Ionic concentration estimation from electrocardiogram for chronic kidney disease and diabetes

Chronic kidney disease emerges as a significant health problem in our country and worldwide. In studies conducted in our country by the Turkish Nephrology Society, 15.7% of adults have chronic kidney disease at various stages. This rate means that there are approximately 7.5 million chronic kidney disease patients, which means that one in every 7 adults has kidney disease. The first treatment for these patients is hemodialysis. However, the mortality rate due to cardiovascular diseases in patients undergoing dialysis is 10% to 30% higher than in the general population. One of the reasons for the connection between these two diseases may be the varying ionic concentrations of calcium and potassium in the two patient groups. It appears to be important to continuously monitor the ionic concentrations affected by dialysis in order to analyze the underlying effects and predict potential risks. Especially in patients undergoing hemodialysis, the estimation of blood potassium (K+) and calcium (Ca2+) levels based on electrocardiogram (ECG) is a promising non-invasive method for monitoring electrolyte imbalances. This method can predict serum electrolyte concentrations by utilizing morphological changes in ECG signals such as the QRS complex and T wave, thereby helping to reduce the need for frequent blood tests. Similar to chronic kidney disease, diabetes is a condition characterized by hyperglycemia due to disturbances in insulin secretion, insulin action, or both, which leads to organ damage in the long term. Pathogenic processes range from the autoimmune destruction of pancreatic β-cells to insulin resistance and insufficient insulin secretion. Among the symptoms of diabetes patients are weight loss and blurred vision. Chronic hyperglycemia leads to increased susceptibility to growth disorders and infections. Among the acute long-term complications, there is a risk of retinopathy, nephropathy, peripheral and autonomic nerve disease, as well as cardiovascular diseases. The relationship between EKG and blood sugar has recently become a significant focus of research. It is known that changes in blood sugar levels affect the electrical activity of the heart. Periods of hypoglycemia (low blood sugar) and hyperglycemia (high blood sugar) have significant effects on certain parameters in the ECG. Especially, the ST, QT, PR intervals, and heart rate variability (HRV) parameters may be sensitive to these changes. Studies have shown that there is a linear relationship between these parameters and blood sugar levels. This finding paves the way for the development of new methods that will allow blood sugar levels to be predicted non-invasively through ECG data. Thus, it is thought that continuous glucose monitoring could be provided in a painless, low-cost, and user-friendly manner, especially for diabetic patients. To manage common health issues such as diabetes and chronic kidney disease (CKD), it is very important to accurately monitor electrolyte imbalances in order to improve patients' treatment processes. Monitoring these imbalances is generally done through blood tests in current clinical practices. However, the invasiveness of these tests can discomfort patients and may be inadequate in situations requiring continuous monitoring. As a result, contemporary medicine should develop alternative methods that are less invasive and do not require continuous monitoring. This thesis examines how biosignals, particularly ECG signals, can be used to monitor ionic concentrations. ECG records the electrical activity of the heart, and changes in electrolyte levels in the body can be associated with these signals. Changes in electrolyte levels can cause morphological changes in specific waves on the electrocardiogram. This relationship between electrolytes and ECG suggests that ECG signals can be used to predict ionic concentrations. In the study, ECG signals from 116, 101, and 1062 individuals were used for Potassium, Calcium, and Glucose, respectively. Artificial intelligence-based algorithms have been designed using the ECG signal for classification and regression. For the study, 61 features were extracted from the ECG. It was attempted to determine whether the extracted features are statistically distinguishable for normal and abnormal conditions or concentration levels using the ANOVA and F-test methods. Additionally, the features were reduced three times using ANOVA and F-test feature selection methods and applied to classification and regression algorithms. In the first part of the study, it was discussed how machine learning methods can be used in this field. Using K-Nearest Neighbors (k-NN), Support Vector Machines (SVM), and Ensemble Learning Methods, classification models were examined to predict abnormal and normal conditions in potassium, calcium, and glucose concentrations. The performance of these models was compared using a series of criteria, including accuracy, sensitivity, specificity, F-measure, Matthew's correlation coefficient (MCC), and kappa score. According to the classification results, normal and abnormal conditions for calcium data were successfully classified using the k-NN classifier with 14 features, achieving 75.93% sensitivity, 55.32% specificity, and a classification accuracy rate of 66.34%. For potassium data, normal and abnormal conditions were successfully classified using the Ensemble classifier with 6 features, achieving 47.83% sensitivity, 86.02% specificity, and a classification accuracy rate of 78.45%. For glucose data, the conditions of hypoglycemia, hyperglycemia, and euglycemia were successfully classified with 61 features, achieving 56.47% sensitivity, 80.52% specificity, and 80.13% classification accuracy. The models determined for the ensemble showed high accuracy, particularly in predicting potassium and glucose levels. However, due to the low sensitivity rate, true positive results are overlooked. This situation indicates that the model can be used for potassium and glucose predictions where high accuracy is required, but false negative results can be tolerated. The k-NN model, which works more balanced on calcium levels, received the highest F-measure score. This indicates that k-NN could be a suitable model for applications requiring a balance between false positive and false negative results. Ensemble methods have provided a balanced approach by offering average values across all criteria without excelling in any specific area. This emphasizes that the model should be selected according to the requirements of the application. In the second part of the study, regression models and methods for predicting calcium, potassium, and glucose levels in patients with chronic kidney disease and diabetes are proposed. For these patients, it is very important to accurately monitor ionic concentrations. Changes in these concentration levels can increase the risk of heart disease in patients undergoing hemodialysis or cause health problems related to sudden blood sugar spikes in diabetic patients. Therefore, patients' quality of life and chances of survival can be improved through continuous monitoring and rapid intervention. For this purpose, important features from ECG signals are compared and evaluated using Ensemble regression, Artificial Neural Network regression model, and fuzzy inference systems (FIS) regression algorithms. The performance of the proposed approach was evaluated using metrics such as RMSE and R². According to the regression results, in the prediction of concentration values, the RMSE value was 0.646 and the R² value was 0.3495, indicating success for potassium using the Ensemble regression method. For the calcium dataset, it shows high performance with NN-R. According to these results, with the selected 4 features, an R² value of 0.3732 and an RMSE value of 0.71 were obtained. For glucose, Ensemble methods, with the Bagging model, showed the best performance in the 13-feature model, achieving the lowest RMSE value of 2.3061 and an R² of 0.4354. This shows that feature selection has an impact on accuracy. In the NN-R algorithm, the 2-layer model with sigmoid activation yielded the best result (RMSE 2.8375, R² 0.1737), but overall, the R² values of the NN-R models remained low. This situation indicates that NN-R models are limited in explaining this data. In conclusion, this doctoral thesis presents a new method for monitoring and predicting ionic concentrations in the field of health. The use of ECG signals as biomarkers has great potential as a cheap, reliable, and continuous monitoring alternative. The results of this study could serve as a foundation for broader clinical applications and new technologies in healthcare in the future. Additionally, the quality of life of patients and the effectiveness of healthcare services will improve through the further development of these methods and their integration into existing health monitoring systems. Therefore, it is expected that machine learning and signal processing techniques will be used more widely in the field of healthcare.