University of Twente Student Theses
Neurological Prognostication of Postanoxic Coma Patients
Schrijver, R. (2025) Neurological Prognostication of Postanoxic Coma Patients.
Full text not available from this repository.
| Full Text Status: | Access to this publication is restricted |
| Embargo date: | 31 January 2027 |
| Abstract: | In recent times, the medical sector has begun to embrace artificial intelligence and data driven technologies to improve the standard of care. These tools can assist hospital staff in making decisions and have the potential to reduce their workload in the future. Given the demographic situation facing the Netherlands, with a growing elderly population, healthcare costs must be seriously curbed, and some form of automation will help to achieve this goal. This thesis is aimed at employing statistical models to aid in the prognostication of the neurological outcome of postanoxic coma patients. The models serve to address shortcomings in the existing prognostic protocol used at Amsterdam UMC, particularly regarding quantifying uncertainty for an individual patient. The dataset underlying the modelling provided by said hospital is plagued with missing values. Multiple imputation surfaced as the most statistically sound way of tackling missingness. Subsequently, five Bayesian logistic regression models were fit. Each of the models corresponded to a certain time stamp after cardiopulmonary resuscitation (CPR) had been performed, where CPR traits and medical test outcomes are considered. The models were applied to three patients to view the evolution of their predictive distribution over time. The predictive performance of the models constructed are judged with several evaluation metrics. Four of these models are also compared to similar models developed on a different study population by researchers affiliated with the University of Pittsburgh Medical Center (UPMC). A limited number of evaluation metrics facilitate the comparisons. The analysis of the three patient’s predictive distribution showed that the first and second model mostly convey uncertainty. Qualitatively, the uncertainty practically fully dissipates in the fourth model. The performance at the group level of the internally developed models steadily increases as the pool of variables is enlarged, meaning that overfitting has not yet occurred. The fourth model, which includes variables collected at twelve hours after CPR and beyond, is found to perform competitively to the ERC/ECISM prognostic algorithm with a sensitivity of 40.3% at a specificity of 99.4%. When attempting to reach a sensitivity of 100%, as desired, the sensitivity is expected to drop to a value between 16.8% and 40.3%. The fourth model therefore offers the most clinical utility. Prognostication at an earlier stage seems a more futile endeavour. The externally supplied models initially showed similar performance, but by the third and fourth comparison discrepancies emerged. This indicates a lack of generalisability of models in this domain. Future research could be aimed at turning the proof of concept outlined in this thesis into an application that can be used in parallel with the current prognostic routine at the Amsterdam UMC. |
| Item Type: | Essay (Master) |
| Clients: | Amsterdam UMC, Amsterdam, Netherlands |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 31 mathematics, 44 medicine |
| Programme: | Applied Mathematics MSc (60348) |
| Link to this item: | https://purl.utwente.nl/essays/104971 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page