Analysis of the NORA Anaphylaxis Database: Towards Identifying Severity Factors using Machine Learning
- Anaphylaxis is a severe allergic reaction, which occurs instantaneously with potential life-threatening symptoms by affecting different organ systems. The reaction could be triggered by an exposure to various types of allergens including foods, drugs, insects, venom and others. The European Anaphylaxis Registry (NORA) is a large database gathered to collect information on anaphylaxis cases through the allergy centers in European and other countries. This questionnaire-based survey data was designed in a way to provide information on patients’ demographic characteristics, eliciting allergens, symptoms, reaction severity, emergency treatment and long-term management of the disease and aims to improve the knowledge on reactions and provide treatment methods to affected patients. Unfortunately, the data is partly incomplete. However, incomplete data hinders some algorithms to work properly and may lead to misleading analysis results. Another issue regarding the dataset is, it contains a large number of variables, which introduces complex linear or/and non-linear relationships to the data and makes it difficult to reveal those relationships, which makes analyzing and driving conclusions from the data a challenging task. The problems mentioned above, leads us to the questions asked in this thesis. The fact that the data is incomplete raises the main question, which is whether and how it is possible to impute the missing data and how this affects subsequent steps, such as classification. The size and the complexity of the data raises the second question: Are there certain relationships within variables that can be used to predict the severity of the disease? In case there is, are there any potential augmentation factors, also called cofactors in the dataset that may increase the severity of a reaction to an allergen that would normally not occur when the cofactor does not exist? As explained, the goal of this thesis is to find answers to the questions indicated above. For that reason, initially, we searched for a suitable method for data imputation and compared performances of two different tree-based multiple imputation methods on the dataset. After obtaining multiple complete datasets, we tried to predict the severity of anaphylaxis from the dataset to determine hidden risk factors behind anaphylaxis using high power gradient boosting methods. It is concluded that, multiple imputation with chained equations (MICE) based on classification and regression trees can be useful for imputing NORA dataset. However, we could not detect enough associations between study variables and anaphylaxis severity in the data, to predict anaphylaxis severity from the dataset.
| Author: | Ecegül Mengüllüoglu |
|---|---|
| Document Type: | Master's Thesis |
| Granting Institution: | Freie Universität Berlin |
| Advisor: | Tim Conrad, Stefan Zachow |
| Date of final exam: | 2022/08/10 |
| Year of first publication: | 2022 |
| Page Number: | 47 |

