Predicting functional results of percutaneous coronary intervention using machine learning modelling

Comprehensive Summary

Fezzi et al. developed and internally validated machine-learning models to predict post-procedural coronal physiology, measured as Murray’s law-based quantitative flow ratio (μFR), using only pre-PCI clinical, angiographic, and physiological data. The study pooled two prospective European cohorts, including 343 vessels from 291 patients undergoing clinically indicated PCI. Four machine-learning approaches were trained and compared using 1,000 bootstrap iterations, with model selection based on lowest root mean square error for continuous prediction. The selected model accurately predicted post-PCI μFR (RMSE: 0.036, MAE: 0.030, MAPE: 3.2%), despite limited correlation between pre- and post-PCI μFR values. Predicted μFR values were then used to classify PCI outcomes as optimal (μFR ≥ 0.91) or sub-optimal, achieving moderate discrimination (accuracy: 0.72, AUC: 0.72) with high sensitivity (0.90) but low specificity (0.29). Physiological disease pattern and vessel geometry, particularly focal disease and larger reference vessel diameter, were strongly associated with higher post-PCI μFR.

Outcomes and Implications

This proof-of-concept study suggests that machine-learning models can reliably estimate post-PCI coronary physiology before intervention, supporting the concept of “virtual PCI” planning. The high sensitivity of predicted μFR for identifying physiologically optimal outcomes may be useful for upfront procedural decision-making, helping operators anticipate suboptimal results and tailor preparation or treatment strategy accordingly. But the modest specificity reflects the limitations of dichotomozing a continuous physiological variable near a fixed threshold, especially in cohorts with typically high post-PCI μFR values. The lack of external validation, limited outcome events, and narrow post-PCI μFR values. Further prospective studies are needed to determine whether pre-procedural μFR prediction improves procedural strategy, reduces residual ischemia, or translates into better long-term patient outcomes.