InsighTx: A Machine-Learning Model That Accurately Predicts Transplant Outcomes During Ex Vivo Lung Perfusion

Abstract

<h3>Purpose</h3> Artificial Intelligence (AI) continues to have a transformative impact on clinical decision-making; however, it has not yet been thoroughly investigated for use during ex vivo lung perfusion (EVLP). In this study we sought to develop and validate a machine learning model, that we termed <i>InsighTx</i>, that predicts transplant outcomes following EVLP. <h3>Methods</h3> A total of n=602 clinical EVLP cases were used to train and test an Extreme Gradient Boosting (XGBoost) model derived from donor features along with biological, physiological, and biochemical assessments made during EVLP. For each feature that was assessed longitudinally during EVLP, a minimum, maximum, and trend value was extracted for a total of n=112 input features to the InsighTx model. EVLPs performed from 2008-2019 (n=505) were used to train the model to predict one of three-outcome classifications: (i) lungs unsuitable for transplantation, or EVLP transplants resulting in a time to extubation of (ii) ≥72 hours or (iii) <72 hours. 5-fold cross validation was used to derive model parameters in the training dataset. A separate dataset of EVLP cases conducted from 2019-2020 was used to test the InsighTx model (n=97). <h3>Results</h3> The InsighTx model had an area under the receiver operating characteristic curve (AUROC) of 79% [76-82%] and 75% [75-76%] in the training and test datasets respectively. The model performed extremely well in predicting lungs that were unsuitable for transplantation (AUROC: 90% [86-94%] training; 88% [88-89%] test) and in EVLP transplant recipients that were extubated <72-hours post-transplant (AUROC: 80% [76-84%] training; 76% [75-76%] test). Physiological assessments were important features driving the prediction of short time to extubation and lung suitability, whereas biological and biochemical features were important predictors of prolonged ventilation. <h3>Conclusion</h3> Herein we describe a novel, large-scale study of a machine learning approach to biomarker prediction models during clinical EVLP. The InsighTx model reports a three-outcome classification for EVLP which more accurately reflects the clinical situation. The model was tested in the largest EVLP study conducted to date and showed excellent predictive performance. The InsighTx model represents a cutting-edge approach to use AI to augment clinical decision-making for transplant surgeons using EVLP.