ARTIFICIAL INTELLIGENCE IN PERCUTANEOUS CORONARY INTERVENTION: IMPROVED PREDICTION OF PCI-RELATED COMPLICATIONS USING AN ARTIFICIAL NEURAL NETWORK

Importance: Complications after percutaneous coronary intervention (PCI) are common and costly. Risk models for predicting the likelihood of acute kidney injury (AKI), bleeding, stroke and death are limited by accuracy and inability to use non-linear relationships among predictors. Additionally, if non-linear relationships among predictors can be leveraged, then the prediction of any adverse event (i.e. the patient who will not do well with PCI) is perhaps of greater interest to clinicians than prediction of adverse events in isolation. Objective: To develop and validate a set of artificial neural networks (ANN) models to predict five adverse outcomes after PCI - AKI, bleeding, stroke, death and one or more of these four (any adverse outcome). Design: Cross-sectional study, using institutional NCDR CathPCI data. Setting and participants: 28,005 patients undergoing PCI at five hospitals in the Barnes-Jewish Hospital system. Main Outcome(s): AKI, bleeding, stroke, death, and one or more of these four (any adverse outcome). We divided 28,005 PCI patients into a training cohort of 21,004 (75%) and a test cohort of 7,001 (25%). We used an artificial neural network (ANN) multilayer perceptron (MLP) model to predict each outcome based on a set of 278 encoded and preprocessed variables. Model accuracy was tested using area under the receiver-operating-characteristic curve (AUC). Performance and validation of the MLP model was compared with existing regression models using integrated discrimination improvement (IDI) and continuous net reclassification index (NRI). Results: The prevalence of AKI, bleeding, stroke and death in the study cohort was 4.6%, 3.6%, 0.3% and 1.1%, respectively. The fully trained MLP model achieved convergence quickly (<10 epochs) and could predict accurately predict AKI (77.9%), bleeding (86.5%), death (90.3%) and any adverse outcome (80.6%) in the independent test set. However, prediction of stroke was not satisfactory (69.9%). Compared to the currently used models for AKI, bleeding and death prediction, our models showed a significantly higher AUC (range 1.6% - 5.6%), IDI (range 4.9% - 7.2%) and NRI (range 0.07 - 0.61). Conclusions and Relevance: By using neural network-based models, we accurately predict major adverse events after PCI. Larger studies for replicability and longitudinal studies for evidence of impact are needed to establish these artificial intelligence methods in current PCI practice. Objective: To develop and validate a set of artificial neural networks (ANN) models to predict five adverse outcomes after PCI - AKI, bleeding, stroke, death and one or more of these four (any adverse outcome). Design: Cross-sectional study, using institutional NCDR CathPCI data. Setting and participants: 28,005 patients undergoing PCI at five hospitals in the Barnes-Jewish Hospital system. Main Outcome(s): AKI, bleeding, stroke, death, and one or more of these four (any adverse outcome). We divided 28,005 PCI patients into a training cohort of 21,004 (75%) and a test cohort of 7,001 (25%). We used an artificial neural network (ANN) multilayer perceptron (MLP) model to predict each outcome based on a set of 278 encoded and preprocessed variables. Model accuracy was tested using area under the receiver-operating-characteristic curve (AUC). Performance and validation of the MLP model was compared with existing regression models using integrated discrimination improvement (IDI) and continuous net reclassification index (NRI). Results: The prevalence of AKI, bleeding, stroke and death in the study cohort was 4.6%, 3.6%, 0.3% and 1.1%, respectively. The fully trained MLP model achieved convergence quickly (<10 epochs) and could predict accurately predict AKI (77.9%), bleeding (86.5%), death (90.3%) and any adverse outcome (80.6%) in the independent test set. However, prediction of stroke was not satisfactory (69.9%). Compared to the currently used models for AKI, bleeding and death prediction, our models showed a significantly higher AUC (range 1.6% - 5.6%), IDI (range 4.9% - 7.2%) and NRI (range 0.07 - 0.61). Conclusions and Relevance: By using neural network-based models, we accurately predict major adverse events after PCI. Larger studies for replicability and longitudinal studies for evidence of impact are needed to establish these artificial intelligence methods in current PCI practice.