Cardiac resynchronization: a digital twin to choose where to place the lead in the coronary sinus — a 74-patient pilot, nothing validated prospectively

A team from the Ural universities (Russia) proposes a decision-support tool for cardiac resynchronization therapy: from a cardiac digital twin rebuilt by imaging, electrophysiological simulations and an explainable classifier, it assigns to each possible left-ventricular lead location within the coronary sinus a predicted probability of response to treatment. On 74 candidates, the model beats a reference clinical calculator (accuracy 0.78 vs 0.58; AUC 0.78 in internal validation). But validation is retrospective and internal, the cohort is small, two heterogeneous datasets were merged, and the central result — for five non-responders, "another lead position would have produced a response" — remains an in-silico prediction that the authors themselves decline to present as a clinical fact.

The context

Cardiac resynchronization therapy (CRT) treats heart failure with left bundle branch block: a pacemaker is implanted to fire both ventricles together, correcting their desynchronized contraction. The field's known problem is that this treatment fails in about 30% of patients, often because the lead pacing the left ventricle is placed in a poor spot. That lead runs through the coronary sinus, the network of veins on the heart's surface, and the surgeon has only a limited choice of accessible branches.

How to choose the best branch before the procedure? Strategies proposed so far target the latest-activated region of the left ventricle, avoid areas of fibrosis (scar tissue that conducts electricity poorly), or maximize an electrical delay measured at implantation. None, the authors note, combines within a single optimization the patient's own venous anatomy, their fibrosis map, and the simulated electrical effect of pacing at one precise point or another. And no official guideline mandates any site-optimization method. It is this gap the paper seeks to fill.

The method

The centerpiece is a digital twin: a computational copy of a patient's heart, built from their CT scan (anatomy of the ventricles and coronary veins) and a late-gadolinium-enhancement MRI (LGE-MRI, which reveals fibrosis). Ventricles, lungs, torso and coronary sinus are segmented manually; fibrosis is mapped and then embedded into a volumetric mesh. On this model, the propagation of the electrical wave is simulated by an Eikonal model — an equation computing the arrival time of the activation front at each point — with slowed conduction (factor 0.3) in fibrotic zones. The model is calibrated by fitting the simulated QRS duration to the patient's actual electrocardiogram.

On this twin, biventricular pacing is simulated at multiple candidate sites along the accessible veins, and electrophysiological indices are extracted. These simulated indices are fused with clinical variables (ejection fraction, ventricular volumes, QRS duration) and geometric measures (distance from lead to fibrosis, distance to the late-activation area) into a "hybrid" feature set. A machine-learning classifier receives this vector and produces a score between 0 and 1: the predicted probability of CRT response for that precise site. Some features are recomputed at each location, enabling site-by-site prediction.

The cohort gathers 74 candidates from the Almazov National Medical Research Centre (St Petersburg): 55 patients from earlier work (without detailed venous anatomy) and 19 new patients with CT venography and fine fibrosis mapping. Response is defined by echocardiography as an ejection-fraction improvement of more than 10% at one year. The set counts 30 responders (40.5%) and 44 non-responders (59.5%). The pipeline compared five algorithms; the best is a support vector machine (SVM) with a Gaussian kernel, retaining 11 variables. Evaluation uses leave-one-out cross-validation (LOO-CV: train on 73 patients, test on the 74th, and repeat for each), with all preprocessing steps sealed inside each fold to avoid data leakage. A SHAP analysis (assigning each variable its contribution to the score) provides the rationale for each prediction.

The results

In internal validation, the classifier reaches an AUC of 0.78, accuracy 0.78, sensitivity 0.80 and specificity 0.77. The AUC (area under the ROC curve) measures the ability to rank a responder ahead of a non-responder: 0.78 means that, drawing at random one responder and one non-responder, the model gives the higher score to the right one about 78 times out of 100. Compared on the same cohort against the reference clinical calculator (Feeny et al.), the model does markedly better: accuracy 0.78 vs 0.58, F1-score 0.75 vs 0.43. A bootstrap analysis (200 resamples) yields a mean AUC of 0.85 (95% CI: 0.70–0.95). The most influential variables, per SHAP, are baseline ejection fraction (the higher it is, the lower the improvement score), heart-failure class, BMI, then the lead's distance to fibrosis and to the late-activation area.

The headline result concerns the 19 patients with reconstructed venous anatomy, of whom 13 were actually non-responders. For 8 of them, the model finds no accessible site in the coronary sinus giving a positive score: conventional CRT would be doomed regardless of location, which would argue for a different pacing strategy. For the other 5, the model designates an alternative site where it predicts a response. Clinical translation: there is none yet, and the authors say so. These 5 "reclassified as virtually positive" were never implanted at the suggested site; it is a computer prediction, not a result measured in a patient. No count of false positives and false negatives in a real population, no prospective validation yet supports this promise.

What is good

A real clinical problem, and an approach that finally ties prediction to anatomy. Predicting CRT response is an old topic, but the paper does something new: it attaches the prediction to a precise location in the veins actually accessible to the surgeon, not to an abstract ventricular surface. Mapping response probability along the coronary sinus, before the procedure, answers a question the practitioner truly asks.

A serious ML methodology, honest about its own figures. Data leakage is actively avoided (preprocessing inside each fold), hyperparameter optimization is nested, and the authors report that the permutation test is significant for AUC but not significant for the F1-score (p = 0.26) — meaning the decision threshold is fragile at this sample size, a rare and welcome admission. The SHAP interpretability is no varnish: it gives, patient by patient, the anatomical reason for the score.

Explicit restraint about scope. The text itself calls the reclassifications "in silico, not clinically validated" predictions, notes that two misclassified responders are corrected by no alternative site, and concludes that prospective validation and regulatory evaluation are needed before any deployment. This restraint contrasts with the usual tone of health-AI announcements.

What is less good

Purely internal validation, on a small single-center cohort. Everything rests on 74 patients from a single Russian hospital, in LOO-CV: no external validation, no prospective cohort. This is an unaddressed population bias — we do not know whether the model holds on another center, another scanner, another population. The bootstrap AUC of 0.85 is more flattering than the LOO AUC of 0.78 because it resamples the same data, and its very wide confidence interval (0.70–0.95) betrays the small sample.

The core result is untested — it is circular. The classifier is trained on the outcome observed at the site actually implanted, then used to score hypothetical sites where no lead was placed. Claiming another location "would have produced a response" assumes exactly what needs proving: that moving the lead changes the outcome. Without a trial that actually implants at the suggested site, the clinical contribution remains a conjecture — the authors acknowledge it, but this is the very limit that hollows out the title's promise.

Two heterogeneous datasets merged, a single comparator, a heavy and non-open pipeline. The 55 old and 19 new patients differ significantly in volumes, QRS and fibrosis: a distribution shift that the sole Quantile Transformer mitigates without truly correcting ("does not constitute formal domain adaptation"). The only comparator is a clinical calculator fed fewer variables, which makes the win expected. Finally, segmentation is manual, the required imaging (CT venography + late-enhancement MRI) is not routine, and while the building blocks are open-source (FEniCS, scikit-learn), the pipeline is not released and the data are available only "on request." Note too: the first author is a former employee of XSpline S.p.A., whose digital-twin engine is used here, and the work is funded by the Russian Science Foundation.

What it changes

For the research community, the contribution is clear: a way to constrain response prediction to real venous geometry, and a feasibility proof of a digital twin coupled with an explainable classifier for pre-operative planning. The credible next step is mapped out in advance: external multicenter validation, then a trial that actually implants at the model-designated site and measures the response. Without that step, the mapping remains a fine picture with no proof of usefulness.

For clinicians, nothing is deployable today: a research tool, internally validated, unapproved, requiring imaging and simulation outside the usual workflow. The most immediately useful idea may not be to choose a site, but to flag in advance the patients for whom no lead position promises benefit — those who should be steered straight to conduction-system pacing. For patients and the general public, the nuance is the same as for any planning AI: a plausible promise, a careful method, but this is a pilot on 74 people, not a tool validated in the operating room, and the most striking figure ("another position would have worked") describes a simulation, not a successful operation.

To go further

The preprint is available on medRxiv (10.64898/2026.07.01.26356827). On other prognostic models in cardiology and their calibration, see our decryptage of ischemic-stroke risk prediction and its multisite calibration; on the limits of a model validated only internally, that of a random forest versus deep learning on acute pancreatitis; and on what SHAP-explainable scores are really worth, that of an SVM interpreted by SHAP in oncology.

Editorial transparency: French version written and signed by the Tatakoto editorial team based on a reading of the preprint. English, Spanish and Chinese translations produced with AI assistance and reviewed.