Liu, Kan

Background: Risk stratification for suspected pulmonary emboli (PE) remains an area of active research. Current risk stratification algorithms have moderate discriminatory capabilities and are guideline-recommended, but they are poorly utilized in practice, with most clinicians relying predominantly on their clinical training and personal “gestalt.” Our group recently derived an Artificial Intelligence (AI) algorithm using electrocardiograms (ECG) to predict the presence of PE (AUC 0.69) on computed tomography scans of the pulmonary arteries (CTPA) in a cohort of 79K patients. Aim: Validate the AI-ECG algorithm in a contemporary dataset of patients with suspected PE. Methods: Emergency Department (ED) visits between 9/1/2022 and 12/31/2023 across the Mayo Clinic Enterprise were searched using a unified data platform. ED visits in patients over 18 years of age with a discretely documented Wells' PE or PERC Rule, a D-dimer result, or a CTPA were included. Only CTPA imaging and D-dimers performed within 48 hours of ED admission were eligible to be included. Patients were included in the final cohort if an ECG was performed within 6 hours of ED admission, D-dimer testing, or CTPA. Patients were considered negative for PE if CTPA was negative, and when this wasn't performed, a negative D-dimer (high sensitivity) or PERC rule of 0 was considered negative. A previously derived Artificial Neural Network algorithm to analyze the ECG to predict PE was applied to the dataset for validation. Results: A total of 27,028 ED visits in 24,288 unique patients meet inclusion criteria. Among this group, the first ED visit with a CTPA and an ECG within 6 hours was included, leaving 18,936 patients for the final validation cohort. The mean age of the cohort was 54.7 (SD 18.7), and 57% were male. Among the entire cohort, 13,573 underwent D-dimer testing for which 4,895 (36%) were positive at standard thresholds. A total of 10,486 patients underwent CTPA (55% of patients evaluated), of which 455 were positive for PE (4.3%). Using the AI-ECG model from the derivation cohort applied to this validation cohort resulted in an AUC of 0.69, identical to the derivation cohort. Using ECG model estimates alone, low (0.74%, n=2447, NPV 99.3%), elevated (2.27%, n=15404, NPV, 97.72%), and high-risk (8.02%, n=1085, NPV 91.98%) categories for PE can be created. A negative D-dimer in the elevated risk group increased the negative predictive value (NPV) to >99.9% and could be used to reduce imaging in this group further. Using the proposed ECG to D-dimer cascade risk stratification could further reduce the need for PE imaging for an overall NPV of 99.9% and sensitivity of 95.2%. We then used additional clinical parameters to train a new algorithm with additional clinical data and the AI-ECG algorithm estimates. Multiple machine learning algorithms were then evaluated in a dataset divided into training (2/3) and testing (1/3) datasets, after downsampling (to address class imbalance). The best results were obtained with AI-ECG results and D-dimer: Ada Boost (AUC 0.93, F1 score 0.895) compared to standard logistic regression (AUC 0.79, F1 score 0.746). The addition of age, sex, and oxygen saturation did not improve model performance. Conclusion: We have now validated the previously derived AI-ANN ECG algorithm for PE in a contemporary and independent cohort and demonstrate potential for integration into clinical PE assessment algorithms. Unlike prior risk stratification processes, ECGs are invariably performed in patients with cardiopulmonary symptoms and have the potential to prevent missed or delayed PE diagnosis by proactively identifying high-risk patients. Additionally, we demonstrate that the results of this algorithm can be combined with D-dimer values, resulting in a simple, powerful, predictive algorithm (AUC 0.93) that pairs cardiovascular electrophysiology with laboratory biomarkers and could lead to a new paradigm in clinical risk stratification for PE.

Background: Many medications, electrolyte perturbations, and several diseases can prolong the QTc beyond its 99 th percentile value resulting in acquired QT prolongation (AQTP). In contrast, approximately 1 in 2,000 people have congenital long QT syndrome (LQTS) hallmarked by pathological QT prolongation. Recently, we developed an AI-ECG model to distinguish these two groups, however this model does not elucidate which ECG segments distinguish LQTS from AQTP. Methods: Using the cohort from our previous study, all patients with LQTS evaluated in the Mayo Clinic Genetic Heart Rhythm Clinic and controls were included. Next, every patient/control with ≥1 ECG above age- and sex-specific 99 th percentile for QTc was selected and matched at 1:100 ratio. Subsequently, 5 ML algorithms (logistic regression, Naïve Bayes, random forest, Gradient boosting and XGBoost) were trained using 52 features representing segments of the 12-lead ECG (e.g. wave peak times, amplitudes, areas under wave, duration; 52x12 features). Training and testing sets were spilt at 80:20 ratio and 5-fold cross validation was performed to prevent overfitting. Following analysis, the top 100 features were re-analyzed to refine results. Results: Among >1,600 patients with LQTS, 690 had ≥ 1 ECG with a QTc above the established threshold compared to 28,186 controls. Following age- and sex-matching and splitting, 22,410 (training), and 5,776 (testing) ECGs were used. Of the 5 ML algorithms, XGBoost demonstrated best performance (AUC 0.913, accuracy 0.836, sensitivity 0.857, specificity 0.835) to distinguish LQTS from AQTP. Following filtering for the top 100 features, performance of the algorithm remained high with AUC of 0.912 (accuracy 0.84, sensitivity 0.841, specificity 0.84). The features with the greatest impact on classification output included time from T wave onset to T wave peak (leads V1, V4, V4), QRS interval duration (leads V1, V3, V4, aVR), and amplitude at end of ST segment (aVF). Conclusions: Distinguishing monogenetically driven LQTS from multi-factorial AQTP is critical in clinical practice for proper management, mitigation of offending QT-prolonging factors, and treatment. Building on our previously developed AI-ECG-LQTS ‘mutation detector’, this model identifies some of the elements within the ‘black box’ that distinguish an ECG and QTc stemming from a patient with LQTS compared to a similar prolonged QTc value that is arising due to non-genetic, acquired QT-prolonging factors.

Abstract Background Existing clinical preoperative risk assessment tools such as the Revised Cardiac Risk Index (RCRI) have been used extensively to predict perioperative major adverse cardiovascular events (MACE). Electrocardiogram (ECG) analysis using Deep Learning can identify hidden features that human eyes cannot identify and might help further risk stratify patients undergoing noncardiac surgery. Objective We aimed to develop a Deep Learning model that predicts perioperative MACE in patients undergoing elective noncardiac surgeries. Methods We included patients ≥ 18-year-old who underwent elective non-cardiac surgery between 2000-2021. MACE was validated in duplicate and were defined as a composite of in-hospital myocardial infarction, cardiac arrest or mortality. A convolutional neural network was developed using ECGs within 30 days before surgery. We randomly split subjects into training, internal validation, and testing datasets in a 7:1:2 ratio. Performance was evaluated using the area under the receiver operator characteristics curve (AUC) values on the testing dataset. Models trained included AI-ECG models to predict in-hospital MACE and to predict in-hospital mortality, and AI-ECG+RCRI models to predict in-hospital MACE and to predict in-hospital mortality. All models were validated using 30-day validated outcomes (MACE or all-cause mortality) in a community-based cohort and were compared with the Revised Cardiac Risk Index (RCRI) score using AUC values. Findings: We included 195,214 patients who underwent a total of 241,999 surgeries. 169,214 ECGs were used in the training dataset, 24,234 in the internal validation dataset, and 48,551 in the testing dataset. In the test cohort, the AI-ECG algorithm discriminated risk for in-hospital MACE with an AUC of 0.82 (95% CI 0.79-0.85). The algorithm similarly discriminated 30-day MACE with an AUC value of 0.79 (95% CI 0.77-0.81) surpassing the discrimination of the RCRI score AUC of 0.69 (95% CI 0.66-0.72). All models’ performances are shown in Figure1A. The AI-ECG model to predict in-hospital mortality had an AUC of 0.83 (95% CI 0.80-0.86) and predicted 30-day mortality with an AUC of 0.80 (95% CI 0.77-0.82). Other models’ performances are shown Figure1B. Training the AI-ECG models including RCRI data did not improve the AI-ECG models’ performance. Conclusion AI-ECG models with no additional clinical data can improve perioperative risk prediction in non-cardiac surgery, outperforming a conventional and widely used risk stratification tool such as the RCRI.