Screening for Pulmonary Embolism Using Single-channel Electrocardiogram

The aim of the study: to create and evaluate the diagnostic efficiency of an algorithm for detecting pulmonary embolism using digital analysis of a single-channel ECG using elements of artificial intelligence. It is a prospective, controlled, single-center, observational, non-randomized study. The study is planned to include at least 500 patients 18 years old and older (300 patients in the training sample and 200 patients in the test sample). The study will include all patients requiring exclusion of the diagnosis of acute pulmonary embolism. Patients will be examined according to clinical guidelines to confirm the diagnosis of pulmonary embolism (laboratory, clinical and instrumental). During the course of the study, the authors of the work do not interfere with the scope of the examination, which is caried out on patients in accordance with clinical guidelines. All patients included in the study will undergo ECG recording in standard lead I for 1 minute, followed by spectral analysis of the obtained data, which will be stored at the remote monitoring center of Sechenov University without being linked to the personal data of patients. Single-channel ECG will be recorded using the portable single-lead ECG monitor CardioQvark. It is designed as an iPhone cover. It is registered with the Federal Service for Health Supervision on February 15, 2019. RZN No. 2019/8124. The patient's personal data (last name, first name, patronymic, date of birth, contact information) will not be transferred or taken into account. Each patient is assigned an individual number that is not associated with his/her personal data. Then a spectral analysis of the electrocardiogram will be performed using a continuous wavelet transformation, the principles of which are based on the Fourier transform method. The analysis involves the evaluation of the following parameters (the parameters listed below will be calculated as the median of the tact-cycle):• TpTe - time from peak to end of the T-wave• VAT - time from the beginning of the QRS to the R-peak• QTc - corrected QT interval.• QT / TQ - the ratio of QT length to TQ length (from the end of T to the beginning of the QRS of the next complex).• QRS\_E - the total energy of the QRS wave based on the wavelet transform• T\_E - T-wave total energy based on wavelet transform• TP\_E- energy of the main tooth of the T-wave based on the wavelet transform• BETA, BETA\_S - T-wave asymmetry coefficients (simple and smooth versions)• BAD\_T - flag of T-wave quality (whether expressed in the current lead• QRS\_D1\_ons - energy of the leading edge of the R-wave (based on the "first derivative" wavelet transform)• QRS\_D1\_offs - energy of the trailing edge of the R-wave (based on the "first derivative" wavelet transform)• QRS\_D2 - peak energy of the R-wave (based on the "second derivative" wavelet transform)• QRS\_Ei (i = 1,2,3,4) - QRS-wave energy in 4 frequency ranges (2-4-8-16-32 Hz) based on wavelet transform• T\_Ei (i = 1,2,3,4) - T-wave energy in 4 frequency ranges (2-4-6-8-10 Hz) based on wavelet transform• HFQRS - the amplitude of the RF components of the QRS wave. Additionally used parameters:• TpTe, VAT, QTc - are duplicated to control the correctness of the record processing (the value of the UCC should be approximately equal to the median of the tick-by-bar).• QRSw - QRS width.• RA, SA, TA - the amplitudes of the R, S, T-waves, respectively, are used to normalize the parameters listed above. Statistical analysis and modeling will be performed using Python V3.8.8 and R V.4.0, as well as SPSS v.17. The correlation between various combinations of ECG time, amplitude, energy, and frequency parameters and the presence or absence of PE will be analyzed. Specific parameters will be incorporated into various multivariate analysis and machine learning models: Lasso regression, random forest, multilayer perceptron, support vector machine, and decision tree. The model with the highest diagnostic accuracy will be selected and used to test the algorithm. The outcome of this study will be the development and testing of an algorithm for pulmonary embolism detection using digital analysis of a single-channel ECG witj elements of artificial intelligence. The endpoints of the study are the parameters of diagnostic accuracy of the developed model: * specificity, * sensitivity, * prognostic significance of a positive and negative result, * diagnostic accuracy. Тhese metrics will be calculated using receiver operating characteristic (ROC) analysis and confusion matrices on a held-out test set (30% of the dataset) after training multifactorial models (logistic regression, random forest, or neural networks) on single-lead ECG features. Sensitivity, specificity, positive/negative predictive values, and overall accuracy will be derived by comparing model predictions of pulmonary embolism against the gold standard, with cross-validation (k=5 folds) to ensure robustness and bootstrap resampling for 95% confidence intervals.