Machine learning predicted index of microvascular resistance with 99% variance explained and coronary flow reserve with 68% variance explained in unobstructed coronaries.
Does a machine learning model using resting angiographic and physiological data accurately predict invasive indices of microvascular function in patients without obstructive coronary artery disease?
Machine learning models using resting angiographic and physiological parameters can accurately predict invasive indices of microvascular function, potentially simplifying diagnostics.
Abstract Background Invasive coronary physiology is the reference standard in the assessment of microvascular function. This is done using either combined thermistor/pressure or Doppler/pressure diagnostic guidewires, to assess coronary physiology during rest and hyperaemia. Wire-free techniques to assess coronary physiology represent accurate, safe and potentially time-efficient alternatives. The cost of these technologies represent a barrier to more wide-spread adoption in clinical practice. This study investigates the potential complimentary role of machine learning and predictive modelling in coronary microvascular assessment in patients without obstructive coronary artery disease. Methods Retrospective analysis of 300 patients undergoing invasive microvascular function testing using a combined temperature/pressure diagnostic guidewire. Cases were identified from existing studies in our centre, including patients with angiographically unobstructed coronary arteries who had provided written informed consent for pooled data analysis. A blinded assessment of resting angiographic TIMI frame count was combined with invasive physiology measurements in a training dataset of 300 unique records. A feature scaled linear regression model predicted coronary transit times from TIMI frame counts at rest. This was combined with a standardised multi-layer perceptron neural network regression (MLPR) model to predict non-linear parameters including hyperaemic response for distal coronary pressures and transit times. Cross validation to establish the optimal hyperparameters was undertaken for neural network models prior to analysis. Results Resting proximal aortic pressure (Pa) predicted hyperaemic distal coronary pressure (Pd) in unobstructed coronaries using MLPR with R2 0.75, mean absolute error (MAE) 4.26 mmHg, root mean square error (RMSE) 5.70 mmHg. TIMI frame count predicted coronary transit time at rest using linear regression with R2 0.33, MAE 0.32, RMSE 0.39. The predicted resting coronary transit time, training data resting transit times and TIMI frame count using MLPR predicted hyperaemic transit time with R2 0.45, MAE 0.09, RMSE 0.12. The combined machine learning model demonstrated excellent performance in predicting the index of microvascular resistance (IMR) and coronary flow reserve with 99% and 68% of the variance explained respectively. Conclusions These results demonstrate that a machine learning model is feasible for the assessment of coronary microvascular function and potentially enhances our understanding of coronary physiology.Model development Model performance summary
Sykes et al. (2025) studied this question. Machine learning predicted index of microvascular resistance with 99% variance explained and coronary flow reserve with 68% variance explained in unobstructed coronaries.