Characteristic	UKB Train	UKB Validation	UKB Test	MIMIC	-	CKB
Number of participants	47478	15826	15825	155721	–	19278
Age at ECG, mean (SD)	66.48 (7.77)	66.37 (7.72)	66.36 (7.82)	56.8 (19.4)	–	58.3 (9.6)
Male sex, n (%)	22648 (47.70%)	7547 (47.69%)	7592 (47.97%)	73099 (46.9%)	–	6930 (35.9%)
BMI, mean (SD)	26.62 (4.61)	26.73 (4.59)	26.60 (4.55)	28.8 (7.1)	–	24.3 (3.5)
Smoking status, n (%)	19260 (40.57%)	6477 (40.93%)	6416 (40.54%)	8567 (5.5%)	–	4236 (22.0%)
Breast cancer events, n (%)	173 (0.36%)	57 (0.36%)	58 (0.37%)	1068 (0.7%)	–	98 (0.79%)

Appendix

Table S17 : Baseline characteristics of participants in the training, validation, test, and external validation cohorts

Figure S3 : ATM gene score-associated ECG time points across 12 leads.

Figure S4 : BRCA2 gene score-associated ECG time points across 12 leads.

Figure S5 : CHEK2 gene score-associated ECG time points across 12 leads.

Figure S6 : BRIP1 gene score-associated ECG time points across 12 leads.

Figure S7 : PALB2 gene score-associated ECG time points across 12 leads.

Clinical Risk Prediction

Clinical risk prediction estimates patients’ future health outcomes from medical records, physiological signals, imaging, and other health data.

See all 231 papers

ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation
Nov 1, 2026Zhuoyuan Li, Xidong Wu, Hanyu Cui +4Electrocardiogram Interpretation Electrocardiogram Classification
The Chinese University of Hong Kong, Hong Kong SAR · University of Dundee, Scotland, UK · Northeastern University, Shenyang, China +1
EchoRisk: A Multicentre Echocardiography Dataset and Benchmark for Cardio-Oncology
Jul 1, 2026Grigorios Kalliatakis, Georgia Karanasiou, Georgios Manikis +10Left Ventricular Ejection Fraction Estimation Healthcare
Foundation for Research and Technology Hellas, Greece · University of Ioannina, Greece · Hellenic Mediterranean University, Greece +3
ECG for laboratory values:A comprehensive physiological and technical analysis
Oct 20, 2026Walentin PalanchukClinical Prediction Electrocardiography
Carl von Ossietzky Universität Oldenburg Fach-Bachelor Informatik

Nov 1, 2026J/K move · Enter open · S save

ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation

Zhuoyuan Li, Xidong Wu, Hanyu Cui, Jiansong Gao+3

The Chinese University of Hong Kong, Hong Kong SAR · University of Dundee, Scotland, UK · Northeastern University, Shenyang, China+1

Electrocardiogram Interpretation Electrocardiogram Classification Medical QA

Open paper AskPDF arXiv

Clinical interpretation of a 12-lead electrocardiogram (ECG) often requires answering targeted diagnostic questions with clear, quantifiable decisions. We introduce ECGJev, the first ECG framework built on a decision decoder, a pretrained language model trained to return a probability for every listed option. A question-conditioned bridge compresses frozen waveform and spectrogram features into 16 ECG tokens, and a two-stage procedure adapts the bridge and low-rank adapters on the decoder's native option readout using diagnostic labels alone, scoring all options for a question in one forward pass. Evaluated across four ECG benchmarks, ECGJev achieves the highest macro-F1 on CPSC2018 (80.46%), PTB-XL (71.68%) and Chapman (96.86%), and generalizes competitively to the small-sample PTB cohort. Concept probes and perturbation analyses further show that its predictions rely directly on the ECG input, remain stable under option-order and wording variations, and transfer zero-shot to unseen waveform concepts. Our code is available at https://github.com/zhuoyuanli-lab/ECGJev.

Abstract

Clinical Risk Prediction

Explore similar work

ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation

EchoRisk: A Multicentre Echocardiography Dataset and Benchmark for Cardio-Oncology

ECG for laboratory values:A comprehensive physiological and technical analysis