Procedural success prediction in chronic total occlusion percutaneous coronary intervention (CTO-PCI)—the rise of the machines?

The field of chronic total occlusion percutaneous coronary intervention (CTO-PCI) has experienced a significant evolution driven by advancements in technology, refinements of procedural techniques, the use of specialized algorithms, and increased operator expertise (1-4). Traditionally, operators have relied on crossing algorithms based on limited angiographic features and expert opinions, which, while useful, often lack the accuracy and adaptability needed for consistent success. Crossing the occlusion with a guidewire is paramount. Characteristics such as cap morphology ambiguity, calcification, occlusion length, and vessel tortuosity significantly affect the likelihood of successful crossing. An artificial intelligence (AI) system could incorporate many more variables for the prediction of procedural success, and further more assist operators in the decision-making process by suggesting strategies and selection of material (5).

In this issue of JACC: Cardiovascular Interventions, investigators represented by Rempakos et al. (6) developed a machine learning (ML) model with 14 features and a high predictive capacity for successful primary antegrade wiring (AW) in CTO-PCI. The authors meticulously developed and validated five ML models, identifying extreme gradient boosting (XGBoost) as the most effective one, with an impressive average area under the receiver-operating characteristic curve (AUC) of 0.775. By applying hyperparameter tuning, a process of optimizing specific internal settings of the model (such as learning rate) to improve performance, the researchers fine-tuned XGBoost’s performance, achieving an AUC of 0.780 in the testing set, underscoring its robustness in predicting AW success.

AI refers to the development of computer systems that can perform tasks typically requiring human intelligence, such as learning, reasoning, problem-solving, and decision-making. AI aims to mimic cognitive functions, allowing machines to interpret data, identify patterns, and make informed decisions based on this information.

ML is a subset of AI that focuses on creating algorithms that enable computers to learn from data without being explicitly programmed. Instead of following predefined instructions, ML models identify patterns within large datasets, learn from those patterns, and make predictions or decisions based on new inputs.

In the context of this study, the ML model was trained on historical data from previous CTO-PCI procedures, including variables like patient characteristics, lesion anatomy, and procedural factors. The model learned to associate specific combinations of these variables with successful or unsuccessful outcomes (in this case, whether the guidewire successfully crosses the CTO lesion). As the model is exposed to more data, it becomes better at making accurate predictions.

XGBoost, the ML algorithm used in this study, is highly effective for structured data, as it builds a series of decision trees sequentially to correct previous errors, incorporating regularization to prevent overfitting, parallelization for fast processing, and automatic handling of missing data, making it versatile for tasks like classification, regression, and ranking, and a popular choice in both research and industry due to its speed and effectiveness, especially on large datasets. This model can assist interventional cardiologists by predicting procedural success based on a patient’s specific clinical and anatomical characteristics, ultimately enhancing decision-making during complex CTO-PCI procedures.

The study’s dataset included variables such as occlusion length, proximal cap ambiguity, calcifications, tortuosity, and a prior failed attempt, all of which are components of the J-CTO score (7). Additional factors included in-stent restenosis, the presence of a side branch at the proximal cap, bifurcation at the distal cap, a poor distal landing zone, interventional collateral channels, blunt/no stump, the target vessel, and parameters from the PROGRESS-CTO score (8), as well as aorto-ostial lesion. These comprehensive inputs enabled the ML models to predict outcomes more accurately than traditional scoring systems like J-CTO, CASTLE, and PROGRESS-CTO, which had lower AUCs of 0.678, 0.629, and 0.563, respectively.

To address one of the common limitations of ML models—interpretability—the SHAP (SHapley Additive exPlanations) method was used. SHAP helps to clarify the impact of each variable on the model’s predictions, providing actionable insights. Factors like occlusion length, proximal cap ambiguity, and vessel diameter emerged as significant parameters for predicting the success of anterograde CTO wiring, aligning with known aspects in CTO-PCI field. Interestingly, the presence of interventional collateral channels—a factor not traditionally emphasized in the antegrade approach—was also identified as an important predictor for successful AW. This finding could highlight a limitation of using retrospective data in ML models, as it may reflect operator preferences and feasibility considerations rather than purely unbiased predictions. However, it could also suggest that improved visualization of the distal coronary bed might facilitate the successful crossing of the lesion.

ML systems, particularly neural networks mimic the human brain’s architecture consisting of interconnected layers of nodes or “neurons”, which process input data through weighted connections. Each neuron receives an input, processes it through an activation function, and passes the output to a subsequent layer of neurons. During training, the system adjusts the weights based on the error of its predictions compared to actual outcomes, a process known as backpropagation. Over numerous iterations, the model learns to recognize complex patterns and make accurate predictions. This method allows the neural network to process a diverse range of input variables, making it highly effective for predicting successful lesion crossing in CTO-PCI.

A test of the efficacy of the ML model was previously conducted, involving a substantial number of patients from an international registry who underwent CTO-PCI (6). The findings indicated that while the model performed well, the patients had been treated according to existing operator algorithms, which already account for various anatomical characteristics. Therefore, the efficacy of ML in a direct, prospective setting—where it could independently influence treatment strategies—remains untested. Despite the clear advantages, ML models are not without drawbacks. Increased computational time and the risk of overfitting are notable concerns. However, the study’s use of a well-validated XGBoost model and the SHAP method for interpretability represents a significant stride toward mitigating these issues.

Currently, the ML model focuses primarily on anatomical features, which, while important, do not offer a complete picture of procedural success. Incorporating operator-specific factors—such as experience, case volume, and complication rates—into the ML model, could significantly enhance its predictive accuracy and clinical relevance. However, achieving this will require large datasets that represent a wide range of operator profiles.

Looking forward, AI in this domain could benefit from advanced methods of data acquisition, storage, and management, capable of handling complex datasets. Innovations like blockchain technology or decentralized storage systems (9) could provide secure, scalable, and efficient ways to store and share operator-specific procedural data. These systems could ensure continuous data collection, updates, and sharing while maintaining data integrity and privacy. By integrating real-time procedural data from individual operators, AI systems could evolve into personalized, real-time assistants, not only predicting success rates but also adapting to the unique characteristics of each operator.

Operators bring varied expertise, decision-making styles, and strategic preferences to CTO-PCI. Experienced operators often achieve higher success rates in complex cases due to refined techniques and superior problem-solving skills, while less experienced operators may encounter higher complication rates, even with less challenging anatomies. By integrating data on each operator’s historical performance, including success rates for more advanced techniques like anterograde dissection re-entry (ADR) or retrograde procedures, ML models could offer more personalized predictions.

To further enhance the accuracy of AI-based predictions, incorporating additional non-invasive and invasive imaging modalities could provide more detailed and granular information. For instance, pre-procedural coronary computed tomography angiography (CTA) has been shown to significantly improve procedural success rates and reduce periprocedural complications compared to angiography alone. By offering precise visualization of the coronary anatomy, CTA allows for better planning and stratification in complex interventions (10,11).

Similarly, intravascular ultrasound offers superior imaging of vessel size and plaque morphology, which are critical for optimal results, and integrating these additional layers of anatomical information could greatly enhance such an AI system (12).

Ensuring the models are user-friendly and integrating them into clinical workflows will be crucial for widespread adoption. Artificial prediction of technical success can certainly help in the procedural planning of CTO-PCI, but a significant drawback of relying on ML systems is the risk that operators may become overly dependent on these tools. This reliance could lead to less personal involvement in analyzing and understanding angiographic details and anatomy, which in return may increase periprocedural complications and hamper success.

The current study’s findings represent a noteworthy advancement in CTO-PCI, highlighting the promising role of ML models in refining and enhancing the accuracy of lesion-crossing success predictions. Integrating comprehensive anatomical lesion characteristics and employing sophisticated analytical techniques offers a more precise, multi-modality data-driven foundation for clinical decision-making. As ML continues to evolve, its application in interventional cardiology is poised to refine and revolutionize patient care, ushering in an era of personalized, efficient, and highly effective treatment strategies.

Although far from complete, the journey from mere expert opinion to data-driven precision ML models could be a potentially helpful direction in the CTO-PCI field. Implementation of these new technologies in the next few years may ultimately allow clinicians to effectively individualize and apply optimal treatment strategies in patients with a CTO, particularly those undergoing PCI.

Future prediction models could incorporate AI technologies exceeding ML and add automation for CTO-PCI strategy suggestions and outcome prediction on an individual level.

At the heart of this evolution lies the quest for data-driven precision and automation, much like the transition depicted in Terminator 3, where machines gradually evolve, gaining autonomy and intelligence. However, as in the film, the integration of such advanced systems requires vigilance to ensure they enhance human expertise rather than replace it, maintaining a balanced synergy between man and machine.

Acknowledgments

Funding: None.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Cardiovascular Diagnosis and Therapy. The article has undergone external peer review.

Peer Review File: Available at https://cdt.amegroups.com/article/view/10.21037/cdt-24-392/prf

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://cdt.amegroups.com/article/view/10.21037/cdt-24-392/coif). C.U. reports grants or contracts, consulting fees, and support for travel from Biotronik, Abbott, Boston, and Top MEDICAL; royalties or licenses, and honoraria for lectures from Cordis. G.L. reports grants from Biotronik, Boston Scientific; support for attending meetings and/or travel by Terumo and Schockwave; consulting fees from Biotronik, Schockwave, Cordis, and Orbus Neich; payment or honoraria for lectures, presentations, or educational events by BBraun, OrbusNeich, Schockwave, Terumo; support for attending meetings and/or travel by Terumo, Schockwave. The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Brilakis ES, Mashayekhi K, Tsuchikane E, et al. Guiding Principles for Chronic Total Occlusion Percutaneous Coronary Intervention. Circulation 2019;140:420-33. [Crossref] [PubMed]
2. Othman H, Seth M, Zein R, et al. Percutaneous Coronary Intervention for Chronic Total Occlusion-The Michigan Experience: Insights From the BMC2 Registry. JACC Cardiovasc Interv 2020;13:1357-68. [Crossref] [PubMed]
3. Werner GS, Hildick-Smith D, Martin Yuste V, et al. Three-year outcomes of A Randomized Multicentre Trial Comparing Revascularization and Optimal Medical Therapy for Chronic Total Coronary Occlusions (EuroCTO). EuroIntervention 2023;19:571-9. [Crossref] [PubMed]
4. Juricic SA, Stojkovic SM, Galassi AR, et al. Long-term follow-up of patients with chronic total coronary artery occlusion previously randomized to treatment with optimal drug therapy or percutaneous revascularization of chronic total occlusion (COMET-CTO). Front Cardiovasc Med 2023;9:1014664. [Crossref] [PubMed]
5. Nakachi T, Yamane M, Kishi K, et al. Machine Learning for Prediction of Technical Results of Percutaneous Coronary Intervention for Chronic Total Occlusion. J Clin Med 2023;12:3354. [Crossref] [PubMed]
6. Rempakos A, Alexandrou M, Mutlu D, et al. Predicting Successful Chronic Total Occlusion Crossing With Primary Antegrade Wiring Using Machine Learning. JACC Cardiovasc Interv 2024;17:1707-16. [Crossref] [PubMed]
7. Morino Y, Abe M, Morimoto T, et al. Predicting successful guidewire crossing through chronic total occlusion of native coronary lesions within 30 minutes: the J-CTO (Multicenter CTO Registry in Japan) score as a difficulty grading and time assessment tool. JACC Cardiovasc Interv 2011;4:213-21. [Crossref] [PubMed]
8. Christopoulos G, Kandzari DE, Yeh RW, et al. Development and Validation of a Novel Scoring System for Predicting Technical Success of Chronic Total Occlusion Percutaneous Coronary Interventions: The PROGRESS CTO (Prospective Global Registry for the Study of Chronic Total Occlusion Intervention) Score. JACC Cardiovasc Interv 2016;9:1-9. [Crossref] [PubMed]
9. Shamsan Saleh AM. Blockchain for secure and decentralized artificial intelligence in cybersecurity: A comprehensive review. Blockchain Research and Applications 2024;5:100193. [Crossref]
10. Panuccio G, Werner GS, De Rosa S, et al. Full-Moon Coronary Calcification as Detected With Computed Tomography Angiography in Chronic Total Occlusion Percutaneous Coronary Intervention. Am J Cardiol 2024;222:149-56. [Crossref] [PubMed]
11. Hong SJ, Kim BK, Cho I, et al. Effect of Coronary CTA on Chronic Total Occlusion Percutaneous Coronary Intervention: A Randomized Trial. JACC Cardiovasc Imaging 2021;14:1993-2004. [Crossref] [PubMed]
12. Panuccio G, Abdelwahed YS, Carabetta N, et al. Clinical and Procedural Outcomes of IVUS-Guided vs. Angiography-Guided CTO-PCI: A Systematic Review and Meta-Analysis. J Clin Med 2023;12:4947. [Crossref] [PubMed]