Development and validation of a dynamic deep learning algorithm using electrocardiogram to predict dyskalaemias in patients with multiple visits

Figures

• 

  Figure 1. Procedure of patient-level prediction with the deep learning model and the calculation of baseline for prediction. (A) The workflow of dynamic revision with the best linear unbiased prediction. The deep learning model predictions via electrocardiogram are revised with the personal best linear unbiased prediction that involved using previous predictions and corresponding ground truth. The black box indicates the dynamic revision of deep learning model prediction. (B) The calculation of the previous mean value for baseline. The previous mean value is the average value of the previous ground truth values from each patient, which is only on the follow-up electrocardiograms.

• 

  Figure 2. Performance of diagnosing hypokalaemia and hyperkalaemia and estimating serum potassium. Hypokalaemia was defined as a serum potassium level <3.5 mEq/L. Hyperkalaemia was defined as a serum potassium level >5.5 mEq/L. The baseline value predicted the mean value for all patients. The previous value was calculated using the most recent examination result for each patient. The previous mean value was calculated using the mean value of the previous examination result for each patient. The deep learning model (directly) is the original prediction from deep learning models, and the deep learning model (dynamic) is the dynamic revision of deep learning model prediction with the personal best linear unbiased prediction.

• 

  Figure 3. Area under the precision recall curves and F1 scores of the models for diagnosing hypokalaemia and hyperkalaemia. Hypokalaemia was defined as a serum potassium level <3.5 mEq/L. Hyperkalaemia was defined as a serum potassium level >5.5 mEq/L. The baseline value predicted the mean value for all patients. The previous value was calculated using the most recent examination result for each patient. The previous mean value was calculated using the mean value of the previous examination result for each patient. The deep learning model (directly) is the original prediction from deep learning models, and the deep learning model (dynamic) is the dynamic revision of deep learning model prediction with the personal best linear unbiased prediction.

• 

  Figure 4. Impact of the number of follow-up electrocardiograms for diagnosing hypokalaemia and hyperkalaemia using best linear unbiased prediction. The performances of the deep learning model on the sample less than or equal to the indicated numbers were analysed. The green colour indicates the prediction directly from the deep learning model, and the blue colours indicate the prediction that is dynamically revised with the personal best linear unbiased prediction.