Time series predictions in healthcare applications

• المؤلف الرئيسي: Zhong, Shaojie
• مؤلفون آخرون: Fan Xiuyi
• التنسيق: Final Year Project
• اللغة: English
• منشور في: Nanyang Technological University 2024
• الموضوعات: Computer and Information Science
• الوصول للمادة أونلاين: https://hdl.handle.net/10356/175261
• المؤسسة: Nanyang Technological University

In recent years, deep learning has been applied in the medical field to perform at sorts of tasks such as the prediction of in-hospital mortality. Due to the inconsistency of follow-up appointments and recordings of vital signs, irregularity is inherently present in medical data. This irregularity inherent in the data poses substantial hurdles to traditional machine-learning techniques as these techniques can only be employed on time-series data with regular intervals. Over the years, researchers have devoted considerable efforts to tackling this challenge which has led to the development of methods falling into two main categories: interpolation-based and non-interpolation- based models. In this study, we propose a novel approach, STraTS-mTAND, which integrates techniques from STraTS, a non-interpolation model, and mTAND, an inter- polation model, to address the binary in-hospital mortality classification problem using irregularly sampled time-series data. Our approach leverages the strengths of each technique while mitigating their respective limitations. We evaluate the effectiveness of STraTS-mTAND using the PhysioNet Challenge 2012 dataset and the MIMIC-III dataset. Our experimental results demonstrate that our approach outperforms existing methods in both PR-AUC and ROC-AUC. Additionally, our approach has also shown to perform better with lesser training data and with sparser and more irregular time-series. Furthermore, we analysed the PhysioNet Challenge 2012 dataset to provide valuable insights, which may be used to improve medical resource allocation for patients in the ICUs.