Artificial Intelligence Quickly And Automatically Screen Hypokalemia From Electrocardiogram

Backgrounds:Hypokalemia is a common electrolyte disturbance,which poses a certain threat to special populations such as heart failure.It is of practical significance to quickly identify hypokalemia.Biochemical testing is the golden standard for diagnosis,but it requires certain hardware requirements,and the return of results often takes more than half an hour.The electrocardiogram?ECG?is easily obtained,and hypokalemia has some non-specific ECG performance.However,the accuracy of manually identifying hypokalemia from the ECG is low.Objective:To quickly and automatically screen hypokalemia by applying artificial intelligence to the ECG.Methods:Extract the standard 12-leads ECG resting position recorded?5 seconds synchronized 6-lead recording ECG?and serum potassium value from September2017 to May 2019 in a hospital.The atrial fibrillation?AF?,complete left bundle branch block?CLBBB?,Complete right bundle branch block?CRBBB?and pacing ECGs were known as confounding ECGs.Exclude death ECGs and blood samples that were given potassium or diuretic treatment during the collection of ECGs and serum potassium samples.According to the principle that patient name,admission number and the time difference between collecting ECGs and serum potassium sample was within 3 hours?if the number of serum potassium value corresponding to ECG>1,the nearest time interval was taken?to match^[1]to obtain effective ECG-blood potassium pairs data sets.The data set was randomly grouped by patients before excluding confounding ECG.Among them,80%of the electrocardiogram was used as the training data set,and 12 leads of the electrocardiogram were collected to obtain a deep learning model?DLM?.Verification data set for cross-validation to evaluate the performance of the model?algorithm?in screening hypokalemia by identifying ECG signals.After eliminating confounding ECG,the training set and verification set were obtained according to the above method,and 12 leads and 2leads of the ECGs were collected to establish the model and verify its effectiveness.The electrocardiograms in the verification set that exclude confounding ECGs are classified to statistical the effect of manual analysis analyzed electrocardiographic to diagnose hypokalemia.Results:?1?A total of 12,450 ECGs were used in our study,In the validation of the12-lead ECG model including confounding factos,the 12-lead ECG model with excluding confounding factors and the 2 lead?I,II?ECG model with excluding confounding factors,the incidence of hypokalemia was 37.0%,47.0%,and 47.0%respectively.The 12-leads ECG model concluding confounding ECGs detected hypokalemia with an AUC value of 0.771?95%CI,0.754-0.787?in its validation data set.When the cut-off value was 0.4212,the sensitivity was 70.0%and the specificity was 69.1%,The positive predictive value was 57.3%,the negative predictive value was 79.6%,and the overall accuracy rate was 69.4%.?2?The 12-leads ECG model excluding confounding ECGs detected hypokalemia with an AUC value of 0.796?95%CI,0.766-0.815?in its validation data set.When the cut-off value was 0.536,the sensitivity was 71.4%and the specificity was 77.1%,The positive predictive value was 73.5%,the negative predictive value was 75.1%,and the overall accuracy rate was 74.4%.?3?The 2-leads?I,II?ECG model excluding confounding ECGs detected hypokalemia with an AUC value of 0.643?0.618-0.668?in its validation data set.When the cut-off point was 0.967,the sensitivity was 70.9%and the specificity was51%.The negative predictive value was 66.2%,the positive predictive value was56.3%,and the overall accuracy rate was 60.3%.?4?Among the electrocardiograms?n=1748?in the validation set that excluded confounding factors,824 cases of hypokalemia were biochemically verified,but only 180?21.9%?of these true hypokalemia electrocardiograms were diagnosed as hypokalemia by manual analysis.In addition,there were 26 cases?3.1%?with diagnosis of obvious U wave,and the biochemical test corresponding to the electrocardiogram of artificially diagnosed hypokalemia was determined to be 116cases of non-hypokalemia.The sensitivity of artificial electrocardiogram to identify hypokalemia is 21.9%and positive predictive value is 60.8%.Conclusions:?1?AI recognition ECG can quickly and non-invasively and effectively screen hypokalemia;?2?Electrocardiograms of mixed factors such as CLBBB,CRBBB and pacing may reduce the performance of AI in screening for hypokalemia?3?Compared with the model of collecting 12-lead ECG signals,the efficiency of the deep learning model of collecting 2 leads?I,II?signals is relatively poor in screening for hypokalemia.