Can ML Help To Predict Long Term Mortality In TAVI Patients?
November 1, 2021
Transcatheter aortic valve implantation (TAVI) is used for aortic valve stenosis treatment in high-risk patients and has now been extended to intermediate-risk patients. With the procedure still having an elevated mortality rate at five years, researchers from Italy’s Centro Cardiologico Monzino, University of Milan, and Politecnico di Milano set out to develop a new machine learning (ML) method that could identify the best predictors of five-year mortality after TAVI.
The study used an ML approach to ascertain the best performing predictors from numerous clinical and echocardiographic variables, with the aim of improving long-term clinical prognosis.
TAVI long-term mortality prediction: Addressing the unknown
Since 2002, TAVI has proved effective in improving symptoms and survival in patients with symptomatic severe aortic valve stenosis (AS) who are considered to be at high surgical risk. However, despite being recognized as the gold standard of treatment for these high-risk patients, the all-cause mortality rate ranges from 6.7% to 14.5% at one year, up to around 47% at five years.
With TAVI long-term mortality prediction still being unknown, the evaluation of mortality predictors was described by the study as of “utmost importance for patient selection, risk stratification, tailoring therapy, and correctly informing the patient about long-term prognosis after the procedure”.
After noting the promise shown by ML solutions in various medical fields, the study theorized that “learning algorithms may allow predictive features undetected by conventional statistical methods to be discovered in order to improve risk definition and prognosis after TAVI procedure”.
A novel risk prediction approach
The researchers used an ML model to develop a new risk prediction approach for predicting the mortality rate at five-year follow-up after TAVI.
A study population was chosen from patients with severe symptomatic AS who underwent TAVI at Centro Cardiologico Monzino IRCCS between 2008 and 2014. Baseline patient data that was analyzed included echocardiographic data, laboratory results, and diagnosis, as well as clinical status and symptoms.
Some 83 pre-TAVI variables were considered for each patient, along with descriptive statistics. The study evaluated three widely used supervised classification ML algorithms - random forest (RF), extreme gradient boosting (XGBoost), and multilayer perceptron (MLP) - using different classifiers to predict the occurrence of all-cause of death at five-year mortality after TAVI. A logistic regression (LR) model was also implemented.
The dataset was pre-processed for data optimization and consistency ahead of the analysis, and different feature selection models were evaluated, with feature selection being performed for each algorithm using the least absolute shrinkage and selection operator (LASSO), gradient-boosting machine (GBM), Boruta, and RF.
The evaluation of ML performances on the testing set used the area under the receiver-operating curve (AUC), along with the computing of additional metrics including accuracy, sensitivity, positive predictive value (PPV), and F1-score, as well as a comparison with the EuroSCORE II, TAVI’s most used score.
The major relevant predictors of the study outcome for the best ML model were determined using the permutation feature importance (PFI) algorithm, which measures the association of individual variables with model accuracy.
The application of recursive feature elimination (RFE) identified 14 pre-treatment variables as the most relevant mortality predictors in TAVI patients at five-year follow-up. These are:
- MR etiology
- Stroke volume index
- Interventricular septal thickness
- Left atrium area
- Aortic valve area
- Mean aortic pressure gradient
- Alanine aminotransferase
- International normalized ratio
- EuroSCORE II
Researchers presented their two main findings as follows: The MLP model achieved the best AUC (0.79) in predicting mortality five years after TAVI. And that novel features, never considered in premature mortality risk scores in TAVI patients, had been identified, including the association of low stroke volume (SV) with higher mortality.
While the findings did not demonstrate a significant disparity in AUC between the ML models of MLP and LR to estimate mortality five years after TAVI, MLP showed the better predictive abilities of the two.
Among the limitations highlighted by the researchers were:
- The size of the dataset, which may affect the performance of the model
- The research being a single-center study
- The dataset only including high and intermediate-risk patients, therefore not offering the opportunity to extrapolate our results to lower-risk cases
The results of the study were discussed with expert medical cardiologists, and clinical explanations were reported.
The future for ML models in prediction of 5-year mortality after TAVI
This study has shown that ML may help clinicians in integrating a multitude of predictors in the assessment of risk factors for long-term mortality after TAVI and that this, in turn, can improve clinical decision-making and prognosis. ML-based tools can potentially be key to the identification of new predictors, through the discovery of “unexpected variables and interactions”.
Even using ML techniques, the study has shown that predicting five-year mortality after TAVI is challenging. Compared to previous risk scores used, the new approach to long-term mortality prediction in TAVI patients which the study presented was based on a wider range of analytic methods variables. This allowed new variables to be highlighted as potentially influencing long-term prognosis.
As ML algorithms continue to improve and data continues to be digitized, it is likely that the relationship between ML and medicine will strengthen further. In years to come, ML models could have an important role in the evaluation of long-term mortality risk after TAVI, offering a way to incorporate various pieces of information to accurately represent clinical scenarios being investigated. This, the study suggests, could pave the way for an enhanced evaluation of treatment options and improve the selection process for intermediate and low-risk patients.
Indications for TAVI are expanding, and it is hoped that these findings can support clinicians in assessing prognosis after TAVI, paving the way for more accurate patient information on the outcome of the procedure.
cvi42: the future of cardiovascular imaging for TAVI patients
Circle CVI’s most innovative and recent software, cvi42, is capable of evaluating, analyzing and assessing access of the aortic valve for TAVI procedures. The latest software was developed for providing fast and efficient planning of structural heart interventions.
Leading the way in cardiac imaging solutions, cvi42 interventional planning capabilities extends to providing comprehensive overviews of structures during the septal approach, as well as providing a fast and complete overview of the vascular access.
To try the latest software in cardiac imaging, why not download your free cvi42 trial.