Primary mitral regurgitation and risk models for predicting outcome: time to move to artificial intelligence?

This editorial refers to ‘Data-driven mortality risk prediction of severe degenerative mitral regurgitation patients undergoing mitral valve surgery’, by S. Kwak et al., https://doi.org/10.1093/ehjci/jead077. Mitral regurgitation (MR) is the most prevalent valve disease.1 Two distinct entities of MR, which differ in their aetiology, consequences, and prognosis, have been defined. Secondary MR develops in patients with left ventricular (LV) or left atrial (LA) dysfunction.2 Its prevalence is underestimated as in secondary MR; the intensity of the murmur does not correlate with the severity of MR.3 Its management remains a conundrum. The present study focused on primary MR implies valvular lesions and are usually degenerative. The most frequent lesions are present in the leaflets and the chordae, inducing prolapse or flail mitral valve. The indications of intervention apply to patients with severe MR [effective regurgitant orifice area ≥0.4 cm²; regurgitant volume ≥ 60 mL, and a regurgitant fraction (RF) ≥ 50%]. An intervention is indicated if the patient is symptomatic and when the symptoms are related to the valve disease. In asymptomatic patients, surgery should be considered when volume overload and/or dysfunction are present. The cut-off values are LV end-systolic dimension ≥40 mm, LV ejection fraction ≤ 60%, or LA volume ≥60 mL/m².4 The Heart Team is essential for appropriate decisions. When surgery is indicated, mitral repair is far more preferable than valve replacement.5 The percentage of repair depends on the aetiology of MR, the mechanism of dysfunction, the expertise of the surgeon, and the complexity of valvular lesions. The short and long-term prognosis of the patients do not only relate to the efficacy of the repair but also to the age and co-morbidities of the patients. Combined surgery or absence of treatment of tricuspid regurgitation may have a role on prognosis. In high-risk patients, transcatheter edge-to-edge repair may be considered.6 Therefore, risk prediction models for time-to-event outcomes have an important role for clinical decision making. Different methods are available. Frequently, the Cox landmarking method is used, which relies on proportional hazards model. Another approach consists in random survival forests (RSF), which is a machine learning algorithm to predict time-to-event outcome.7 This method can capture complex relationships between the predictors and survival without requiring specification. It has been shown to have higher predictive accuracy than Cox models. However, several limitations of the RSF must also be acknowledged: interpretability, depending on the large number of decision trees; the bias-variance tradeoff towards the training data, which can lead to overfitting; the need of very large sample size; the sensitivity of the model to outliers; and the number of censored observations in the dataset. Not all of these parameters were provided. Ideally, the validation cohort should be external, which was also not the case in the present study. In this issue of the Journal, Kwak et al. used this method to construct a risk model for predicting 3-year all-cause mortality after surgical repair or replacement.8 They included 1521 patients; 70% of the population serve as a training set and the other 30% as a test set. The most important predictors were age, haemoglobin level, mitral valve replacement, glomerular filtration rate, LA dimension, and LV end-systolic diameter. This simple risk model using these six variables predicted outcome with a 3-year C-index of 0.880 in the test set. The risk of mortality increased strongly when LA dimension was >55 mm and LV end-systolic diameter was >45 mm. The prediction persisted when the type of surgery was not included in the analysis. The mortality was low in patients with a score of 0 or 1; in the patients with a mitral valve regurgitation (MVR) score of ≥6 points, the 1-year and 3-year mortalities were 25.0% and 36.5%, respectively. These interesting results compare well with the Mitral International Database (MIDA) score, although being significantly more predictive. The MIDA population included patients with degenerative MR (mainly flail leaflet) submitted to medical management or surgery.9 The MIDA score included 7 parameters: age (3 points), heart failure symptoms (3 points), LA dimension ≥55 mm, LV ejection fraction ≤ 60%, LV end-systolic diameter ≥40 mm, atrial fibrillation and right ventricular systolic pressure ≥50 mmHg. Under medical or surgical management, the MIDA score showed a C = 0.78 in their derivation cohort and a C = 0.81 in the validation cohort. In the study by Kwak et al., LA volume, a better parameter than diameter included in the current Guidelines was not calculated in all study patients. The cut-off value of LV end-systolic dimension is now ≥40 mm, rather than ≥45 mm in the 2021 ESC/EACTS Guidelines.4 However, LV dimension is not always easy to accurately measure as the LV remodelling usually develops in the apical half of the LV. There is not yet a validated cut-off value for LV volumes. A low haemoglobin level can be related to renal dysfunction, inflammation, or a bleeding process. Moreover, bleeding is one of the potential complications after MR intervention and could aggravate a pre-existing low haemoglobin and worsen post-intervention prognosis. Blood transfusion has not been included as a potential risk predictor. Atrial fibrillation (AF) was also not included as a parameter. Preoperative AF has indeed prognostic importance; it is a predictor of long-term mortality and post-operative LV dysfunction.10 In addition, AF is frequently associated with atrial functional MR which is added to the severity of primary MR. In a multivariable Cox model, age >60 years and haemoglobin <12.5 g/dL had the highest values for prediction. Therefore, these two parameters were given 2 points and the other 4 were given one point in the MVR risk model. However, age and haemoglobin are predictors of poor outcome in many clinical settings. Other parameters that have previously shown a significant predictive value should be included in the model: impaired LV function, as indicated by a low ejection fraction or deformation imaging parameters, the various degree of MR, the association with other valvular disease, the presence of comorbidities, such as diabetes and hypertension, which may have a higher risk of complications, and mortality after surgery. The integration of the type of mitral valve surgery performed, such as repair vs. replacement, the surgical volume. Finally, the time from diagnosis to surgery, early intervention vs. late, remains an unanswered question. How should we use this MVR risk score in the daily practice? The authors correctly suggest that patients with a low score of 0 or 1 should be submitted to mitral valve repair, if feasible. However, it could also be an indication if the score is 2, because of an age slightly >60 years. When the MVR score is high, the decisions are more difficult, and the role of the Heart Team is essential. Knowing the high 3-year mortality risk, it is reasonable to prefer a transcatheter edge-to-edge repair, if it is possible, according to echocardiographic data. Such approach should be tested in prospective studies. Both the MIDA and this new MVR scores are prognostically important and should be included in the Heart Team discussions. None declared. No new data were generated or analysed in support of this research.