TL;DR
This study introduces a novel radiomics-based model that, when combined with clinical factors, can predict future ischemic strokes in patients with silent lacunar infarction (SLI) more accurately than clinical factors alone. This non-invasive approach offers a promising tool for identifying high-risk individuals and enabling targeted preventive interventions.
What Problem Does This Paper Address?
Silent lacunar infarction (SLI) is a common finding on brain imaging, particularly in aging populations. While SLI is known to be associated with an increased risk of future symptomatic strokes and dementia, current clinical guidelines lack specialized strategies for stratifying risk within SLI patients. This makes it challenging to determine which individuals would benefit most from aggressive preventive therapies. Traditional risk factors like age and dyslipidemia are considered, but their predictive power for future strokes in SLI patients is not fully understood, and it's recognized that relying solely on these factors might be insufficient. The paper aims to address this gap by developing a more accurate, non-invasive method to predict future ischemic strokes in SLI patients, moving beyond traditional risk factors by incorporating radiomic features from CT scans.
Methods
The study employed a retrospective case-cohort design. Key methodological components include:
- Radiomic Feature Extraction: Radiomic features were extracted from baseline cranial CT images of SLI patients. Regions of interest (ROIs) were defined around lacunar lesions. A comprehensive set of features was extracted, including shape-based, intensity-based, and texture features (GLCM, GLRLM, GLSZM).
- Reproducibility Assessment: The stability of radiomic features was assessed using an interclass consistency test (ICC), with features having an ICC > 0.75 being retained.
- Feature Selection and Model Building: A Least Absolute Shrinkage and Selection Operator (LASSO)-Cox regression analysis was used to select significant radiomic features. This was followed by a forward selection method to determine the final features and their coefficients for constructing radiomic, clinical, and integrative (ModelCR) signatures.
- Model Comparison: Three Cox proportional hazards models were constructed: one based on clinical factors (ModelC), one on radiomic features (ModelR), and one combining both (ModelCR). Model performance was evaluated using the concordance index (C-index) and receiver operating characteristic (ROC) curves.
- Prognostic Nomogram and Calibration: A prognostic nomogram was developed based on the best-performing model (ModelCR) to visualize the predicted probability of future stroke. Calibration curves were used to assess the agreement between predicted and observed survival probabilities.
- Statistical Analysis: Mann-Whitney U tests and chi-squared tests were used for clinical characteristic comparisons. Kaplan-Meier curves and log-rank tests were employed to compare stroke-free survival probabilities.
Datasets
The study utilized a dataset of 148 patients with silent lacunar infarction (SLI) who were retrospectively recruited from Zhuhai Hospital Affiliated with Jinan University between February 2013 and December 2016. Patients were divided into training and validation cohorts at a 1:1 ratio. No new benchmark datasets were introduced; the study relied on existing patient data.
Key Insights and Ideas
- Radiomics as a Predictive Tool: The study demonstrates that radiomic features extracted from CT scans can quantitatively capture subtle tissue characteristics that are predictive of future ischemic stroke risk in SLI patients, even when these features are not visually apparent.
- Synergy of Radiomics and Clinical Factors: The combined model (ModelCR), integrating radiomic features with traditional clinical factors, significantly outperformed models based on either radiomics or clinical factors alone. This highlights the complementary nature of these data sources.
- Heterogeneity as a Risk Indicator: Certain radiomic features, particularly those emphasizing the heterogeneity of neighboring voxels and higher intensity values within the lesion ROI, were associated with a worse prognosis, suggesting that the internal complexity of SLI lesions is a key indicator of future risk.
- Clinical Factors Still Relevant, but Less Powerful: While clinical factors like age and dyslipidemia were identified as independent predictors, their standalone predictive power was less robust compared to the integrative model. This suggests that while important, they are not sufficient for comprehensive risk stratification.
Key Results
- Model Performance: ModelCR achieved the highest accuracy, with 6-, 12-, and 18-month areas under the curve (AUCs) of 0.84, 0.81, and 0.79 for the training cohort, and 0.79, 0.88, and 0.75 for the validation cohort, respectively.
- Independent Predictors: Radiomic signature, age, dyslipidemia, and the number of lesions were identified as independent predictors of future ischemic strokes.
- Prognostic Nomogram: A prognostic nomogram was developed based on ModelCR, allowing for the calculation of an individual's probability of future stroke at 6, 12, and 18 months.
- Stratified Analysis: The integrative signature showed significant associations with stroke-free survival across subgroups defined by age, dyslipidemia, and the number of lesions, reinforcing its broad applicability.
Limitations and Future Work
The authors acknowledge several limitations:
- Retrospective Design: The retrospective nature of the study limits the investigation of certain factors (e.g., chronic kidney disease, hyperhomocysteinemia) and detailed medical interventions.
- Selection Bias: The exclusion of patients with cardioembolic stroke to avoid uncertainty might have introduced selection bias.
- CT vs. MRI: The study relied solely on CT-derived radiomic features. MRI is a more sensitive modality for detecting lacunar infarction and could potentially yield more robust radiomic insights.
- Short-Term Follow-up: The study focused on short-term prognoses, and longer-term follow-up data would be beneficial.
- Sample Size: While 148 patients were included, larger, multi-center studies are recommended to further validate the findings and improve the robustness of the model.
Future work suggested includes exploring various imaging modalities and expanding the analysis to larger, multi-center datasets to enhance the accuracy and generalizability of the radiomic-based predictive model.
Summary updated