Machine learning-based survival models integrating quantitative flow ratio to predict in-stent restenosis after percutaneous coronary intervention
H Lu, D Zhang, K Y Chen, T Liu, C LiuAbstract
Background
In-stent restenosis (ISR) remains a significant adverse event following percutaneous coronary intervention (PCI). Quantitative flow ratio (QFR) and radial wall strain (RWS) provide functional and mechanical assessments of coronary lesions. Machine learning (ML) offers advantages in integrating multi-dimensional features to improve the accuracy of ISR risk stratification.
Purpose
To develop and validate ML-based survival models for predicting the risk of ISR following PCI, and to compare their performance.
Methods
Consecutive adult patients who underwent PCI were enrolled. The primary endpoint was angiographically confirmed ISR during follow-up. Clinical, laboratory, lesion characteristics, and procedural variables were recorded. Key variables included pre-stent and post-stent QFR measurements (e.g., QFR, △QFR, △QFR/stent length (mm), maximum percentage diameter stenosis (Max%DS)), the Lu-index (residual QFR/residual length), and QFR and RWS measurements (e.g., distal △QFR, distal Max%DS, distal RWS, distal RWS%) within 5 mm from the stent edges. Eleven machine learning algorithms (including AKE, random survival forest, XGBoost, LightGBM, and DeepSurv) and a Cox proportional hazards (CoxPH) model were employed. The dataset was randomly split into training (70%) and testing (30%) sets. Model performance was evaluated using the concordance index (C-index), integrated Brier score, and time-dependent receiver operating characteristic curves (time-ROC). Time-dependent feature importance plots were used to interpret predictor contributions over time.
Results
A total of 363 patients (mean age 63.5 years, 63.74% male) were included. During a median follow-up of 1.44 years, 131 cases (36.09%) of ISR were observed. Following feature selection, eight key predictors were retained, namely post-stent QFR, △QFR, stent length, Max%DS, distal RWS, distal RWS%, △QFR/stent length (mm), and the Lu-index. In the testing set, the AKE model demonstrated superior discriminative performance (C-index: 0.730, 95% CI: 0.680-0.780) compared to the CoxPH model (C-index: 0.645, 95% CI: 0.595-0.695). The AKE model also achieved a lower integrated Brier score (0.256 vs. 0.274) and a time-dependent AUC of 0.741 (95% CI: 0.639-0.840) at 3 years. Time-dependent analysis identified distal RWS%, stent length, and the Lu-index as the most consistently important predictors for ISR risk.
Conclusion
The AKE machine learning model effectively predicts the risk of ISR in post-PCI patients and outperforms the conventional Cox regression model. This model shows potential as a precise risk stratification tool to support clinical decision-making.C-index and ROCfeature importance and KM curves