Predictive modeling of hematoma expansion from non-contrast computed tomography in spontaneous intracerebral hemorrhage patients

Reviewer #1 (Public review):

Summary:

The study explores the use of Transport-based morphometry (TBM) to predict hematoma expansion and growth 24 hours post-event, leveraging Non-Contrast Computed Tomography (NCCT) scans combined with clinical and location-based information. The research holds significant clinical potential, as it could enable early intervention for patients at high risk of hematoma expansion, thereby improving outcomes. The study is well-structured, with detailed methodological descriptions and a clear presentation of results. However, the practical utility of the predictive tool requires further validation, as the current findings are based on retrospective data. Additionally, the impact of this tool on clinical decision-making and patient outcomes needs to be further investigated.

Strengths

(1) Clinical Relevance: The study addresses a critical need in clinical practice by providing a tool that could enhance diagnostic accuracy and guide early interventions, potentially improving patient outcomes.

(2) Feature Visualization: The visualization and interpretation of features associated with hematoma expansion risk are highly valuable for clinicians, aiding in the understanding of model-derived insights and facilitating clinical application.

(3) Methodological Rigor: The study provides a thorough description of methods, results, and discussions, ensuring transparency and reproducibility.

Weaknesses:

(1) The limited sample size in this study raises concerns about potential model overfitting. While the reported AUCROC of 0.71 may be acceptable for clinical use, the robustness of the model could be further enhanced by employing techniques such as k-fold cross-validation. This approach, which aggregates predictive results across multiple folds, mimics the consensus of diagnoses from multiple clinicians and could improve the model's reliability for clinical application. Additionally, in clinical practice, the utility of the model may depend on specific conditions, such as achieving high specificity to identify patients at risk of hematoma expansion, thereby enabling timely interventions. Consequently, while AUC is a commonly used metric, it may not fully capture the model's clinical applicability. The authors should consider discussing alternative performance metrics, such as specificity and sensitivity, which are more aligned with clinical needs. Furthermore, evaluating the model's performance in real-world clinical scenarios would provide valuable insights into its practical utility and potential impact on patient outcomes.

(2) The authors compared the performance of TBM with clinical and location-based information, as well as other machine learning methods. While this comparison highlights the relative strengths of TBM, the study would benefit from providing concrete evidence on how this tool could enhance clinicians' ability to assess hematoma expansion in practice. For instance, it remains unclear whether integrating the model's output with a clinician's own assessment would lead to improved diagnostic accuracy or decision-making. Investigating this aspect-such as through studies evaluating the combined performance of clinician judgment and model predictions-could significantly enhance the tool's practical value.

Reviewer #2 (Public review):

Summary:

The author presents a transport-based morphometry (TBM) approach for the discovery of non-contrast computed tomography (NCCT) markers of hematoma expansion risk in spontaneous intracerebral hemorrhage (ICH) patients. The findings demonstrate that TBM can quantify hematoma morphological features and outperforms existing clinical scoring systems in predicting 24-hour hematoma expansion. In addition, the inversion model can visualize features, which makes it interpretable. In conclusion, this research has clinical potential for ICH risk stratification, improving the precision of early interventions.

Strengths:

TBM quantifies hematoma morphological changes using the Wasserstein distance, which has a well-defined physical meaning. It identifies features that are difficult to detect through conventional visual inspection (such as peripheral density distribution and density heterogeneity), which provides evidence supporting the "avalanche effect" hypothesis in hematoma expansion pathophysiology.

Weaknesses:

(1) As a methodology-focused study, the description of the methods section somewhat lacks depth and focus, which may make it challenging for readers to fully grasp the overall structure and workflow of the approach. For instance, the manuscript lacks a systematic overview of the entire process, from NCCT image input to the final prediction output. A potential improvement would be to include a workflow figure at the beginning of the manuscript, summarizing the proposed method and subsequent analytical procedures. This would help readers better understand the mechanism of the model.

(2) The description of the comparison algorithms could be more detailed. Since TBM directly utilizes NCCT images as input for prediction, while SVM and K-means are not inherently designed to process raw imaging data, it would be beneficial to clarify which specific features or input data were used for these comparison models. This would better highlight the effectiveness and advantages of the TBM method.

(3) The relatively small training and testing dataset may limit the model's performance and generalizability. Notably, while the study mentions that 1,066 patients from the ERICH dataset met the inclusion criteria, only 170 were randomly selected for the test set. Leveraging the full 1,066 ERICH cases for model training and internal validation might potentially enhance the model's robustness and performance.

(4) Some minor textual issues need to be checked and corrected, such as line 16 in the abstract "Incorporating these traits into a v achieved an AUROC of 0.71 ...".

(5) Some figures need to be reformatted (e.g., the x-axis in Figure 2 a is blocked).