Deep Learning in Prediction of Late Major Bleeding After Transcatheter Aortic Valve Replacement

Abstract

Late major bleeding is one of the main complications after transcatheter aortic valve replacement (TAVR). We aimed to develop a risk prediction model based on deep learning to predict major or life-threatening bleeding complications (MLBCs) after TAVR. Patients and Methods: This was a retrospective study including TAVR patients from West China Hospital of Sichuan University Transcatheter Aortic Valve Replacement Registry (ChiCTR2000033419) between April 17, 2012 and May 27, 2020. A deep learning-based model named BLeNet was developed with 56 features covering baseline, procedural, and post-procedural characteristics. The model was validated with the bootstrap method and evaluated using Harrell's concordance index (c-index), receiver operating characteristics (ROC) curve, calibration curve, and Kaplan-Meier estimate. Captum interpretation library was applied to identify feature importance. The BLeNet model was compared with the traditional Cox proportional hazard (Cox-PH) model and the random survival forest model in the metrics mentioned above. Results: The BLeNet model outperformed the Cox-PH and random survival forest models significantly in discrimination [optimism-corrected c-index of BLeNet vs Cox-PH vs random survival forest: 0.81 (95% CI: 0.79-0.92) vs 0.72 (95% CI: 0.63-0.77) vs 0.70 (95% CI: 0.61-0.74)] and calibration (integrated calibration index of BLeNet vs Cox-PH vs random survival forest: 0.007 vs 0.015 vs 0.019). In Kaplan-Meier analysis, BLeNet model had great performance in stratifying high-and low-bleeding risk patients (p < 0.0001). Conclusion:Deep learning is a feasible way to build prediction models concerning TAVR prognosis. A dedicated bleeding risk prediction model was developed for TAVR patients to facilitate well-informed clinical decisions.

Recent advancement in integrating artificial intelligence and information technology with real‐world data for clinical decision‐making in China: A scoping review

“…Three models were derived from existing population-based cohorts, which included the DATADRYAD dataset, 44 the China Kadoorie Biobank, 36 the dataset built by a prospective cohort study. 45 For clinical EMRs, single-center EMRs were adopted by seven studies, 37,41,42,47,[50][51][52] and another seven studies used multicenter EMRs. [38][39][40]46,48,49,53 One additional study leveraged disease registry, the China Acute Myocardial Infarction registry.…”

“…The retrospective use of RWD was the predominate method of data collection. [36][37][38][39][40][41][42][43][44][45][46][47][48][49][50][51][52][53] Similar to diagnostic tools, sample sizes varied by the types of RWD sources. For tools derived from population-based cohorts and registry, the sample sizes ranged from 15k 44 to 503k.…”

“…52 Evidential reasoning rule and focused-CNN with Boruta algorithm were used respectively to select features of ICU admission and in-hospital mortality of trauma patients 50 and patients who were at risk of progression to severe COVID-19. 49 Furthermore, two algorithms were developed specifically to address time-to-event data, including one deep learning technique, BLeNet, 41 and one supervised learning technique, XGBoost-Surv. 46…”

“…Some studies have highlighted the importance of model interpretability for the purpose of real-world clinical application. Captum interpretation library 41 and SHapley Additive exPlanations (SHAP) tool 37,42,46 were two commonly adopted tools for clinical interpretability. It should be stressed that a trade-off between predictive ability and clinical explainability was suggested by Li et al 43 in choosing the optimal in-hospital mortality prediction model for patients with ST-elevation myocardial infarction.…”

“…The retrospective use of RWD was the predominate method of data collection 36–53 . Similar to diagnostic tools, sample sizes varied by the types of RWD sources.…”