Standard heart-risk calculators use a handful of numbers: age, sex, blood pressure, cholesterol, smoking and diabetes. They routinely label people "low risk" who go on to have heart attacks. Deep networks trained on the raw 12-lead electrocardiogram (ECG) look like the natural fix, because the ECG is cheap, universal and rich. The best published attempt, however, fuses ECG with clinical factors, beats the traditional score at the hospital where it was trained, and then falls to exact parity with that score at a second hospital. Our explanation is that much of what an ECG network learns is a noisy re-encoding of the risk factors the calculator already uses: age, sex and hypertensive remodeling. That shortcut helps at the training site and carries no extra information anywhere else.
We propose to train the ECG model to predict only the part of risk that the traditional score cannot explain. The calculator's output enters the model as a fixed starting point, and the network is penalized for encoding the calculator's own inputs. The claim is testable. A model trained this way should give up a little in-site accuracy, keep its gain over the traditional score when moved to a new population, and deliver most of that gain inside the "low-risk" group where traditional scores fail.
Core Claim
Fusion adds the most value over a traditional score when the added signal is statistically independent of that score. The clearest evidence is genetic. An integrated polygenic risk score correlated at $r = 0.01$ with the Pooled Cohort Equations, the guideline 10-year risk calculator, and still raised the C-statistic (the probability that the model ranks a random case above a random non-case) from 0.718 to 0.753 in the UK Biobank (Polygenic risk score improves the accuracy of a clinical risk score for coronary artery disease). The ECG offers no such independence for free. In Deep learning on electrocardiogram waveforms to stratify risk of obstructive stable coronary artery disease, the ECG-plus-clinical network (DL-MM) reached an AUC of 0.807 internally but only 0.716 on the external Stanford cohort, against 0.715 for the simple CAD2 score. The ECG-only network fell further, from 0.741 to 0.633.
We hypothesize that forcing the ECG pathway to be complementary to the traditional score will produce three results. It will shrink the gap between internal and external performance. It will concentrate reclassification among events the score called low risk. And it will do better than simply stacking an unconstrained ECG score on top of the calculator. If stacking matches it, the hypothesis is wrong in an informative way, and the experiments are designed to show that.
Research Program
Let $z$ be the traditional risk-factor vector and $s(z)$ the traditional score's linear predictor on the log-hazard scale. That score is the Pooled Cohort Equations for population prediction, or CAD2 for patients referred with chest pain. Let $e$ be the raw 12-lead ECG, and let $g_\theta(e)$ be a one-dimensional convolutional network with parameters $\theta$ that outputs a scalar through an embedding $h_\theta(e)$. The model's risk for patient $i$ is
$$\eta_i = \beta\, s(z_i) + g_\theta(e_i),$$
with $\beta$ refit once on the training cohort and then frozen. $g_\theta$ is trained under the Cox partial likelihood $\mathcal{L}_{\text{Cox}}$, the standard time-to-event loss, with the traditional score held fixed as an offset. Because $s(z)$ already explains what the risk factors explain, the gradient reaching the ECG network carries only the residual signal. To stop the network from re-encoding risk factors through the back door, which matters when the risk factors enter nonlinearly, we add a dependence penalty:
$$\min_\theta\; \mathcal{L}_{\text{Cox}}(\eta) + \lambda\, \mathrm{HSIC}\big(h_\theta(e), z\big).$$
Here HSIC is the Hilbert–Schmidt independence criterion, a kernel measure of statistical dependence that is zero when the two variables are independent. $\lambda$ controls how strictly the ECG pathway must avoid the risk-factor information. Setting $\lambda = 0$ gives the offset-only ablation. We sweep $\lambda$ and select it on internal validation data only.
Complete independence is not the goal. An ECG finding may mean something different at 40 than at 70. We therefore keep one low-capacity interaction term between $g_\theta$ and age, and we report its contribution separately.
Two lines of work supply the method, and one constrains it. The first method source is Constraint-Based Model in Multimodal Learning to Improve Ventricular Arrhythmia Prediction. It trains a clinical XGBoost model and then upweights the patients that model misclassifies when training a graph network on CT myocardial thickness maps. That raised sensitivity for post-infarction arrhythmia from 62.8% (clinical only) and 65.0% (early fusion) to 80.7%, in 600 patients at a single center. We borrow the idea that the imaging model should learn where the clinical model fails. Our version differs in three ways: the correction is a continuous likelihood offset rather than sample reweighting, independence from risk factors is enforced explicitly, and the target is portability across sites, which that work never tested.
The second method source is CARDIAC-FM. It aligns ECG and cardiac MRI encoders with a contrastive objective and reports an AUROC of 0.82 for atrial fibrillation using ECG plus risk scores without MRI. If its encoder weights are obtainable, they are a candidate initialization for $g_\theta$. That paper adds risk scores through a standard prediction head and reports no coronary disease endpoint.
The constraint comes from High-dimensional machine learning models for prediction of heart failure in more than 400 000 men and women from the UK Biobank. There, gradient boosting and random survival forests matched Cox regression, with C-indices of 0.78–0.80 in men and 0.81–0.83 in women, and Cox was better calibrated. Nonlinear tabular modeling is therefore unlikely to be where the gain lies. We keep the tabular side a Cox model on purpose and put all learned capacity into the ECG.
The main training cohort is the UK Biobank. It has resting ECGs, Pooled Cohort Equations inputs and incident coronary disease follow-up. External testing uses MESA, which CARDIAC-FM already used for transfer. The endpoint is incident coronary artery disease or myocardial infarction, defined identically in both cohorts. A second arm uses the Cedars-Sinai-to-Stanford angiography pair from Trivedi et al., with CAD2 as $s(z)$, if a data-sharing arrangement is possible. That arm allows a direct test on the exact collapse we are trying to prevent.
Baselines
The naive reference is the traditional score alone. This is the Pooled Cohort Equations on UK Biobank incident coronary disease, where the reported C-statistic is 0.718 (Polygenic risk score improves the accuracy of a clinical risk score for coronary artery disease), and CAD2 on the angiography cohorts, reported at AUC 0.733 internally and 0.715 externally (Deep learning on electrocardiogram waveforms to stratify risk of obstructive stable coronary artery disease). It is refit on the same training split. Any method that cannot beat it externally has failed.
The strongest existing fusion method on this question is DL-MM from Trivedi et al. It passes the ECG through a convolutional network and the clinical factors through a feedforward network, then concatenates the two before classification. Its reported numbers, AUC 0.807 internally and 0.716 externally against CAD2's 0.715, define the target. We reimplement it with the same ECG backbone, preprocessing, data splits and number of hyperparameter trials as our model, so that the only difference is how the traditional information enters. On the angiography arm, beating it means an external AUC gain over CAD2 whose bootstrap 95% interval excludes zero, where DL-MM's gain was 0.001. On UK Biobank to MESA, it means a smaller internal-to-external drop in the gain over the Pooled Cohort Equations than DL-MM shows under identical training. No reported result was found on Lacuna for ECG-based incident coronary disease prediction on UK Biobank or MESA, so all of those numbers are produced by us under matched conditions.
The most dangerous simple competitor is stacking. We train an unconstrained ECG-only survival network, then fit a Cox model on the traditional score plus that network's output. This is the obvious way to reproduce the polygenic-score result with an ECG, and it uses the same compute and tuning budget as our model.
We also adapt sequential error-weighted fusion from Fotso et al. to ECG. Patients are weighted by the traditional model's error, the same backbone is used, and the reweighting temperature receives the same tuning budget.
Two ablations isolate our ingredients. The offset-only model sets $\lambda = 0$. The penalty-only model applies the independence term but trains without the offset. Beating both shows that the two parts do separate work.
Why not just…?
Why not just stack an ECG-only score onto the Pooled Cohort Equations? This is the real threat. If the ECG network's useful signal is already largely independent of the risk factors, stacking will recover it and our constraint adds nothing. The evidence against stacking is indirect. The external collapse of the ECG-only model in Trivedi et al., from 0.741 to 0.633, suggests that much of what an unconstrained ECG network learns is not portable, and stacking inherits whatever that network learned. We settle it directly by comparing external gain over the traditional score for stacking against offset-plus-penalty. We also probe how well age, sex and systolic blood pressure can be decoded from each model's embedding. If stacking matches our model externally and its embedding is no more risk-factor-laden, we will report that the constraint is unnecessary.
Why not just add a few more blood tests to a Cox model? In 621 acute coronary syndrome patients without standard risk factors, adding fibrinogen, D-dimer, NT-proBNP and Lp(a) raised the C-statistic from 0.592 to 0.711 (Prognostic performance of multiple biomarkers in patients with acute coronary syndrome without standard cardiovascular risk factors). In UK Biobank, cystatin C and spirometry were strong non-traditional predictors of heart failure. This is complementary rather than competing, but it could absorb the ECG's gain. We therefore run a second setting in which $s(z)$ is an expanded Cox model with cystatin C, spirometry and the other available UK Biobank biochemistry. We then ask whether the residual ECG pathway still adds C-index. If it does not, the ECG adds nothing beyond cheap laboratory values in that population.
Why not just use a richer fusion architecture, such as cross-attention or TabMixer-style mixing? TabMixer mixes tabular embeddings into video feature maps across space, time and channels, reaching 6.66 mmHg mean absolute error on pulmonary pressure. Richer mixing, though, gives the network more ways to pass clinical information through the ECG pathway, which is the shortcut we suspect caused the external drop. Architecture capacity was not the bottleneck in Trivedi et al.; portability was. We include a cross-attention variant of DL-MM as an additional arm. We predict it will match or exceed DL-MM internally and not externally.
Why not just trust the guideline score? On external data the skeptic has a point: DL-MM bought 0.001 AUC. But the guideline scores' low-risk groups are demonstrably heterogeneous. Among 5,183 symptomatic patients labeled low risk, those with coronary calcium scores above 100 had an adjusted hazard ratio of 13.47 for major cardiac events (Coronary Artery Calcium Score Improves Risk Assessment of Symptomatic Patients in Low-Risk Group Based on Current Guidelines). Separately, patients with heart attacks but none of the standard risk factors rose from about 3% to 7% of cases (Comparison of Mortality Outcomes in Acute Myocardial Infarction Patients With or Without Standard Modifiable Cardiovascular Risk Factors). Whole-population AUC is therefore the wrong yardstick, so our primary metrics are computed inside the low-risk group.
Proposed Experiments
Every model is trained on UK Biobank with identical splits, preprocessing and tuning budgets, and evaluated both internally and on MESA without any refitting except the traditional score's baseline hazard. We report five measures. The first is the C-index gain over the traditional score. The second is calibration slope and intercept, which matters because the heart-failure study found machine-learning models underestimated high risk. The third is the net reclassification of events among people the Pooled Cohort Equations place below 7.5% ten-year risk, the threshold Ingelsson et al. used when 14.2% of cases were reclassified upward by genetics. The fourth is the same analysis restricted to participants with none of the standard modifiable risk factors. The fifth is the ratio of external to internal gain, our portability measure.
Several checks decide the question. If the constrained model keeps a clearly larger fraction of its internal gain on MESA than both DL-MM and stacking, the core claim holds. If its advantage in reclassifying low-risk events is larger than its advantage in the whole population, the method is doing what it was built for.
The mechanism check comes from a dose-response curve over $\lambda$. As $\lambda$ rises, decodability of age and sex from the embedding should fall, while the external gain should rise before eventually declining. A flat external curve would falsify the mechanism even if the headline numbers look good.
The angiography arm, if data sharing allows, repeats the comparison on the exact Cedars-Sinai-to-Stanford transfer where fusion previously collapsed to parity, using CAD2 as the offset. In the CICM-SCP cohort, the model could also be scored against calcium score above 100 within the guideline low-risk group, which tests whether the residual ECG signal tracks occult plaque.
Finally, we report results by ethnicity in MESA. The South Asian excess mortality in the Singapore registry, and the failure of polygenic scores in African-ancestry participants, both warn that a residual learned mostly in white British participants may not transfer across groups. Any claim of portability must be shown within each group, not only on average.
References
- Polygenic risk score improves the accuracy of a clinical risk score for coronary artery disease
- Deep learning on electrocardiogram waveforms to stratify risk of obstructive stable coronary artery disease
- Constraint-Based Model in Multimodal Learning to Improve Ventricular Arrhythmia Prediction
- CARDIAC-FM: A Multimodal Foundation Model for Cardiovascular Risk Prediction Using ECG and Cardiac MRI
- High-dimensional machine learning models for prediction of heart failure in more than 400 000 men and women from the UK Biobank
- Prognostic performance of multiple biomarkers in patients with acute coronary syndrome without standard cardiovascular risk factors
- TabMixer: Noninvasive Estimation of the Mean Pulmonary Artery Pressure via Imaging and Tabular Data Mixing
- Coronary Artery Calcium Score Improves Risk Assessment of Symptomatic Patients in Low-Risk Group Based on Current Guidelines
- Comparison of Mortality Outcomes in Acute Myocardial Infarction Patients With or Without Standard Modifiable Cardiovascular Risk Factors