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

Abstract

Aims: Deep learning models (DLMs) have shown superiority in electrocardiogram (ECG) analysis and have been applied to diagnose dyskalaemias. However, no study has explored the performance of DLM-enabled ECG in continuous follow-up scenarios. Therefore, we proposed a dynamic revision of DLM-enabled ECG to use personal pre-annotated ECGs to enhance the accuracy in patients with multiple visits.

Methods and results: We retrospectively collected 168 450 ECGs with corresponding serum potassium (K⁺) levels from 103 091 patients as development samples. In the internal/external validation sets, the numbers of ECGs with corresponding K⁺ were 37 246/47 604 from 13 555/20 058 patients. Our dynamic revision method showed better performance than the traditional direct prediction for diagnosing hypokalaemia [area under the receiver operating characteristic curve (AUC) = 0.730/0.720-0.788/0.778] and hyperkalaemia (AUC = 0.884/0.888-0.915/0.908) in patients with multiple visits.

Conclusion: Our method has shown a distinguishable improvement in DLMs for diagnosing dyskalaemias in patients with multiple visits, and we also proved its application in ejection fraction prediction, which could further improve daily clinical practice.

Keywords: Best linear unbiased prediction; Deep learning model; Ejection fraction; Electrocardiogram; Linear mixed model; Potassium.

Graphical Abstract

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.