SPEED-TR: a self-distilled and pre-trained transformer model for enhanced ECG detection of tricuspid regurgitation

Figures

• 

  Fig. 1. ROC curves and PRC curves for model development datasets and multi-center evaluation datasets. This figure presents the ROC curves and PRC curves with 95% CIs in the A training set, B validation set, C hold-out testing set, D FW testing set, E YF testing set and F SF testing set, respectively. AUROC area under the receiver operating characteristic curve, AUPRC area under the precision-recall curve, CI confidence interval, FW Fuwai Hospital, PRC the precision-recall curve, ROC the receiver operating characteristic curve, SF Shenzhen Fuwai Hospital, YF Yunnan Fuwai Hospital.

• 

  Fig. 2. Distribution of prediction probabilities of SPEED-TR in different subgroups of TR severity and accuracy of SPEED-TR in grading TR severity across different datasets. The figure illustrates the distribution of the predicted probabilities of the SPEED-TR model on different subgroups of varying TR severity via scatter plots and box plots, in A hold-out testing set, B FW testing set, C YF testing set, D SF testing set and E overall multi-center testing set, respectively; and F the accuracy of SPEED-TR in grading TR severity across the five testing sets, based on the predicted probability for TR detection with the thresholds of 0.008 for mild, 0.255 for moderate and 0.755 for severe TR. FW Fuwai Hospital, SF Shenzhen Fuwai Hospital, SPEED-TR the self-distilled and pre-trained Transformer model for detecting tricuspid regurgitation, TR tricuspid regurgitation, YF Yunnan Fuwai Hospital.

• 

  Fig. 3. ROC curves of SPEED-TR performance in different subgroups on the overall multi-center testing set. A ROC curves of SPEED-TR in patients with different TR risk factors, including PH, AF/AFL, age ≥ 60 years, female, LVEF < 50% and MR, and patients characterized by differing numbers (0, 1, 2, 3 and over 3) of these risk factors; and B ROC curves of SPEED-TR across patients presenting with other typical VHDs, including AS, AR, MS or MR, and with 0, 1, 2 or over 2 types of these VHDs on the overall multi-center testing set. AF atrial fibrillation, AFL atrial flutter, AR aortic regurgitation, AS aortic stenosis, AUROC area under the receiver operating characteristic curve, LVEF left ventricular ejection fraction, MR mitral valve regurgitation, MS mitral valve stenosis, PH pulmonary hypertension, ROC the receiver operating characteristic curve, SPEED-TR the self-distilled and pre-trained Transformer model for detecting tricuspid regurgitation, TR tricuspid regurgitation, VHD valvular heart diseases.

• 

  Fig. 4. Flowchart illustrating patient inclusion and dataset distribution. The flowchart depicts data inclusion and exclusion criteria, as well as the division methods for different datasets. A The flowchart for model development datasets, including pre-training, training, validation, and hold-out testing sets. B The flowchart for multi-center evaluation datasets, including FW, YF and SF testing sets, these three datasets constitute the multi-center testing set. Dec., December; ECG, electrocardiogram; ECHO echocardiography, FW Fuwai Hospital, Jan. January, SF Shenzhen Fuwai Hospital, YF Yunnan Fuwai Hospital.

• 

  Fig. 5. Flowchart for model development and evaluation. Model development comprised a pre-training stage on the pre-training set, followed by a post-training stage using the training set. During pre-training, a self-supervised learning method was employed to extract of intrinsic ECG patterns. The model adopted an encoder-decoder network, tasked with reconstructing the masked ECG segments. In post-training stage, a self-distillation strategy with a teacher-student setup using identical architectures was applied to enhance model performance. Specifically, the pre-trained encoder was transformed into a TR detection model, serving as the architecture for both the teacher and student models, with the pretrained weights providing initialization. After the self-distillation, the student model achieving the highest validation AUPRC was selected as the final SPEED-TR model. Performance evaluation was conducted on the hold-out testing set and FW testing set from Fuwai Hospital, YF testing set from Yunnan Fuwai Hospital and SF testing set from Shenzhen Fuwai Hospital. Comprehensive analysis results of SPEED-TR for TR recognition were presented, including not only overall performance, but also…