By taking advantage of the increasing amount of available electronic health data, applications of machine learning in the intensive care unit have the potential to improve medical diagnostics and risk stratification. This thesis proposes an approach for early onset prediction of pneumothorax with such technique, using time series data extracted from a clinical database. The prevalence of pneumothorax among patients is identified through ICD-9 codes, and labels for the onset are obtained by relying on proxies closely related to the condition. Both simple algorithms and deep learning networks are used in a sliding window-based prediction framework, and the importance of each feature is measured with weighted Shapley values. The results proved the feasibility of this approach using Long Short-Term Memory models, although the number of false positives is noticeably high. Mechanical ventilation was the most contributing feature for the outcome. In future work, the focus should be on addressing the large class imbalance that prevails, along with considering more well-founded methods of target labeling.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:kth-309330 |
| Date | January 2022 |
| Creators | Malm, Emma |
| Publisher | KTH, Skolan för kemi, bioteknologi och hälsa (CBH), KTH Royal Institute of Technology |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | TRITA-CBH-GRU ; 2022:010 |
Page generated in 0.0018 seconds