My Visit to Professor Ziad Obermeyer at UC Berkeley

· 10 minute read

Meeting with Professor Ziad Obermeyer (School of Public Health)

Takeaways

  • Applied Mathematics?
    Cannot be the top in one field? Try to be in the top 5% in several fields!

  • In serious scenarios, like medical systems, how much do we value interpretability?
    Very much. Interpretability in this area means that we want our models to perform well on data they haven’t seen before. Besides, we need to know if it really has a sense of “science”. For example, if I edit the symptom a little, what kind of change will happen in the output, and is the change aligned with the biology?

  • Why do we still need mathematical models? “The bitter lesson” has said: the data and compute will definitely outperform engineered architecture.
    We need them. First, “Garbage in, garbage out”. We need mathematical models to make some data more “high-quality”. Secondly, sometimes we cannot directly investigate the things we want. Hence, we should use mathematical models to find proxies.

An ECG biomarker for sudden cardiac death discovered with deep learning

Ziad Obermeyer, Alexander Schubert, James Ross, Sendhil Mullainathan & Markus Lingman
Nature volume 655, pages 210–218 (2026)

Abstract:
Sudden cardiac death is, in theory, preventable with defibrillators. But every year, many patients die without defibrillators because doctors fail to predict their risk. The only predictive biomarker in wide use, cardiac left ventricular ejection fraction (LVEF), misses most sudden cardiac deaths, and flags many low-risk patients for futile defibrillators that never fire. Here we apply deep learning to a dataset linking all electrocardiograms (ECGs) in a Swedish region to death certificates. The resulting model isolates a high-risk group (2.2% of the sample) with a 7.0% annual rate of sudden cardiac death, higher than those with reduced LVEF (1.9% of the sample; 4.6% annual rate). Notably, 86.1% of the model’s high-risk patients were not flagged by LVEF. High-risk ECG patients with defibrillators implanted were 54.4% less likely to die than expected, suggesting a mortality benefit. We externally validate the model in a US health system, in which it predicts ventricular arrhythmias that cause sudden death; and a Taiwanese hospital registry, in which it specifically predicts future arrhythmic cardiac arrests. To visualize the waveform morphology ‘discovered’ by the predictive model, we pair it with a generative model of the ECG waveform. Together, they reveal a biomarker that is easily visible and robustly predicts sudden cardiac death, but has not to our knowledge been previously described. Tying the biomarker’s shape to electrophysiological first principles, we form and preliminarily test a new hypothesis on the mechanism of sudden cardiac death.

The Logical Structure of the Paper: A Chain of Falsification

This paper isn’t “we ran an experiment and the model worked.” It reads closer to a mathematical proof: state a bold claim, then actively search for the strongest counterexample, and design one targeted experiment to close each gap — repeat until no gap remains. The result is a chain of eliminated alternative explanations. Below is that chain, compressed.

Step 0 — The Problem

Sudden cardiac death (SCD) is fatal but in principle preventable: an implantable defibrillator (ICD) can shock a lethal arrhythmia back to a normal rhythm. The bottleneck is prediction. The only biomarker in clinical use, left ventricular ejection fraction (LVEF), fails in both directions — most SCD victims had normal LVEF (high false-negative rate), and most patients who get an ICD for low LVEF never fire it (high false-positive rate).

Claim to test: a deep learning model reading raw ECG can out-predict LVEF.

Step 1 — Discrimination: does the score rank risk at all?

Metric: AUC, defined as

\[\text{AUC} = P(\text{score}(X_{\text{died}}) > \text{score}(X_{\text{survived}}))\]

the probability that a random SCD case is scored higher than a random survivor. AUC is chosen over accuracy because the base rate is only ~0.6%/year — a trivial “never predict SCD” classifier hits 99.4% accuracy while being useless. AUC is robust to this imbalance because it measures ranking, not hit rate.

Result: AUC = 0.872, vs. 0.697 for the existing AHA/ACC risk score and 0.655 for a prior published ECG deep-learning model.

Gap: the training label is cause of death on a death certificate — a noisy, often guessed label. A model with high AUC here might just predict “cardiac-coded death” broadly, not the specific arrhythmic mechanism an ICD can actually treat.

Step 2 — Mechanism: is it really tracking arrhythmia?

Fix: replace the death-certificate endpoint with VF/VT (ventricular fibrillation/tachycardia), the proximate mechanism an ICD interrupts — observable even in survivors, since not everyone with VF/VT dies.

Result: within the model’s high-risk group, an additional 3.8%/year develop VF/VT (95% CI 2.2–7.0%), rising monotonically with risk score. (Censoring — a patient who dies before a VF/VT event can be logged — is corrected for explicitly.)

Gap: all data comes from one Swedish region, one healthcare system. Is the model exploiting local artifacts rather than a general physiological signal?

Step 3 — Zero-shot transfer: does it survive a new country and device?

Fix: run the frozen model, with no fine-tuning, on an external US cohort — different ECG hardware, different sampling rate.

Result: AUC = 0.822 for VF/VT in the US, higher than Sweden’s 0.717 — plausibly because US billing incentives capture VF/VT events more completely, and the US cohort spans a wider risk range.

Gap: this still only answers “can it rank risk,” not “is the signal specific to arrhythmia.” A model that merely learned “this patient is generally very sick” would also rank well here — and would be clinically useless, since an ICD cannot treat generic illness.

Step 4 — Specificity: a placebo negative control (Taiwan)

The sharpest test in the paper. If the model only detects general sickness, it should perform equally well on any severe illness. Test it on a population where cardiac-arrest cases are chart-reviewed and split into:

  • arrhythmic arrest (n = 96) — the true target
  • non-arrhythmic arrest (stroke, respiratory failure, etc.) — the placebo

Result: AUC = 0.767 for arrhythmic arrest vs. AUC = 0.582 (near chance) for non-arrhythmic arrest, P < 0.001. The model fails exactly where a mechanism-specific model should fail — which is what makes it credible.

Step 5 — From ranking to causation: does the ICD actually help?

Everything above answers “can we identify high-risk patients,” not “does treating them work.” That’s a causal question, and there is no RCT here — only observational data, modeled with an OLS interaction term:

\[\text{SCD} = \beta_0 + \beta_1(\text{high risk}) + \beta_2(\text{has ICD}) + \beta_3(\text{high risk} \times \text{has ICD}) + \text{controls}\]

$\beta_3$ is the quantity of interest: do patients who are both high-risk and have an ICD show less mortality than the two main effects alone would predict?

Result: $\beta_3$ implies 54.4% lower-than-expected mortality (P < 0.001).

Gap, stated plainly by the authors: ICD assignment isn’t randomized. Physicians choose who gets one, and that choice correlates with unmeasured confounders (more attentive overall care, closer follow-up, etc.). $\beta_3$ conflates true causal effect with selection bias — the two cannot be cleanly separated from observational data alone.

Step 6 — Calibrating the framework against a known answer

Since the confounding in Step 5 can’t be removed, the authors instead check whether the method itself is trustworthy: run the identical regression on LVEF, a risk marker whose ICD benefit has already been established by multiple RCTs (effect size 50–88% mortality reduction).

Result: the same regression recovers 67.5% — squarely inside the RCT-validated range.

This doesn’t prove Step 5’s causal estimate is unconfounded; it proves the regression framework isn’t systematically broken. A badly miscalibrated framework would be unlikely to reproduce a known RCT effect size by chance. This is calibration, not confounder removal — a distinction the authors are careful to state.

Step 7 — Opening the black box

A high-AUC model is still a black box unless it can say what it sees. The authors pair the predictor with a generative model (VAE) of the ECG waveform, then take a real low-risk waveform and push it, gradient by gradient, toward higher predicted risk — producing a sequence of synthetic, increasingly “high-risk” heartbeats a cardiologist can read directly.

Findings:

  1. Left axis deviation — an already-known ECG marker (associated with left anterior fascicular block). Its emergence validates the method: the morphing procedure recovers something medically real, not noise.
  2. A previously undescribed notch at the end of the QRS complex in lead aVL — quantified via first/second derivatives of QRS amplitude, shown to be an independent, significant predictor with strength comparable to left axis deviation. The authors hypothesize this reflects myocardial fibrosis (collagen deposits disrupting electrical conduction), with preliminary cMRI support — explicitly framed as hypothesis-generating, since definitive diagnosis would require endomyocardial biopsy, rarely performed clinically.

The chain, compressed

flowchart TD
    P["Problem<br/>LVEF predicts SCD poorly —<br/>can ECG + deep learning do better?"]
    S1["Step 1 — Discrimination<br/>AUC = 0.872"]
    S2["Step 2 — Mechanism<br/>VF/VT tracking"]
    S3["Step 3 — Generalization<br/>US zero-shot transfer, AUC = 0.822"]
    S4["Step 4 — Specificity<br/>Taiwan placebo control<br/>0.767 vs 0.582"]
    S5["Step 5 — Causation<br/>OLS interaction: ICD × high-risk<br/>54.4% lower mortality"]
    S6["Step 6 — Calibration<br/>same regression on LVEF<br/>recovers RCT range 50–88%"]
    S7["Step 7 — Interpretability<br/>generative model →<br/>novel aVL notch, fibrosis hypothesis"]

    P --> S1
    S1 -->|"gap: noisy death-certificate label"| S2
    S2 -->|"gap: single-country data"| S3
    S3 -->|"gap: arrhythmia-specific, or just 'sick'?"| S4
    S4 -->|"specificity resolved"| S5
    S5 -->|"gap: observational, confounded"| S6
    S6 -->|"framework validated"| S7

Why this structure is worth studying

The mathematical habit on display is close to proof by contradiction paired with adversarial peer review: after every claim, ask “what is the strongest way this could still be wrong?”, then design the one experiment that would falsify it. A less careful paper stops at Step 1 (AUC = 0.872) and calls it done. But a single high AUC proves almost nothing clinically — it can hide label noise, geographic overfitting, mechanism confounding, and selection bias, all at once. What makes this paper credible isn’t model strength; it’s that every one of those failure modes was hunted down and closed off, one at a time, with the authors stating plainly where each fix still falls short.