ML models integrating clinical, demographic, and CYP2C19 genetic information can identify patients at high ischemic risk after PCI, with a support vector machine achieving the best external performance (AUC 0.667; sensitivity 0.871; specificity 0.282) and XGBoost providing a more balanced profile (AUC 0.619; sensitivity 0.442; specificity 0.688).
Key Findings
Results
A support vector machine with polynomial kernel achieved the best external validation performance for predicting 1-year ischemic outcomes after PCI.
The SVM polynomial kernel model achieved AUC 0.667, sensitivity 0.871, and specificity 0.282 in external validation on the Precision PCI registry.
The outcome was a composite of cardiovascular death, myocardial infarction, stroke, and stent thrombosis at 1 year.
Models were trained on the TAILOR-PCI trial (n=4572) and externally validated on the Precision PCI registry (n=3745).
Cross-validation and synthetic minority oversampling (SMOTE) were used to address class imbalance given the rarity of ischemic events.
Results
XGBoost provided a more balanced sensitivity-specificity profile compared to the SVM polynomial model in external validation.
XGBoost achieved AUC 0.619, sensitivity 0.442, and specificity 0.688 in external validation.
The SVM polynomial model favored high sensitivity (0.871) at the cost of low specificity (0.282), while XGBoost offered a more balanced trade-off.
Both models were assessed by area under the receiver operating characteristic curve (AUC), sensitivity, and specificity.
Methods
Boruta feature selection identified 11 predictors from clinical, demographic, and CYP2C19 pharmacogenetic data for model training.
Variable importance for the SVM polynomial model demonstrated that all 11 Boruta feature-selected predictors had relatively high importance (>75).
Predictors integrated clinical, demographic, and CYP2C19 genetic information, representing a pharmacogenetic-inclusive approach not previously incorporated in existing ML risk scores.
Current machine learning model-developed risk scores do not incorporate pharmacogenetic data, making this a novel methodological contribution.
Methods
The study analyzed a total of 8317 patients across two datasets to develop and externally validate the ML models.
The TAILOR-PCI trial contributed n=4572 patients used for model training.
The Precision PCI registry contributed n=3745 patients used for external validation.
The combined dataset represents both a large clinical trial and a real-world registry population.
Background
Ischemic events after contemporary PCI are uncommon but carry high morbidity and mortality, making risk stratification clinically important.
The composite ischemic outcome (cardiovascular death, myocardial infarction, stroke, and stent thrombosis) is described as rare, necessitating use of SMOTE to address class imbalance.
The authors note that 'ischemic events are rare, and most patients may safely de-escalate DAPT to reduce bleeding risk.'
Identifying high-risk patients is described as 'critical to guide dual antiplatelet therapy (DAPT) intensity while minimizing bleeding.'
Conclusions
The ML models were proposed as a tool to support a multimodal approach to risk stratification to guide DAPT de-escalation decisions.
The authors state that 'ML models trained on large clinical trial and real-world registry datasets can help identify the small subset of patients at high ischemic risk after PCI.'
The clinical relevance is framed around maintaining ischemic protection 'through a multimodal approach to risk stratification.'
The models are intended to refine clinical decision making regarding DAPT intensity post-PCI.
What This Means
This research developed and tested computer-based (machine learning) models that combine patients' genetic information — specifically a gene called CYP2C19 that affects how the blood thinner clopidogrel works — with clinical and demographic data to predict which patients are most likely to have a serious heart-related event (such as heart attack, stroke, or death) within one year after a procedure to open blocked arteries (percutaneous coronary intervention, or PCI). The study used data from over 8,000 patients across a large clinical trial and a real-world patient registry. The best-performing model (a support vector machine) was very good at catching patients who would go on to have a bad outcome (high sensitivity of 87%), though it also flagged many patients who ultimately did not have such events (low specificity of 28%). A different model (XGBoost) offered a more balanced trade-off between catching true high-risk cases and avoiding false alarms.
This research suggests that incorporating genetic data alongside traditional clinical information into machine learning models can help doctors identify the small group of patients who face the highest risk of serious complications after artery-opening procedures. This distinction matters because these serious events are relatively rare — meaning most patients are at low risk and could potentially reduce their blood-thinning medication regimen (dual antiplatelet therapy, or DAPT) to lower their chances of bleeding, while truly high-risk patients might benefit from more intensive treatment.
The practical implication is that tools like these, once further validated, could support more personalized decisions about how long and at what intensity patients should remain on blood-thinning medications after PCI. Rather than a one-size-fits-all approach, integrating genetic and clinical data through machine learning may allow clinicians to better tailor therapy — protecting high-risk patients from heart events while sparing lower-risk patients from unnecessary bleeding risk.
Grant C, Ingraham B, Lennon R, Raina A, Tian S, Kowlgi G, et al.. (2026). Machine Learning With Genetic and Clinical Data to Predict Ischemic Outcomes After PCI.. Clinical and translational science. https://doi.org/10.1111/cts.70655