| Author: | Shao, Huiling |
| Title: | A clinical decision support tool to identify eligible stroke patients for Alteplase |
| Advisors: | Chan, Lawrence (HTI) Chen, Fiona XY (HTI) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Cerebrovascular disease -- Diagnosis Cerebrovascular disease -- Treatment Critical care medicine Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Health Technology and Informatics |
| Pages: | 109 pages : color illustrations |
| Language: | English |
| Abstract: | In the treatment of ischemic stroke, timely and efficient recanalization of occluded brain arteries can effectively save the brain. The first-line treatment for ischemic stroke is thrombolysis. Nonetheless, a number of patients present unpredictable prognosis after Intravenous (IV) thrombolysis treatment, including symptomatic hemorrhage complications (5% among patients receiving Alteplase (rtPA)) and failed systemic reperfusion (60% among patients receiving rtPA). Models based on machine learning have the potential to perform more delicate stratifications of candidates for thrombolysis. In Chapter 2, literature on the accuracy and feasibility of current machine learning models to assist in stroke thrombolysis was reviewed. And based on the literature review findings, in Chapter 3, we identified the proposed improvements and organized them into several research directions. For this thesis, we only focused on one of the research directions with the following three research objectives: 1. To design a feature representing penumbra using modified clinical-diffusion mismatch and investigate the usefulness of this feature in predicting responsiveness to thrombolysis for stroke patients 2. To develop a deep learning empowered, automatic and accurate pipeline to calculate the feature representing penumbra 3. To develop a new machine learning algorithm with high accuracy and interpretability to predict thrombolysis efficiency, based on baseline clinical information and the feature representing penumbra In Chapter 4, we focused on the first research question. Penumbra evaluation is important in predicting alteplase responsiveness and the study aimed to propose a feature representing penumbra and evaluate its effectiveness in predicting responsiveness to thrombolysis. Consecutive stroke patients undergoing thrombolysis were screened and recruited from November 2013 to May 2020. Baseline information and Diffusion-weighted imaging (DWI) were recorded before the start of thrombolysis. One-week National Institutes of Health Stroke Scale (NIHSS) and three-month modified Rankin Scale (mRS) were recorded to assess responsiveness to thrombolysis. The penumbra features were defined using traditional and modified clinical-diffusion mismatch, respectively. After adjusting for covariates, logistic regression analysis revealed that a small value of traditional clinical-diffusion mismatch feature was both independently associated with favorable response (OR = -3.13; 95% CI: -5.52 – -0.74; P = 0.0102) and favorable outcome (OR = -3.57; 95% CI: -6.75 – 0.38; P = 0.0280). A small value of modified clinical-diffusion mismatch feature was both independently associated with favorable response (OR = -3.23; 95% CI: -5.63 – -0.83; P = 0.0082) and favorable outcome (OR = -3.73; 95% CI: -6.94 – -0.52; P = 0.0228). In the prediction model validation, compared to traditional mismatch, modified mismatch feature showed superior accuracy and stability in the five-folder cross-validation experiments using some algorithms, and equivalent accuracy and stability in the five-folder cross-validation experiments using other algorithms. For response prediction, models based on traditional mismatch feature achieved Area under the receiver operating characteristic curve (AUC) of 0.55 ± 0.07 using gradient boosted trees, AUC of 0.56 ± 0.06 using support vector machines and AUC of 0.57 ± 0.06 using k-nearest neighbors; while models based on modified mismatch feature achieved AUC of 0.56 ± 0.06 using gradient boosted trees, AUC of 0.57 ± 0.05 using support vector machines and AUC of 0.59 ± 0.05 using k-nearest neighbors. For outcome prediction, models based on traditional mismatch feature achieved AUC of 0.70 ± 0.08 using gradient boosted trees, AUC of 0.72 ± 0.08 using support vector machines and AUC of 0.73 ± 0.09 using k-nearest neighbors; while models based on modified mismatch feature achieved AUC of 0.70 ± 0.03 using gradient boosted trees, AUC of 0.73 ± 0.08 using support vector machines and AUC of 0.73 ± 0.08 using k-nearest neighbors. In Chapter 5, we focused on the second research question. Longer thrombolytic door-to-needle time is associated with higher mortality. To shorten the pre-treatment assessment time, we designed a fully automatic pipeline to generate penumbra features proposed in the previous chapter. We re-used the DWI and Apparent diffusion coefficient (ADC) images in the previous chapter. The diffusion volume of defined penumbra features were calculated manually and by our convolutional neural network. The automatic pipeline achieved acceptable performance with a dice score coefficient of 0.86. In Chapter 6, we focused on the third research question. When designing machine learning algorithms to predict thrombolysis prognosis, the majority of prior algorithms traded interpretability for predictive power, making it difficult for neurologists to trust and implement the algorithms in clinical practice. Our proposed algorithm is an advanced version of traditional K-nearest neighbors (KNN). By modifying the isotropy in feature space of classical KNN, we were able to attain high interpretability. We re-used the previous patient cohort to demonstrate that our algorithm maintains the interpretability of previous models while enhancing the predictive power compared with the existing algorithms. The results showed that for outcome prediction problem, compared with the following algorithms, our advanced KNN maintained high interpretability while not compromising the predictive power (classical KNN: AUC 0.88 versus 0.68, p = 0.00256, adaptive KNN: AUC 0.88 versus 0.69, p = 0.00639, weight adjusted KNN: AUC 0.88 versus 0.69, p = 0.0044, decision trees: AUC 0.88 versus 0.78, p = 0.04746). Model inference revealed that three variables: modified clinical-diffusion mismatch, age, and baseline NIHSS proved significant feature importance in outcome prediction, which was consistent with previous clinical trials/observational studies. While for response prediction model we didn’t reach conclusive results on whether advanced KNN outperformed the other commonly used algorithms. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/12884