| 摘要: | 中風後癲癇是中風後最常見的神經系統併發症之一，對患者的生活質量和長期預後產生了重大影響。儘管已經進行了大量的研究，但對於中風後發生癲癇的準確預測和預防仍然存在挑戰。本研究旨在綜合分析中風後癲癇的發生率、風險因素以及臨床特徵，以建立一個預測模型，幫助醫護人員更好地識別高風險患者並及早介入。通過對大量中風後癲癇患者的回顧性分析，我們發現年齡、中風類型、患者病史中是否有癲癇等因素與中風後發生癲癇之間存在明顯的相關性。我們建立了一個基於這些因素的預測模型，其準確性經過了充分的驗證。該模型不僅有助於預測個體患者發生癲癇的風險，還可以為臨床醫生提供指導，提高中風後癲癇的預防和管理水平。
本研究採用深度神經網路結構，旨在建立一個預測模型，以協助臨床對於中風後是否會發生癲癇的預測能力。對從2012年至2017年共132位病患的神經影像資料以及分為四組實驗，分別為中風後癲癇與中風後無癲癇、中風後早發性癲癇與中風後無癲癇、中風後遲發性癲癇與中風後無癲癇、中風後早發性癲癇與中風後遲發性癲癇，透過交叉驗證的方式進行模型預測，可得出各組實驗結果，在中風後癲癇之預測準確率70.44%±8.78% (68.94-71.94)、敏感度60.44%±20.19% (57.00-63.88)、精確度77.68%±13.77% (75.33-80.03) 以及AUC達到70.53%±8.63% (69.06-72.00)。中風後早發性癲癇之預測準確率為78.35%±3.15% (77.73-78.96)、敏感度70.91%±1.58 % (70.6-71.22)、精確度65.00%±6.01% (63.83-66.17)以及AUC達到81.27%±3.56% (80.57-81.96)。中風後遲發性癲癇之預測準確率為77.15%±6.62% (75.84-78.47)、敏感度59.09%±28.79% (53.36-64.82)、精確度68.89%±10.18% (66.86-70.92)以及AUC達到72.41%±17.62% (68.9-75.91)。最後，中風後早發性癲癇與中風後遲發性癲癇之預測準確率為81.82%±7.87% (80.47-83.16)、敏感度91.67%±8.34% (90.24-93.09)、精確度78.21%±6.6% (77.08-79.33)以及AUC達到88.51%±5.43% (87.58-89.43)。
研究結果顯示，本研究基於深度學習並經由神經影像進行預測，提出了一種關於臨床預測中風後是否發生癲癇的方式，並有助於臨床醫師在評估病患發生癲癇的風險率。
;Post-stroke epilepsy is one of the most common neurological complications following a stroke, significantly impacting patients′ quality of life and long-term prognosis. Despite extensive researches, accurately predicting and preventing post-stroke epilepsy remains challenging. This study aims to comprehensively analyze the incidence, risk factors, and clinical features of post-stroke epilepsy to develop a predictive model that helps healthcare professionals better identify high-risk patients and intervene early. Through a retrospective analysis of numerous patients with post-stroke epilepsy, we found significant correlations between factors such as age, stroke type, and a history of epilepsy with the occurrence of post-stroke epilepsy. We developed a predictive model based on these factors, and its accuracy has been thoroughly validated. This model not only aids in predicting the risk of epilepsy in individual patients but also provides guidance for clinicians, enhancing the prevention and management of post-stroke epilepsy.
This study adopts a deep neural network structure with the aim of establishing a predictive model to assist in the clinical prediction of post-stroke epilepsy. The study uses neuroimaging data from a total of 132 patients from 2012 to 2017, divided into four experimental groups: post-stroke epilepsy vs. no post-stroke epilepsy, early-onset post-stroke epilepsy vs. no post-stroke epilepsy, late-onset post-stroke epilepsy vs. no post-stroke epilepsy, and early-onset post-stroke epilepsy vs. late-onset post-stroke epilepsy. The model predictions are performed through cross-validation, and the results of each group of experiments are obtained.
For predicting post-stroke epilepsy, the accuracy is 70.44%±8.78% (68.94-71.94), sensitivity is 60.44%±20.19% (57.00-63.88), precision is 77.68%±13.77% (75.33-80.03), and AUC reaches 70.53%±8.63% (69.06-72.00). For predicting early-onset post-stroke epilepsy, the accuracy is 79.20%±0.36% (79.13-79.27), sensitivity is 71.21%±6.56% (69.93-72.49), precision is 69.49%±0.49% (69.40-69.57), and AUC reaches 79.72%±4.17% (78.90-80.53). For predicting late-onset post-stroke epilepsy, the accuracy is 74.76%±1.13% (74.54-74.99), sensitivity is 66.77%±3.18% (66.13-67.40), precision is 61.17%±2.67% (60.64-61.70), and AUC reaches 79.02%±2.46% (78.53-79.51). Lastly, for predicting early-onset vs. late-onset post-stroke epilepsy, the accuracy is 81.82%±7.87% (80.47-83.16), sensitivity is 91.67%±8.34% (90.24-93.09), precision is 78.21%±6.6% (77.08-79.33), and AUC reaches 88.51%±5.43% (87.58-89.43).
The results demonstrate that our study, based on deep learning and neuroimaging, proposes a method for clinically predicting the occurrence of epilepsy after a stroke, aiding clinicians in assessing the risk of epilepsy in patients. |
