Computer-Aided Diagnosis and Treatment for Hemorrhagic Stroke

1430 

Abstract Submission 

Minyue Liu¹, Hengyuan Ma², Jiawei Zhang², Tianye Jia²

¹Fudan University, Shanghai, shanghai, ²Fudan University, Shanghai, Shanghai

Minyue Liu  
Fudan University
Shanghai, shanghai

Hengyuan Ma  
Fudan University
Shanghai, Shanghai

Jiawei Zhang  
Fudan University
Shanghai, Shanghai

Tianye Jia  
Fudan University
Shanghai, Shanghai

Hemorrhagic stroke is a disease caused by the sudden rupture of a cerebral aneurysm or leakage due to weakened blood vessels, leading to disturbances in cerebral blood circulation. In patients with vascular diseases, various factors can induce narrowing, occlusion, or rupture of cerebral arteries, resulting in acute cerebral hemorrhage. (Caplan et al, 2023) Clinically, this manifests as temporary or permanent symptoms and signs of brain dysfunction.Hemorrhagic stroke onset is rapid, progression is rapid, and prognosis is poor, with a mortality rate as high as 45-50% during the acute phase. About 80% of patients will leave behind significant neurological dysfunction, bringing a heavy health and financial burden to society and the patients' families.(Thakur et al, 2022)Therefore, it is of significant clinical importance to identify the risk factors of hemorrhagic stroke, integrate imaging characteristics, clinical patient information, and clinical treatment plans to precisely predict patient prognosis and optimize clinical decisions accordingly.

The aim of this study was to utilize clinical information and imaging data from stroke patients to analyze and predict the expansion of hematomas and the progression of edema, and to reveal the correlation between treatment modalities and the degree of patient recovery.We acquired 100 CT images from the hospital, in which they were patients with the different stage of stroke with clinical data.Initially, we applied the Naive Bayes method to predict the expansion of hematomas.Concurrently, a dual-exponential model was employed to model the temporal evolution of brain edema volume in patients, serving as the basis for patient classification.The time constant in the model reflected the time-course characteristics of the two edema mechanisms, which was corroborated by clinical data.We applied grouping as a prior for non-parametric probability inference, and further discussed the correlation between different treatment modalities and edema volume groups.Additionally, we utilized ordered logistic regression to predict the clinical prognosis within 90 days based on patient population-level clinical data and initial and follow-up imaging data.Finally, we employed the LIME method for personalized clinical prognostic factor analysis, obtaining an ordering of prognostic key factors applicable to individual patients.

The average accuracy of our model on the validation set just with the clinical data is 0.7400, with a standard deviation of 0.3040, indicating that our model possesses a certain degree of generalization capability.After learning, the accuracy of our model on the training set is 0.91.By integrating the model with clinical data, it was discovered that the edema in the mild group was predominantly caused by cytotoxic edema, while the edema in the moderate and severe groups was primarily due to secondary vasogenic edema. Additionally, it was found that incorporating follow-up imaging data could reduce the prediction error by 30% on the validation set, and provided conclusions consistent with existing literature and testable new clinical recommendations.

In our statistical prediction at the group level, certain clinical and imaging information significantly affect the 90-day mRS scores of stroke patients.However, the specific impact of each factor and the prioritization of their importance vary from individual to individual.Therefore, it is imperative to utilize big data to train statistical models at the group level and then employ the LIME method for personalized analysis at the individual level. Such an approach is essential for the prediction and diagnosis of mRS. After training the model using big data, employing personalized and interpretative methods such as the LIME method for precision medicine is also a future trend in the development of medical artificial intelligence.

Classification and Predictive Modeling ¹

Methods Development

Imaging Methods Other ²

Computed Tomography (CT)

Data analysis

Statistical Methods