الفهرس 
Title pageOnly 14 pages are availabe for public view
 |  | from | 79 |  |  |
| | | from | 79 | | |
Abstract
This section summarizes the previousThis study was motivated by the need for reducing the mortality rate due to the complications that happened during the dialysis session. Therefore, working on good quality data of 6000 patients collected from Egypt aided in achieving the objective goal. A comparative study of various individual and ensemble classifiers was implemented. The performance of both multiclassification and binary analyses was calculated. Then, applying feature selection using simple filter and RF approaches for selecting the highly ranked 25, 12, and 6 features out of the 50 features were considered.
In this thesis, the early and accurate prediction of intra-dialytic complications with a minimum number of features is proposed. The results proved that the performance remains stable with 25 features in multiclassification and only 12 features from 50 features in binary analysis. The BpV, Age, Duration of the dialysis session, Heart rate variability, meal, UFR, mean hemoglobin, room humidity, white blood counts, dialysate sodium, urea reduction ratio, and room temperature were the most important 12 features in the prediction. The results also showed that the ensemble techniques achieved higher performance than the individual classifiers for the binary datasets. The RF achieved 98% with the least training time in predicting the occurrence of intra-dialytic side effects.
According to the study performed in this thesis, the derived conclusions can be summarized as follows.
• In multi-classification, the best performance measures were achieved when 25 features were applied by ANN with an accuracy of 82%.
• In binary classification, the best accuracy was obtained when 12 features were applied with the filter feature selection technique.
• When predicting the occurrence of the complications, the results proved that the ensemble technique using RF achieved the highest performance of 98%.
• When predicting intra-dialytic hypotension, hypertension, and dyspnea, the highest F1-score of 94%,92%, and 78% respectively was achieved by GB with a suitable training time.
• The ensemble-based DT had the best accuracy higher than other ensemble and individual classifiers, although the individual DT had the worst performance other than individual classifiers. This proves the power of using the ensemble technique in clinical prediction.
In conclusion, the results in this thesis proved that the ML classifiers can predict with high performance the occurrence of intra-dialytic complications with only 12 features. Furthermore, specifying its type from the seven most common happened side effects. This will reduce the mortality rate of patients on dialysis by the early detection of intra-dialytic
complications. This thesis gives promising results that the model can be used in the dialysis units in Egypt. Therefore, the manufactures of the dialysis machines can add a smart feature to the device, which is to early predict the happening of any comorbidity during the session and notify for emergency alerts, so nephrologists and nurses can take a suitable decision during the session.
5.2 Future work
This research outlines a viable outcome; however, it had some limitations. Therefore, an extension is recommended, and the future work will include the following.
• Consider using a deep learning algorithm in the prediction of HD complications (like CNN and RNN approaches).
• Study the performance of ensemble techniques performance when adding Deep Learning as one of the classifiers.
• Consider using another technique of feature extraction by employing wrapper or hybrid methods.
• Consider enhancing the ensemble model’s performance by applying further hyperparameters tuning.
• Consider the problem of the imbalanced datasets by using the augmentation technique.