Andreia Vaconcellos Faria's team
These are my team's members:
Andreia V. Faria, Johns Hopkins University, Associate Professor of Radiology,
Shun Liu, Johns Hopkins University, Post-doctoral fellow,
Wen Zhang, Johns Hopkins University, Research assistant,
Dr. Hisham Abdeltawab (listed in Phase I) moved to another institution and is no longer member of our team.
SOOP: 10.17605/OSF.IO/YQKTJ
StrokeFAIR: https://doi.org/10.3886/ICPSR38464.v5
We used public datasets to train and test models to predict functional domains that compose NIHSS in patients with acute ischemic strokes. NIHSS is a key metric for assessing the clinical impact of acute strokes and is critical for treatment decisions and prognostic modeling. its automated calculation supports practical aspects of patient care and serves as a basis for evaluating functional impairment. Predicting functional outcomes by integrating identify the features that define it advances our understanding of the brain functions and aids in identifying biomarkers of brain health. We developed public, user-friendly tools to calculate NIHSS domains, providing real-time results. From nearly 3,500 patients, we generated and released tabular lesion-loading data to support hypothesis testing, reproducibility, and new resource development. We created a platform that enables analysis of new data within this knowledgebase. These tools promote accessibility and the democratization of science
We analyzed two large public datasets of patients with acute ischemic stroke. The first, StrokeFAIR (Annotated Clinical MRIs and Linked Metadata of Patients with Acute Stroke), includes 2,888 images and metadata, publicly available at ICPSR (https://www.icpsr.umich.edu/web/ICPSR/studies/38464). Together with other resources we developed for automatic stroke image analysis, such as the Acute Stroke Detection and Segmentation (ADS) tool (https://www.nitrc.org/projects/ads) and the Digital Atlas of Brain Arterial Territories (https://www.nitrc.org/projects/arterialatlas), this work received the 2023 DataWorks! Exemplary Achievement Award. The second dataset, the Stroke Outcome Optimization Project (SOOP), includes 1,737 patients and is available on the Open Science Framework (OSF).
We used ADS to harmonize diffusion-weighted images (DWIs), where ischemic injury was manually delineated, aligning them in intensity and space to the JHU_SS_MNI182. This template supports several public segmentation schemes that we previously developed, validated and distributed through ADS, MRIStudio.org, and GitHub (Eve_Atlas). Using ADS, we applied these parcellations to individual brains to quantify the percentage of each region affected by stroke. In this study, we used the JHU_MNI_SS_BPM_TypeI_v2.1 whole-brain segmentation, defining 92 ROIs covering white and gray matter. These lesion loadings served as model inputs to predict acute functional impairment (NIHSS domain scores). From StrokeFAIR, 913 unique cases with same-day MRI and domain-specific NIHSS labels were used for training and testing.
We used several unsupervised machine learning methods (e.g., Random Forest, XGBoost) and supervised neural networks (e.g., tabular deep learning models such as TabPFN) for model training and validation. All models were implemented with public, well-established Python packages. Training followed a two-step approach: (1) binary classification to predict whether each NIHSS domain was affected, and (2) prediction of NIHSS scores within affected domains, using either regression or cascaded categorical classification. Twenty percent of the data was reserved as an independent test set, unseen during training and validation, with balanced NIHSS score distributions across subsets. Model performance was evaluated using balanced accuracy (BACC), F1-score, and confusion matrices to assess agreement between predicted and true scores. Feature importance was analyzed using Mean Decrease Impurity (MDI) for unsupervised models and SHAP values for supervised ones, identifying both global and individual predictors of functional impairment.
We achieved moderate to high accuracy across NIHSS domains, establishing a foundation for an automated system that outputs domain-level NIHSS predictions and the most influential features contributing to each result. This approach ensures interpretability and quality control. We are now refining these models using complementary strategies. One strategy integrates disconnectome maps—representing the probability of structural disconnection caused by stroke—computed with the public BCBtoolkit. These maps have been processed and dimensionally reduced for model use and will be publicly released. Another strategy incorporates multimodal MRI data, already publicly available, to capture not only the acute lesion but also broader brain characteristics such as preexisting atrophy or microvascular white matter disease. All imaging modalities are being integrated using our modality-agnostic multimodal encoder (VAE), which supports image compression, de-identification, secure transfer, and synthesis. The encoder will be released after final validation (estimated within six months).
A key next step is to include non-imaging data, such as demographic and clinical profiles, available in StrokeFAIR. These variables can improve prediction accuracy and highlight clinical factors most associated with impairment severity. Integrating these multidomain inputs (imaging, demographic, and clinical) presents a challenge, but recent deep learning methods such as TabPFN allow this joint analysis efficiently.
Finally, we project these multimodal inputs into a shared latent feature space that compactly represents the full patient profile. A prototype of the integrated analysis and visualization platform is available at https://www.strokefair.org/. This platform supports data exploration, image-based search, and comparison of new cases to similar individuals in the knowledgebase. It is the basis of a public, continuously expanding stroke knowledgebase, enabling reproducible research, model refinement, subgroup analysis, and ultimately, the advancement of personalized medicine.
This project followed an Open Science approach, using open datasets, open-source tools, and producing several public, accessible resources aligned with FAIR principles:
1. Prediction models: We developed a model to predict acute functional impairment (NIHSS) after stroke. Using lesion loadings, we achieved moderate accuracy in predicting affected domains and NIHSS by domain, with balanced accuracy (BACC) ranging from 0.5–0.75. BACC is a more reliable metric than traditional accuracy, which in our case reached 0.8–0.9, as it accounts for sample distribution. Prediction of specific NIHSS subdomains remains limited by sample size; for example, some subdomains have fewer than 20 subjects with high scores (3–4). This highlights the need for data sharing, since even large datasets may lack sufficient variability across all tasks. Although all classifiers provided comparable results, given its ease of use and ability to integrate data from multiple domains, we plan to use TabPFN in the next stage of combining non-imaging information to further refine our models.
2. Feature identification: We identified injury to brain structures influencing acute impairment, providing a basis for biomarker discovery and neuroscientific studies. Feature analysis revealed key brain regions linked to specific dysfunctions. For example, the contralateral internal capsule strongly predicted motor deficits in leg and arm, while left temporal regions predicted higher “best language” scores. These results align with established brain–function relationships and serve as a biological quality control of our models. Importantly, the feature pools identified by Random Forest and TabPFN were highly consistent, confirming that our models rely on biologically meaningful features rather than noise, a key concern for black-box deep learning models.
3. Data resources: We generated lesion loadings from nearly 3,500 acute ischemic stroke patients (StrokeFAIR and SOOP), mapped in three parcellation schemes (arterial territories, lobar atlas, and JHU atlas). These open resources can be combined with demographic and clinical data (already public, within the source dataset) for replication, validation, or new hypothesis testing. We made these data is available at https://zenodo.org/records/17281967.
4. Software outputs: We are extending ADS tool with NIHSS estimation by domain, probability scores for each prediction, and feature interpretation at the individual level, in addition to the already existing functions such as output of lesion loadings, harmonized images, ASPECTS indices, and radiology-style reports. ADS was refactored for modularity and deployment across platforms, being now available as software, container (operating in command line or GUI), and cloud service (https://www.strokefair.org/openads).
5. Web platform: We developed an R Shiny-based web service (prototype in https://www.strokefair.org/) where researchers can analyze their data in the context of our knowledgebase. Users will be able to explore lesion–outcome associations, group patients by lesion or metadata similarity, and generate reports such as NIHSS or ASPECTS predictions. By enabling comparison against population-level data, the platform promotes user contribution, progressively enlarging the dataset and reducing bias. Public release is expected within 6 months.
All outputs (scripts, matrices, models, software, services, platforms) are shared via GREI sand recommended scientific repositories such as OSF, Zenodo, NITRC, and GitHub, or shared in open cloud environments. By releasing both data and tools in user-friendly formats, within a FAIR and secure framework, we enable both computational scientists and clinical researchers to reproduce our findings, validate models, and build upon them. Resources can be used as test sets or for developing new hypotheses. The emphasis on open, standardized, and well-documented outputs ensures replicability, reproducibility, and broad community use.
The impact of this project on human health occurs across multiple domains:
1. Clinical application (diagnosis and management): We developed an automated, user-friendly system to estimate NIHSS by domain. NIHSS is required in stroke care but takes ~10 minutes to administer and shows inter-rater variability of up to 5 points. Automated NIHSS can provide real-time scoring, act as a second evaluator, assist in triage when specialists are unavailable, and generate structured reports. This accelerates workflow, reduces variability, and supports downstream AI applications for prognosis and care planning.
2. Scientific contribution (brain–function relationships): By linking domain-specific NIHSS scores to lesion patterns, we identified features that predict functional impairment in specific domains. This goes beyond the global NIHSS score, offering new insights into brain function, disconnection, and compensation. Including demographic and clinical metadata allows investigation of how factors such as age, vascular disease, or brain atrophy affect outcomes, supporting research into “brain health” and resilience.
3. Data-driven interpretation: We created a public platform integrating thousands of acute ischemic stroke cases that supports data exploration, image-based search, and comparison of new cases to similar individuals in a feature space. This resource contributes to neuroscience, radiology, and data science by providing harmonized, accessible, and comprehensive, yet safe and privacy-preserved image, demographic, and clinical data from patients with stroke. By enabling researchers and clinicians to explore, compare, and learn from large-scale stroke data, this platform might potentially advance precision diagnostics, guide treatment strategies, and lay the foundation for personalized medicine in cerebrovascular disease. It will serve as a continuously expanding, open knowledgebase for reproducible and data-driven stroke research.
4. Resource democratization (biobank and tools): All derived resources (for example, lesion loadings) are openly shareable, as they do not constitute restricted human data. Combined with existing open datasets (StrokeFAIR, SOOP) and tools (ADS, arterial atlas), these outputs provide a reusable testbed for neuroscientists, clinicians, engineers, and developers. Researchers can train, test, and validate models; explore hypotheses; or develop new computational tools. This democratization expands access, promotes innovation, and accelerates translation of methods into clinical practice.
By uniting open data, robust models, and accessible platforms, this project advances neuroscience, improves reproducibility, and fosters tools that directly impact diagnosis, prognosis, and treatment planning in stroke care, ultimately supporting precision medicine.
The project was completed within the award period and met, and in some aspects exceeded, the proposed goals. We used public datasets and software to generate computable data objects from nearly 3,500 patients with acute ischemic stroke. The resulting dataset is publicly available and supports hypothesis testing, replication, and reproducibility studies. We developed models to predict NIHSS domains using lesion loadings, identifying key brain regions whose damage explains acute deficits, thus providing in vivo mapping of brain function. We are integrating these models into a public software platform that processes MRIs, segments and quantifies lesions, extracts lesion loadings across multiple parcellation schemes, and computes clinically relevant indices, including those related to radiological reports and NIHSS scores. We created a visualization system which allows users to explore their data within the knowledgebase, retrieve similar cases, and perform subgroup or stratification analyses to refine personalized models.
No major revisions were made to the project scope. A main constraint was that the SOOP dataset did not provide NIHSS scores by domain. We addressed this by validating our models on total NIHSS and initiating collaboration with the SOOP developers, who plan to release domain-level scores after the publication of their own work using these data, enabling full validation of our models.
We implemented multiple levels of quantitative and qualitative quality control. Metadata were checked for completeness, missing values, and outliers. Image quality was assessed using intensity-based indices and registration metrics (e.g., SSIM), and segmentations were evaluated against individual brain anatomy using quantitative indices described in our previous publications. The calculation and visualization of quality control indices and figures for each case are integrated as outputs of our public software. A trained neuroradiologist visually inspected all images and stroke masks in native and standard space to ensure accuracy and consistency with acute ischemic stroke diagnoses, especially in the SOOP dataset, which includes subacute cases.
The main limitation was incomplete clinical detail in public datasets, sometimes requiring manual review or developer access. These challenges underscore the need for secure, interoperable systems for responsible clinical data sharing and reuse.
The main challenges in data reuse involved access to detailed clinical information and computational demands during large-scale analyses. To obtain NIHSS scores broken down by domain, we paid a third part service (the Core for Clinical Research Data Acquisition) to retrieve them directly from medical records. For the SOOP dataset, we contacted its developers to request domain-level NIHSS scores, which are not yet publicly released; they agreed to share them after their own publication. Another challenge was the computational cost of processing thousands of imaging cases. We addressed this by collaborating for access to high-performance computing (HPC) resources during data processing and model training. Importantly, the resulting software and workflows were optimized to run efficiently on standard computers and cloud systems in real time, ensuring that future analyses and third-party use will not require specialized hardware.
