Reconstruction of Potassium Concentrations with the ECG on Imbalanced Datasets

9 Jun 2020  ·  Nicolas Pilia, Cristiana Corsi, Stefano Severi, Olaf Dössel, Axel Loewe ·

End-stage chronic kidney disease (CKD) patients are facing a 30% rise for the risk of lethal cardiac events (LCE) compared to non-CKD patients. At the same time, these patients undergoing dialysis experience shifts in the potassium concentrations. The increased risk of LCE paired with the concentration changes suggest a connection between LCE and concentration disbalances. To prove this link, a continuous monitoring device for the ionic concentrations, e.g. the ECG, is needed. In this work, we want to answer if an optimised signal processing chain can improve the result quantify the influence of a disbalanced training dataset on the final estimation result. The study was performed on a dataset consisting of 12-lead ECGs recorded during dialysis sessions of 32 patients. We selected three features to find a mapping from ECG features to [K+]o: T-wave ascending slope, T-wave descending slope and T-wave amplitude. A polynomial model of 3rd order was used to reconstruct the concentrations from these features. We solved a regularised weighted least squares problem with a weighting matrix dependent on the frequency of each concentration in the dataset (frequent concentration weighted less). By doing so, we tried to generate a model being suitable for the whole range of the concentrations.With weighting, errors are increasing for the whole dataset. For the data partition with [K+]o<5 mmol/l, errors are increasing, for [K+]o$\geq$5 mmol/l, errors are decreasing. However, and apart from the exact reconstruction results, we can conclude that a model being valid for all patients and not only the majority, needs to be learned with a more homogeneous dataset. This can be achieved by leaving out data points or by weighting the errors during the model fitting. With increasing weighting, we increase the performance on the part of the [K+]o that are less frequent which was desired in our case.

PDF Abstract
No code implementations yet. Submit your code now

Tasks


Datasets


  Add Datasets introduced or used in this paper

Results from the Paper


  Submit results from this paper to get state-of-the-art GitHub badges and help the community compare results to other papers.

Methods


No methods listed for this paper. Add relevant methods here