A transformer-based artificial intelligence model generated individualized estimates of ischemic and bleeding risk at six intervals during the first year after percutaneous coronary intervention and generally outperformed two established deep learning survival models in internal and external testing, according to a retrospective study published in npj Digital Medicine. 

The model, called Transformer-DAPT, used structured electronic health record (EHR) data to estimate a patient's risk at one, two, three, six, nine, and 12 months after percutaneous coronary intervention (PCI), rather than producing a single, fixed-risk estimate.  

Although the model was developed for patients undergoing PCI rather than endocrine disease, it illustrates how artificial intelligence (AI) can use structured EHR data to generate individualized risk estimates across multiple follow-up intervals rather than a single fixed prediction. For endocrinologists, the study provides an example of how longitudinal AI models could eventually support repeated risk assessment during ongoing management of chronic diseases, in which treatment decisions are revisited as patient risk evolves.  

Dual antiplatelet therapy (DAPT) reduces ischemic complications after PCI, but longer treatment can increase bleeding risk. Existing instruments address specific decisions: PRECISE-DAPT estimates bleeding risk at discharge, whereas the DAPT score was developed to help assess whether therapy should be extended beyond 12 months. Neither provides simultaneous ischemic and bleeding estimates at multiple points during the first year, according to the investigators.  

The retrospective study used EHR data from Mayo Clinic to develop the Transformer-DAPT model and the OneFlorida+ Clinical Research Consortium to perform external validation. The Mayo cohort included 29,032 adults who underwent PCI with drug-eluting stent implantation between 2002 and 2024 and received DAPT after the index procedure. External validation identified 19,173 eligible patients from the OneFlorida+ network between 2012 and 2020 using the same inclusion and exclusion criteria. Model performance was evaluated in a randomly selected external test set comprising 20% (3,835 patients) of this cohort. 

The investigators randomly divided the Mayo cohort into training (70%), validation (10%), and testing (20%) datasets using stratified sampling to preserve event rates. 

The primary outcomes were ischemic and bleeding events occurring within one year after PCI. The ischemic endpoint comprised a composite of acute ischemic heart disease—including acute myocardial infarction, unstable angina or intermediate coronary syndromes, and cardiac arrest—plus ischemic stroke or transient ischemic attack, repeat drug-eluting stent implantation after the index PCI, and coronary artery bypass grafting. Bleeding events included spontaneous bleeding and blood transfusion. 

The investigators compared Transformer-DAPT with two established deep learning survival models, DeepSurv and DeepHit. Model discrimination was assessed using the time-dependent concordance index (Ctd-index), with calibration evaluated separately using isotonic regression. 

Among the 29,032 patients in the Mayo cohort, 10,054 (35%) experienced ischemic events and 4,077 (14%) experienced bleeding events during the first year after PCI. 

For ischemic events, Transformer-DAPT achieved a Ctd-index of 0.84 at one month. Its performance was significantly better than DeepHit, which had a Ctd-index of 0.80, but was comparable to DeepSurv, which had a Ctd-index of 0.83. From two through 12 months, Transformer-DAPT significantly outperformed both benchmark models. Its Ctd-index was 0.86 at two months, 0.87 at three and six months, 0.86 at nine months, and 0.85 at 12 months. 

For bleeding events, Transformer-DAPT achieved a one-month Ctd-index of 0.88, compared with 0.76 for DeepSurv and 0.84 for DeepHit. The model significantly outperformed DeepSurv but was not significantly different from DeepHit at one month. From two months onward, Transformer-DAPT significantly outperformed both models, with Ctd-indices decreasing from 0.87 at two, three, and six months to 0.84 at nine months and 0.81 at 12 months.  

In a secondary 12-month binary classification analysis in the Mayo cohort, the model achieved an AUROC of 0.89 for ischemic event prediction and 0.81 for bleeding event prediction.  

Performance was lower in the OneFlorida+ external test set, but Transformer-DAPT maintained higher Ctd-indices than the benchmark models at most intervals. For ischemic events, Transformer-DAPT achieved a one-month Ctd-index of 0.84, compared with 0.72 for DeepSurv and 0.75 for DeepHit. Neither first-month comparison reached statistical significance. From two through 12 months, Transformer-DAPT significantly outperformed both models, with Ctd-indices ranging from 0.77 at two, three, and six months to 0.75 at nine months and 0.74 at 12 months. 

For bleeding events, the model achieved a one-month Ctd-index of 0.83, compared with 0.79 for DeepSurv and 0.80 for DeepHit; neither difference was statistically significant. From two months onward, Transformer-DAPT significantly outperformed both benchmark models. Its Ctd-index was 0.79 at two, three, and six months, 0.76 at nine months, and 0.75 at 12 months. 

In the Mayo test set, the investigators also evaluated whether predicted event probabilities aligned with observed outcomes. After isotonic regression calibration, agreement between predicted and observed 12-month event rates improved substantially for both ischemic and bleeding outcomes. 

For ischemic events, the expected calibration error (ECE) improved from 0.050 before calibration to 0.016 afterward. Bleeding predictions initially demonstrated greater miscalibration, particularly at higher predicted risk levels, with an ECE of 0.188 that improved to 0.011 following calibration. Time-dependent calibration also showed closer agreement between model-predicted survival curves and Kaplan-Meier estimates throughout the first year after PCI. 

The study had several limitations. The investigators relied on structured electronic health record data that lacked detailed information on stent characteristics, bleeding severity, and precise medication start and stop dates. Medication prescriptions documented in the electronic health record also could not confirm patient adherence.  

Outcome ascertainment was limited to encounters captured within participating health systems, and the ischemic composite endpoint was broader than those commonly reported in clinical trials because it included unstable angina, transient ischemic attack, cardiac arrest, and repeat coronary revascularization in addition to myocardial infarction and ischemic stroke. 

The authors also acknowledged potential temporal changes in PCI practice over the study period, the high exclusion rate in the external validation cohort because of incomplete data, and suboptimal calibration of bleeding predictions at the highest predicted risk levels. Finally, the study did not prospectively evaluate whether using Transformer-DAPT to guide treatment decisions improved patient outcomes. 

No conflicts of interest were reported.  

(Editor's Note: The study is an unedited article-in-press version that will undergo further editing before final publication and may contain errors that affect the content. All legal disclaimers apply.)