For patients with severe aortic stenosis, replacing the aortic valve can improve more than the obstruction that prompted the procedure. Some people also experience an improvement in mitral regurgitation, a separate valve problem in which blood leaks backwards through the mitral valve. The difficulty for clinicians is that this secondary improvement is far from guaranteed.
Research published in Scientific Reports on 1 October 2026 suggests that a relatively small set of routinely available preoperative measurements may help identify which patients are more likely to see their mitral regurgitation improve after transcatheter aortic valve implantation, or TAVI. The study compared 11 machine-learning approaches, but one of its most notable findings was that conventional logistic regression performed as well as, or slightly better than, more complex algorithms when tested on a later group of patients.
Why improvement is difficult to predict
Aortic stenosis restricts blood leaving the heart because the aortic valve has become narrowed. TAVI treats the problem by placing a replacement valve through a catheter rather than through conventional open-heart valve surgery. Mitral regurgitation can coexist with aortic stenosis, but what happens to it after TAVI varies substantially from one patient to another.
That uncertainty matters because improvement could influence how clinicians think about the need for additional mitral-valve treatment. A prediction made before TAVI could therefore contribute to risk stratification and treatment planning, provided that the model is sufficiently reliable and validated beyond the population in which it was developed.
Yazheng Shan and colleagues studied patients with moderate-to-severe mitral regurgitation who underwent TAVI at a single centre. Their development cohort contained 324 patients treated between 2019 and 2024. Instead of relying only on a random split of those patients to assess performance, the researchers also assembled a temporal validation cohort of 120 patients treated in 2025. In total, the analysis therefore covered 444 patients.
The later cohort is an important feature of the design. Temporal validation asks a model trained on earlier patients to make predictions in patients treated later, which more closely resembles the direction in which a clinical prediction tool would ultimately be used. It is still not the same as testing the model at a different hospital, however, because both cohorts came from the same centre.
Eleven algorithms were put to the test
The researchers compared 11 machine-learning algorithms. To narrow the available clinical information to a more stable set of predictors, they used two feature-selection techniques: the Boruta algorithm and least absolute shrinkage and selection operator, commonly known as LASSO. Variables that survived this dual-screening process formed the basis of the predictive models.
Six factors emerged as robust predictors of whether mitral regurgitation would improve: whether the regurgitation was functional, interventricular septal thickness, atrial fibrillation, left ventricular ejection fraction, the surgical approach and left atrial diameter. These are clinically interpretable variables rather than opaque measurements requiring a specialised experimental test.
Performance was assessed using the area under the receiver operating characteristic curve, or AUC, alongside calibration curves and decision-curve analysis. An AUC of 0.5 indicates discrimination no better than chance, while 1.0 represents perfect separation between outcomes. AUC should not be read as a literal percentage of patients correctly classified, but it provides a standard measure of how well a model ranks patients who experience different outcomes.
The simpler model held up best in later patients
In the 120-patient temporal validation cohort, logistic regression achieved the highest reported AUC at 0.788, with a 95% confidence interval from 0.694 to 0.883. LASSO regression was almost identical at 0.787, with a 95% confidence interval from 0.693 to 0.882. CatBoost followed with an AUC of 0.778 and a 95% confidence interval from 0.683 to 0.872.
The differences between those headline figures are small, and their confidence intervals overlap substantially. The result should therefore not be interpreted as proof that logistic regression is inherently superior to the other approaches. Instead, it shows that a comparatively straightforward statistical model remained competitive after the researchers applied the same selected clinical features to a later patient cohort.
That is useful in a field where greater algorithmic complexity does not automatically translate into greater clinical value. A model that relies on familiar inputs and can be interpreted relatively easily may be more practical to scrutinise, validate and eventually integrate into clinical workflows than a black-box system offering only marginally different discrimination.
Heart-wall thickness showed a nonlinear pattern
The analysis also used SHapley Additive exPlanations, or SHAP, to examine how individual predictors contributed to the models’ predictions. One particularly interesting signal involved interventricular septal thickness, the thickness of the muscular wall separating the heart’s two ventricles.
Its relationship with mitral-regurgitation improvement was nonlinear rather than a simple steady increase or decrease. Exploratory analysis suggested a possible breakpoint at approximately 1.1 centimetres. The authors describe this as provisional, which is an important distinction. A threshold discovered retrospectively in one centre should not be treated as a clinical cut-off until it has been reproduced prospectively and in independent populations.
Together, the six predictors also make physiological sense as markers of the condition and structure of the heart before intervention. Left ventricular ejection fraction reflects pumping performance, left atrial diameter captures structural change associated with pressure and volume burden, and atrial fibrillation often accompanies more advanced cardiac remodelling. Functional mitral regurgitation is itself driven by changes in heart geometry rather than a primary structural defect of the mitral valve.
Promising prediction, but not a clinical decision rule yet
The study’s strongest practical feature is also a reason for caution. Both the 324-patient development group and the 120-patient validation group came from the same centre. A later cohort can reveal whether performance survives changes over time, but it cannot establish that the same relationships will hold across hospitals with different patient populations, imaging practices, clinical protocols or procedural techniques.
The retrospective design creates further limits. Researchers can analyse the information recorded during routine care, but they cannot control data collection in the same way as a prospective study designed around a prediction model. The sample is also modest by machine-learning standards, particularly when 11 algorithms are being evaluated.
The authors accordingly frame the results as preliminary and call for prospective multicentre validation before clinical implementation. That qualification is especially important when a model could influence decisions about invasive cardiac treatment. Discrimination approaching an AUC of 0.8 is encouraging, but it does not eliminate false-positive and false-negative predictions, nor does it demonstrate that using the model improves patient outcomes.
What the research does provide is a focused candidate model built from information clinicians can obtain before TAVI. If independent studies reproduce the result, prediction of mitral-regurgitation improvement may not require an exceptionally complex artificial-intelligence system. In this dataset, six readily available characteristics and a conventional logistic model were enough to produce the strongest temporal-validation result among the approaches tested.
Source Information
Study Title: Machine learning prediction of mitral regurgitation improvement after TAVI
Authors: Yazheng Shan, Peng Liu, Zirui Zhao, Hanzhe Wang, Junlin Lai, Xiang Qiu, Xiaojing Li, Shijie Wang and Yin Wang
Journal: Scientific Reports
Year: 2026
DOI: 10.1038/s41598-026-74084-4
