4713
Prediction of Lifetime Acute Ischemic Stroke Risk using Multimodal Models and SHAP-based Interpretability Methods1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, 2Department of Radiology, Zhongshan Hospital, Fudan University, Shanghai, China, 3Simens Healthineers Ltd., Shanghai, China, 4Human Phenome Institute, Fudan University, Shanghai, China
Synopsis
Keywords: Diagnosis/Prediction, Stroke, Stroke occurence prediction; Deep learning
Motivation:
Accurately predicting the lifetime risk of acute ischemic stroke (AIS) remains a significant challenge, and there is a notable scarcity of multimodal models that effectively integrate medical imaging with clinical factors.
Goal(s):
To propose an effective multi-modality deep learning model based on both MR images and clinical factors for improved prediction of AIS occurrence.
Approach:
The model leveraged a clinical-factor-agnostic module to extract clinical features from clinical factors and employed Shapley methods to scrutinize the significance of features.
Results:
The proposed method achieved higher performance than conventional models for the prediction of lifetime AIS occurrence.
Impact: Our model's predictive outcomes could pinpoint individuals at high risk for AIS, allowing clinicians to advise them on self-health vigilance.
Introduction
Stroke is the second leading cause of death and the third leading cause of disability worldwide1, 2, among which acute ischemic stroke (AIS) is particularly prevalent3, 4. Clinicians have conducted the analysis of factors associated with AIS, including diabetes, hypertension and alcohol addiction5 - 7. Magnetic resonance imaging (MRI) could observe individuals with narrow blood vessels. However, multimodal models for assessing AIS potential are presently inadequate, necessitating enhancements in their predictive outcomes.Methods
Study populationThis study included 678 participants from UK Biobank8 (Table. 1), in which 356 individuals remained stroke-free, while 322 suffered from AIS during their lifetime.
Model input
All subjects received brain MRIs (T1, T2FLAIR, T2* and SWI sequences) as imaging inputs for our model, along with 14 clinical factors listed in Table.1 as clinical inputs. Our model generated a binary prediction of potential lifetime AIS occurrence in individuals.
Model architecture
Figure. 1 depicted the model architecture, featuring an initial Extraction Module where the input MRIs predicted standardized clinical factors, which could yield derived clinical features that clarified brain-clinical factor interactions. Utilizing a clinical-factor-agnostic approach, the backbone in this module shared parameters equally across clinical factors. The subsequent Prediction Module, receiving both MRIs and extracted clinical features, generated the binary classification outputs.
Model evaluation
In evaluation stage, the receiver operating characteristic (ROC) curve was applied in all models and visualized for analysis among all the models, and the area under the curve (AUC) was calculated. Accuracy, positive predictive value (PPV) and negative predictive value (NPV) were also employed to show model performances.
During training and validating, we employed 5-fold cross-validation to evaluate performance. Additionally, we compared our model with two other models: an image-input-only CNN model (Image-CNN) and a fully connected model using only clinical factors (Clinical-FC). SHapley Additive exPlanations (SHAP) values9 were applied in the stage to reveal the importance of indicators for our proposed model.
Results
The model performances were shown in Figure. 2 and Table. 2. Our proposed model outperformed others with accuracy (0.62 ± 0.01), AUC (0.64 ± 0.04), PPV (0.65 ± 0.08) and NPV (0.54 ± 0.09). The combination of medical images and clinical factors facilitated the acquisition of more comprehensive features associated with stroke occurrence. Clinical-FC achieved AUC of 0.59 ± 0.02 against 0.56 ± 0.05 for Image-CNN, and achieved NPV of 0.46 ± 0.13 against 0.38 ± 0.17, revealing that clinical factors provided more extensive information in predicting stroke occurrence.The SHAP values revealed that the self-health score was the foremost predictor of stroke occurrence with mean absolute SHAP values = 1.31, followed by self-employment score, T1 image and smoking addiction, with mean absolute values of 0.64, 0.60 and 0.27 respectively (see Figure3(a)). Positive SHAP values suggested a factor's positive contribution to stroke risk, whereas negative indicated a reduction. It was also noteworthy that alcohol addiction, obesity, and short sleep duration could increase the risk of stroke.
It was demonstrated in Figure 3(b) that, given similar self-employment scores (0.05) and smoking habits (mild smoking addiction), a higher self-health score (-0.27) corresponded to an increased risk of stroke with SHAP values of 0.76, in contrast to a lower score with SHAP values of -4.27. This suggested a higher predisposition to strokes among individuals with elevated self-health scores, implying that maintaining physical and mental health might reduce stroke risk. It was observed that the corresponding clinical feature from the Extraction Module focused on outer edge of the cerebral cortex when AIS risk increased, while the feature covered the entire brain when the individual was stroke-free.
Discussion
Our multimodal model outperformed unimodal methods, excelling at extracting clinically informative features that revealed the interplay between clinical factors and brain regions. Its clinical-factor-agnostic backbone streamlined the model and improved training efficiency. The use of SHAP values boosted our method's interpretability, highlighting the self-health scores routinely. Furthermore, our model showed promise for regression tasks, predicting the age of potential AIS diagnosis.Conclusion
The proposed model's superior performance underscored its predictive capability for lifetime AIS risk, advising individuals to remain vigilant about their self-health.Acknowledgements
No acknowledgement found.References
1. Johnson, W., et al., Stroke: a global response is needed. Bulletin of the World Health Organization, 2016. 94(9): p. 634.
2. Collaborators, G.B.D.S., Global, regional, and national burden of stroke and its risk factors, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Neurol, 2021. 20(10): p. 795-820.
3. Powers, W.J., Acute ischemic stroke. New England Journal of Medicine, 2020. 383(3): p. 252-260.
4. Tsao, C.W., et al., Heart disease and stroke statistics—2023 update: a report from the American Heart Association. Circulation, 2023. 147(8): p. e93-e621.
5. Mankovsky, B.N. and D. Ziegler, Stroke in patients with diabetes mellitus. Diabetes/metabolism research and reviews, 2004. 20(4): p. 268-287.
6. Lawes, C.M., et al., Blood pressure and stroke: an overview of published reviews. Stroke, 2004. 35(3): p. 776-785.
7.Larsson, S.C., et al., Differing association of alcohol consumption with different stroke types: a systematic review and meta-analysis. BMC medicine, 2016. 14(1): p. 1-11.
8. Sudlow, C., et al., UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS medicine, 2015. 12(3): p. e1001779.
9. Lundberg, S.M. and S.-I. Lee, A unified approach to interpreting model predictions. Advances in neural information processing systems, 2017. 30.
Figures
