Automated risk prediction of post-stroke adverse mental outcomes using artificial intelligence and machine learning

• Main Author: Oei, Chien Wei
• Other Authors: Ng Yin Kwee
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Medicine, Health and Life Sciences; Automated risk prediction; Artificial intelligence; Machine learning; Deep learning; Post-stroke adverse mental outcome
• Online Access: https://hdl.handle.net/10356/174230

Depression and anxiety are common comorbidities of stroke. Research has shown that about 30% of stroke survivors develop depression, and about 20% develop anxiety. Stroke survivors with such adverse mental outcomes are often attributed to poorer health outcomes, such as higher mortality rates. The objective of this study is to use Deep Learning (DL) methods to predict the risk of a stroke survivor experiencing post stroke depression and/or post-stroke anxiety, which is collectively known as post stroke adverse mental outcomes (PSAMO). This study studied 179 patients with stroke, who were further classified into PSAMO vs no PSAMO group based on the results of validated depression and anxiety questionaries, which are the industry’s gold standard. This study collected demographic and sociological data, quality of life scores, stroke-related information, medical and medication history, and comorbidities. In addition, sequential data such as daily lab results taken seven consecutive days after admission are also collected. Using a combination of DL algorithms such as Multi-Layer Perceptron (MLP) and Long Short Term Memory (LSTM), the model could ingest and make predictions using static and sequential data. The trained model attained a 79.2% binary classification accuracy and performed better than Machine Learning (ML) models. The combination of using DL algorithms that are able to process complex patterns in the data and the inclusion of new data types, such as sequential data, helped to improve model performance. Accurate prediction of PSAMO helps clinicians to make early intervention care plans and potentially reduce the incidence of PSAMO.