Using AI to Inform TAVR and Reduce Post-Procedure Pacemaker Rates
Authors: Amina H Khalpey, PhD, Kirtana Roopan, Ujjawal Kumar BA & Zain Khalpey, MD, PhD, FACS
Introduction
Transcatheter aortic valve replacement (TAVR) has emerged as a minimally invasive alternative to surgical aortic valve replacement to treat severe aortic stenosis. The procedure involves implanting a prosthetic valve via a catheter, avoiding the need for open-heart surgery. Although TAVR has shown promising results in terms of reduced morbidity and mortality, there still are complications that can occur, one of which is the need for placement of a permanent pacemaker (PPM). Risk stratification tools have been steadily developed to better inform cardiologists about what risk factors are most important to consider when exploring TAVR as a treatment option for patients. Machine learning (ML) algorithms and artificial intelligence have the potential to transform the prediction of complications due to TAVR (24).
AI Algorithms and Predictive Analytics
Simple logistic regression is currently the standard model for prediction in medicine and is therefore most commonly used for post-TAVR PPM prediction. However, newer, more accurate models are being explored and have shown great promise. One such model type is random forest. This algorithm is composed of multiple decision trees that each include a random amount and combination of variables from the dataset. The algorithm then uses a culmination of results from all the trees to make a prediction. Gradient boosting is a form of machine learning used for regression that utilizes a series of decision trees that are scaled and adjusted based on the error of the previous tree. This process continues until additional trees no longer add to the accuracy of the model prediction. XG Boost is a regularized form of gradient boosting that can be generalized to more datasets (1). While XG Boost is more generalizable, it requires a higher n number than gradient boosting does. There are also even more complex models like artificial neural networks, which are made up of layers of nodes rather than the simple nodes found in a decision tree. There is an input and an output layer, sandwiched between which lies several hidden layers that are essential to the processing of the data in order to create an accurate output. Exploring the efficacy of all of these types of algorithms in PPM prediction after TAVR will need to assess how well each one can incorporate the large, diverse, and heterogeneous datasets needed for optimal performance.
Predicting permanent pacemaker placement after TAVR requires consideration of several variables, including parameters from echocardiography, computerized tomography (CT), and electrocardiography along with multiple risk factors. One challenge that AI prediction models face is being able to take into account all of these factors in an unbiased fashion that is generalizable across many datasets.
TAVR Risk Assessment and Stratification
There are a wide variety of factors that lead to conduction abnormalities resulting in PPM placement after TAVR. Some of the risk factors include history of hypertension, diabetes, chronic kidney disease, coronary artery disease, myocardial infarction, prior cardiac surgery, and peripheral artery disease. There are also pre-procedural measurements that can aid in the prediction process, like PR and QRS intervals from electrocardiogram, to identify underlying atrioventricular blocks. Different types of TAVR valves can also influence the risk of permanent pacemaker implantation. One study compared different types of TAVR valves and their association with pacemaker implantation. The study concluded that the type of valve plays a crucial role in determining the need for pacemaker implantation post-TAVR (2).
Additionally, there are pieces of information vital to assessing PPM risk in CT scans and echocardiography, like ejection fraction, mean pressure gradient, and degree of calcification. One study used a combination of factors, including valve seating, to predict the need for a permanent pacemaker with good accuracy, which emphasized the importance of valve positioning in the overall risk assessment (3). A study by De Torres-Alba et al. found that TAVR valve depth in relation to the aortic annulus was significantly associated with the risk of new-onset conduction abnormalities requiring pacemaker implantation. The study found that deeper implantation of the valve was linked to higher rates of pacemaker requirement, as the lower seating increases proximity to the heart’s conduction system (4). Similarly, prosthesis oversizing can cause similar issues and conduction disturbances that lead to the implantation of a permanent pacemaker (5).
Valve oversizing is a key factor in predicting PPM after TAVR (6). ML algorithms can analyze pre-procedural imaging data like computed tomography and echocardiography to predict the optimal valve size and type for each patient or predict whether the patient will need a PPM after the procedure. AI-powered software coupled with new-technology can identify comorbidities during preoperative evaluation of heart function. Gohmann et al analyzed the performance of an ML-based CT-fractional flow reserve (CT-FFR), which was used to evaluate TAVR patients for co-morbid CAD with reasonable success. High-performance image analysis like these opens numerous doors for enhancing complication prediction for TAVR.
The ability of ML algorithms to analyze large amounts of patient data to identify pre-procedural risk factors for complications can aid in the selection of appropriate candidates for TAVR. Evertz et al trained an AI-powered function to analyze ejection fraction and other echocardiographic evidence to pre-operatively stratify patients for TAVR, showing equivalent stratification to manual assessment of heart function (7). Further AI-driven tools such as this one can stratify patients based on their preoperative risk profiles, enabling clinicians to make more informed decisions about treatment options (24).
Current ML Models for Predicting Permanent Pacemaker (PPM) After TAVR
A review of the literature on current models for PPM prediction post-TAVR found eight logistic regressions (14, 1, 15, 16, 17, 18, 19, 8), three random forests (1, 8, 9), one XGBoost (Pandey, one random forest gradient boosting (10), one gradient boosting (11), one Naive Bayes (1), one custom-built neural network (12), and one Levenberg-Marquadt algorithm (19).
Random forest (RF) was used in one study to predict risk of permanent pacemaker implantation after TAVR. Using data from 555 patients, the study compared the performance of an RF algorithm as compared to a traditional logistic regression model. They found that including post-TAVR ECG data led to more accurate prediction, and that the RF algorithm more accurately predicted PPM risk than the logistic regression (8).
One study that performed a gradient boosting machine learning (GBM) prediction model included 163 variables and was optimized using 5-fold cross-validation repeated 10-times (Agasthi). The receiver operator characteristic (ROC) for the GBM model was calculated to predict one-year mortality post TAVR, and then compared to the TAVI2-SCORE and CoreValve score. Based on GBM, the ten most predictive variables for one-year survival were cardiac power index, hemoglobin, systolic blood pressure, INR, diastolic blood pressure, body mass index, valve calcium score, serum creatinine, aortic annulus area, and albumin. The study found the GBM model to be very accurate in predicting one-year mortality post-TAVR, with an area under the curve of 0.72. The GBM model proved to be more accurate than the standard regression models like TAVI2-SCORE and CoreValve Score (10, 24).
Another study evaluated one year morbidity and mortality prediction with external validation of an XGBoost model. The AI-driven remote monitoring systems can track patient data (e.g., vital signs, valve function) and alert clinicians to potential complications (21). Machine learning by extreme gradient boosting was trained and tested with repeated 10-fold hold-out testing using 34 pre- and 25 periprocedural clinical variables. The XGBoost model includes an internal function that can identify the patient features that contributed to a specific prediction, allowing for more transparency between the AI and clinician (21).
ML Algorithms can give insight on which pre-operative parameters can contribute to postoperative complications and presence of these parameters could affect the approach a surgeon suggests. Many of the current models include relevant risk factors and certain anatomical measurements like annulus diameter. But there seems to be a lack of emphasis on several important factors, for instance prosthesis oversizing and implantation depth. Additionally, machine learning image analysis of sources like CT scans and angiograms has the potential to take AI prediction to even greater heights. A multimodal model that considers all of the risk factors, clinical measurements, and images that can contribute to prediction can open doors for predicting PPM implantation after TAVR. AI tools such as Chat GPT and Google’s Med-PaLM 2 have begun the journey on medical image analysis, but they are still a long way from accurate medical image analysis. The need for accurately analyzing and incorporating such diverse, heterogeneous data types poses a challenge for building AI prediction models, but it is already in the works.
Challenges and Future Directions
The shift from emphasis on the individual clinical experience to a “collective medical mind” based on data and published literature is presenting new ways to incorporate learning from accumulated data with artificial intelligence.
AI tools’ accuracy and reliability depends on the quality and standardization of the data used to train the algorithm. To accomplish this, it is vital to have collaboration between those who build and those who use these models. Medical professionals, researchers, and AI developers must work together to establish standardized data collection and reporting methods. Implementing robust data curation processes can help minimize the impact of data discrepancies on the performance of ML algorithms (24).
While incorporating new technologies is expected in health care, AI-enabled technologies possess fundamental characteristics that set them apart from other innovations, characteristics that have the power to change the doctor-patient relationship in both positive and negative ways. The use of AI in medical decision-making raises ethical concerns around algorithmic bias, transparency, and accountability.
Additionally, ensuring patient privacy and data security is vital in the development and implementation of AI-driven TAVR tools. Regulatory frameworks and guidelines should be established to address these concerns and promote the responsible use of AI in TAVR.
Conclusions
AI is slowly being recognized as a powerful tool for analysis and prediction in healthcare. There are already several machine learning models for predicting PPM after TAVR that include a wide range of parameters from TAVR risk factors to ECG and echocardiographic measurements utilizing algorithms like random forest and XGBoost. Building on these developments and enhancing them to create a multi stackable model that can not only analyze numerical data but also images will provide a pathway for enhancing patient outcomes and minimizing complications.
References
1. Tsushima T, Al-Kindi S, Nadeem F, Attizzani GF, Elgudin Y, Markowitz A, Costa MA, Simon DI, Arruda MS, Mackall JA, Thal SG. Machine Learning Algorithms for Prediction of Permanent Pacemaker Implantation After Transcatheter Aortic Valve Replacement. Circ Arrhythm Electrophysiol. 2021 Mar;14(3):e008941. doi: 10.1161/CIRCEP.120.008941. Epub 2021 Mar 9. PMID: 33685208.
2. Levack MM, Kapadia SR, Soltesz EG, Gillinov AM, Houghtaling PL, Navia JL, Krishnaswamy A, Blackstone EH, Svensson LG, Mick SL. Prevalence of and Risk Factors for Permanent Pacemaker Implantation After Aortic Valve Replacement. Ann Thorac Surg. 2019 Sep;108(3):700-707. doi: 10.1016/j.athoracsur.2019.03.056. Epub 2019 Apr 26. PMID: 31034827.
3. Hamdan, A., et al. (2015). Inverse relationship between membranous septal length and the risk of atrioventricular block in patients undergoing transcatheter aortic valve implantation. *JACC: Cardiovascular Interventions*, 8(9), 1218-1228.
4. De Torres-Alba, F., et al. (2018). Changes in the pacemaker rate after transition from Edwards SAPIEN XT to SAPIEN 3 transcatheter aortic valve implantation: The critical role of valve implantation height. *JACC: Cardiovascular Interventions*, 11(18), 1815-1823.
5. Tsoi M, Tandon K, Zimetbaum PJ, Frishman WH. Conduction Disturbances and Permanent Pacemaker Implantation After Transcatheter Aortic Valve Replacement: Predictors and Prevention. Cardiol Rev. 2022 Jul-Aug 01;30(4):179-187. doi: 10.1097/CRD.0000000000000398. Epub 2021 Jun 14. PMID: 34132653.
6. Sheth T, Natarajan MK, Kreatsoulas C, Whitlock R, Parry D, Chu V, Smith A, Velianou JL. Avoiding S3 Valve Over-Sizing by Deployment Balloon Over-Filling: Impact on Rates of Permanent Pacemaker and Other Procedural Complications During TAVR. J Invasive Cardiol. 2018 Jan;30(1):23-27. PMID: 29289947.
7. Evertz R, Lange T, Backhaus SJ, et al. Artificial Intelligence Enabled Fully Automated CMR Function Quantification for Optimized Risk Stratification in Patients Undergoing Transcatheter Aortic Valve Replacement. J Interv Cardiol. 2022;2022:1368878. Published 2022 Apr 20.
8. Truong VT, Beyerbach D, Mazur W, Wigle M, Bateman E, Pallerla A, Ngo TNM, Shreenivas S, Tretter JT, Palmer C, Kereiakes DJ, Chung ES. Machine learning method for predicting pacemaker implantation following transcatheter aortic valve replacement. Pacing Clin Electrophysiol. 2021 Feb;44(2):334-340. doi: 10.1111/pace.14163. Epub 2021 Jan 28. PMID: 33433905.
9. Truong V, Wigle M, Bateman E, et al. PACEMAKER IMPLANTATION FOLLOWING TAVR: USING MACHINE LEARNING TO OPTIMIZE RISK STRATIFICATION. J Am Coll Cardiol. 2020 Mar, 75 (11_Supplement_1) 1478. https://doi.org/10.1016/S0735-1097(20)32105-7
10. Agasthi P, Ashraf H, Pujari SH, Girardo M, Tseng A, Mookadam F, Venepally N, Buras MR, Abraham B, Khetarpal BK, Allam M, Md SKM, Eleid MF, Greason KL, Beohar N, Sweeney J, Fortuin D, Holmes DRJ, Arsanjani R. Prediction of permanent pacemaker implantation after transcatheter aortic valve replacement: The role of machine learning. World J Cardiol. 2023 Mar 26;15(3):95-105. doi: 10.4330/wjc.v15.i3.95. PMID: 37033682; PMCID: PMC10074998.
11. Agasthi, P., Mookadam, F., Venepally, N., Girardo, M., Buras, M., Khetarpal, B. K., … & Arsanjani, R. (2019). Machine learning helps predict permanent pacemaker requirement post transcatheter aortic valve replacement. Circulation, 140(Suppl_1), A15572-A15572.
12. Krittanawong C, Hassan H, Inayat F, et al. CAN WE PREDICT THE PACEMAKER IMPLANTATION AFTER TRANSCATHETER AORTIC VALVE REPLACEMENT USING DEEP LEARNING? AN EXPERIENCE FROM URBAN HEALTHCARE SYSTEM. J Am Coll Cardiol. 2021 May, 77 (18_Supplement_1) 3271. https://doi.org/10.1016/S0735-1097(21)04625-8
13. Gohmann, R. F., et al. (2022). Combined cCTA and TAVR Planning for Ruling Out Significant CAD: Added Value of ML-Based CT-FFR. JACC Cardiovasc Imaging, 15(3), 476-486.
14. Maeno Y, Abramowitz Y, Kawamori H, Kazuno Y, Kubo S, Takahashi N, Mangat G, Okuyama K, Kashif M, Chakravarty T, Nakamura M, Cheng W, Friedman J, Berman D, Makkar RR, Jilaihawi H. A Highly Predictive Risk Model for Pacemaker Implantation After TAVR. JACC Cardiovasc Imaging. 2017 Oct;10(10 Pt A):1139-1147. doi: 10.1016/j.jcmg.2016.11.020. Epub 2017 Apr 12. PMID: 28412434.
15. Barrett, Christopher and Nickel, Andrew and Rosenberg, Michael and Ream, Karen and Tzou, Wendy and Aleong, Ryan G. and Tumolo, Alexis and Garg, Lohit and Zipse, Matthew and West, John J. and Varosy, Paul and Sandhu, Amneet, PRIME Score for Prediction of Permanent Pacemaker Implantation After Transcatheter Aortic Valve Replacement. Available at SSRN: https://ssrn.com/abstract=4100676 or http://dx.doi.org/10.2139/ssrn.4100676
16. Tsushima T, Nadeem F, Al-Kindi S, et al. Risk Prediction Model for Cardiac Implantable Electronic Device Implantation After Transcatheter Aortic Valve Replacement. J Am Coll Cardiol EP. 2020 Mar, 6 (3) 295–303. https://doi.org/10.1016/j.jacep.2019.10.020
17. Vejpongsa, P., Zhang, X., Bhise, V., et al., Risk Prediction Model for Permanent Pacemaker Implantation after Transcatheter Aortic Valve Replacement. 2018 Jul, 328-335. https://doi.org/10.1080/24748706.2018.1467067
18. Wheat, H., Tshusima, T., Al-Kindi, S., Nadeem, F., Attizzani, G., Elgudin, Y., … & Thal, S. (2020). Risk Prediction Model for Cardiac Implantable Electronic Device Implantation After Transcatheter Aortic Valve Replacement in Patients with Pre-Existing Atrial Fibrillation. Authorea Preprints. 19. Parulkar, A. B., Sheikh, W., Ahmed, M. B., Singal, S., Parulkar, S. S., Lima, F., … & Chu, A. (2021). PREDICTING PERMANENT PACEMAKER IMPLANTATION (PPMI) IN PATIENTS UNDERGOING TRANSCATHETER AORTIC VALVE REPLACEMENT (TAVR) UTILIZING DEEP LEARNING TECHNIQUES. Journal of the American College of Cardiology, 77(18_Supplement_1), 413-413.
20. Hermans MC, Van Mourik MS, Hermens HJ, Baan J Jr, Vis MM. Remote Monitoring of Patients Undergoing Transcatheter Aortic Valve Replacement: A Framework for Postprocedural Telemonitoring. JMIR Cardio. 2018;2(1):e9.
21. Kwiecinski J et al. Machine learning for prediction of all-cause mortality after transcatheter aortic valve implantation, European Heart Journal – Quality of Care and Clinical Outcomes, 2023
22. Char DS, Shah NH, Magnus D. Implementing Machine Learning in Health Care – Addressing Ethical Challenges. N Engl J Med. 2018;378(11):981-983.
23. Hayashidani, S., Shiose, A., & Tsutsui, H. (2020, May 25). New-onset left bundle branch block after transcatheter aortic valve implantation ― not a harmless bystander ―. Circulation Journal. https://www.jstage.jst.go.jp/article/circj/84/6/84_CJ-20-0292/_html/-char/en
24. Khalpey, Z., Roopan, K., Deckwa, J., Wilson, P., Burke, R., Riley, R., Rizik, D. Machine Learning Algorithms And AI: Augmenting Safety And Predicting Complications In Transcatheter Aortic Valve Replacement. Khalpey AI Lab.