Stroke, as the second leading cause of death and the leading cause of acquired disability, remains a major burden in Western countries (Katan and Luft, 2018; Gbd 2019 Stroke Collaborators, 2021). Over the past decade, substantial advancements have been made in both diagnostic and therapeutic strategies for acute stroke. Perfusion imaging has emerged as a vital tool, offering insights into potentially salvageable at-risk tissue. Acute treatment modalities, notably thrombolysis therapy and mechanical thrombectomy, hold tremendous promise. However, they are not without risks, necessitating a careful evaluation of the potential benefits against inherent risks.

One of the challenges in acute stroke management is the absence of a reliable method to assess the future clinical implications of a specific lesion. The inability to accurately predict untreated outcomes, both in the immediate aftermath and over the long term, complicates acute treatment decisions. These studies are mostly based on either the National Institutes of Health Stroke Scale (NIHSS score) or the volume of brain tissue at risk (Goyal et al., 2016; Ma et al., 2019). However, more detailed knowledge of the potential harm of a particular lesion would be extremely useful.

The critical need to address this issue has spurred a plethora of studies aiming to predict patients’ clinical outcomes, employing various data types. These approaches can be broadly categorized into three main strategies: (i) analysis of pure clinical data, (ii) assessment of lesion size and location, and (iii) integration of network connectivity information.

Efforts to forecast clinical outcomes based solely on clinical data have yielded moderate results, indicating that patients with more severe initial symptoms generally experience poorer outcomes (Sato et al., 2008; Sablot et al., 2011; Wouters et al., 2018; Kazi et al., 2021). Another approach involves incorporating the size and location of the lesion in outcome prediction models. Initially, this approach was expected to enhance prediction accuracy significantly, as the lesion directly contributes to clinical symptoms and potential persistent disability. However, despite considerable efforts, the results have not reached a level of accuracy applicable in clinical practice (Schiemanck et al., 2005; Yoo et al., 2010). This limitation extends to other parameters in prediction models, such as cortical thickness in certain contralesional cortices (Rojas Albert et al., 2022) and metrics of brain age and resilience (Liew et al., 2023).

To obtain data on functional or structural connectivity, resting state functional magnetic resonance imaging (rs-fMRI), diffusion tensor imaging (DTI) and connectivity analyses have been employed (Horn et al., 2014; Ktena et al., 2019; Kao et al., 2020; Koch et al., 2021; Peng et al., 2023). However, these data (rs-fMRI, DTI) are not typically gathered during routine diagnostics in stroke cases. Moreover, obtaining them requires patient cooperation and extends the examination time, rendering them impractical in the hyperacute phase of stroke, where therapeutic decisions are time-sensitive. In acute stroke diagnostics, a diffusion sequence and FLAIR are necessarily acquired in MRI. Our approach utilizes a T1 and FLAIR data set, whereby even the T1 could be generated from the FLAIR (Iglesias et al., 2023). This emphasis on basic MRI sequences reduces the MRI acquisition time to just several minutes. Consequently, systemic thrombolysis and mechanical thrombectomy can be rapidly initiated during the hyperacute phase of stroke.

Our alternative approach combines individual MRI data with probabilistic information sourced from the Human Connectome Project (HCP). We hypothesized that the clinical state of a patient can be predicted using probabilistic information from the HCP, rather than individual DTI data. Specifically, we posit that this connectome information could enhance the predictability of clinical stroke severity, as measured by the NIHSS score, beyond what can be achieved by considering lesion size alone. To estimate the number of fibers traversing the lesion, we calculated the number of disconnected fibers and used the term disconnection value. In order to evaluate the hypothesis, we assessed the number of disconnected fiber tracts to specific cortical areas (quantified by the disconnection value), drawing on probabilistic data from the HCP alongside individual T1 and T2-weighted fluid-attenuated inversion recovery (FLAIR) imaging data from our cohort of 51 patients.

The study was reviewed and approved by the Ethics Committee of Friedrich Schiller University Jena in accordance with the Declaration of Helsinki.

We employed a uniform and standardized examination design for all patients included in the project. Stroke patients underwent clinical examinations, and their medical histories were reviewed. This procedure, as well as asking specific questions within the framework of the anamnesis, is well-established. In addition, we developed a software that utilizes data provided by the HCP along with a high-resolution image of the patient’s ischemic lesion to create a probabilistic surface atlas of the disconnected areas.

In our research study we included 20 right and 31 left hemispheric strokes (n = 51; 39.2% vs. 60.8%; 24 females). All patients reported acute focal neurological symptoms before first presentation in the Emergency Department. The mean admission NIHSS score was 4.4 ± 3.2 points (range: 0–14), and the average length of hospital stay was 7.0 ± 3.0 days (range: 2–15).

To assess the respective importance of lesion size and disconnection value, we initially performed two separate simple linear regression analyses to test if (1) the lesion size and (2) the disconnection value significantly predict the NIHSS score. The model utilizing the lesion size revealed a significant result, explaining 44% of the measured NIHSS score (R² = 0.44; adjusted R² = 0.43, F-value = 38.0; p < 0.001). We observed that lesion size significantly predicted the NIHSS score (β = 0.66, p < 0.001, see Figure 4 and Table 2 for details). Similarly, the model utilizing the disconnection value also revealed a significant result, explaining 60% of the measured NIHSS score (R² = 0.60; adjusted R² = 0.59, F-value = 72.9; p < 0.001). It was found that the disconnection value significantly predicted the NIHSS score (β = 0.77, p < 0.001, see Figure 4 and Table 3 for details).

To further investigate the value of the disconnection value compared to lesion size, we compared the model that uses log lesion size as the sole predictor with a model that includes both log lesion size and the disconnection value as predictors (Model 3, Table 4). The overall regression in Model 3 was statistically significant (R² = 0.64; adjusted R² = 0.63, F-value = 43.1; p < 0.001). We found that both predictor variables made a significant contribution to the model (lesion size: β = 0.28, p = 0.019; disconnection value: β = 0.60, p < 0.001, Table 4). We then compared both models using ANOVA to test whether there is a significant difference between both models in predicting the NIHSS score. Results demonstrated that the addition of the disconnection value significantly improved the fit of the model (F = 27.6, p < 0.001, Table 5). In other words, the prediction of the NIHSS score could be enhanced significantly by incorporating the disconnection value into the regression model (p < 0.001).

This study underscores the potential of probabilistic connectome data in predicting the clinical status of stroke patients, showcasing its superiority over methods relying solely on lesion size. Previous research has already established that connectivity-based models, which integrate additional insights about the brain’s functional and structural wiring, offer enhanced predictive power compared to models limited to clinical data or lesion metrics (Grefkes and Fink, 2014; Silasi and Murphy, 2014; Dulyan et al., 2022; Rivier et al., 2023). This advantage aligns with the prevailing network-centric perspective of brain function, recognizing that brain functionalities are not isolated but emerge from complex, widely distributed networks that necessitate constant informational exchange (Siegel et al., 2022). Our findings also resonate with this perspective, illustrating how focal brain lesions can profoundly impact local and global network functions by disrupting not only the directly affected area but also its connected regions (Foulon et al., 2018; Saenger et al., 2018; Griffis et al., 2019; Salvalaggio et al., 2020). This is exemplified by studies on subcortical stroke lesions that indicate their impact extends beyond the immediate site, influencing connected cortical areas and leading to cortical thinning due to secondary neuroaxonal degeneration (Cheng et al., 2015; Duering et al., 2015).

While patient-specific data derived from high-resolution individual DTI and rs-fMRI examinations would offer the most accurate insight into structural and functional connectivity, the practical constraints of routine clinical practice often render this unfeasible. However, our approach, leveraging a probabilistic connectome, necessitates only the lesion’s size and location and yet outperforms methods lacking connectome insights. This implies that probabilistic connectome information provided by an atlas offers valuable features for predicting a patient’s clinical condition. To optimize prediction models, integrating these features into machine learning approaches is crucial, as the quality of input features heavily influences model performance (Rehme et al., 2015; Mouridsen et al., 2020; Jayakumar et al., 2022). Utilizing preprocessed connectome features is essential, as machine learning models cannot infer the probabilistic connectivity structure on a population basis from limited data. It is crucial to emphasize that for a system to be viable for clinical application, it must be fundamentally grounded in data that can be routinely obtained in clinical environments. While advanced methodologies like DTI and functional connectivity analyses greatly enhance our scientific understanding, their applicability in clinical settings may be limited. This limitation becomes particularly pronounced during the critical, high-acuity phase of stroke management, where swift and decisive actions regarding recanalization interventions are paramount. The strength of the approach presented in this study lies precisely in its clinical applicability, leveraging data that are readily available and can be efficiently utilized during the time-sensitive decision-making process inherent in stroke treatment. This ensures that the methodology is not just theoretically robust but also practically implementable, aligning with the real-world constraints and urgent needs of acute stroke care. A primary objective for the future is to facilitate the availability and routine individual application of this innovative approach in everyday clinical practice for stroke patients.

We showed that this method, which uses a normative connectome in predefined cortical ROIs, can significantly improve the prediction of stroke outcome compared to traditional measures of lesion size. This finding is consistent with a growing body of evidence suggesting that damage to white matter tracts plays a more important role in stroke outcome than lesion size alone (Moulton et al., 2019; Kessner et al., 2021). Further, there has been previous work on assessing stroke severity based on lesion size and location. One widely utilized approach is voxel-based lesion-symptom mapping (VLSM), which has generally been used for more specific scores such as aphasia scores (Bates et al., 2003; Park et al., 2021), but can also be applied to the NIHSS score (McCullough-Hicks et al., 2020). Other approaches have focussed on different measures, such as corticospinal lesion load, which is attributed with greater explanatory power than lesion size. A comparative study demonstrated that the extent of functional motor deficits following stroke strongly depends on the degree of lesion overlap with the corticospinal tract rather than lesion size itself (Zhu et al., 2010). Further studies confirmed the predictive value of the lesion load of the corticospinal tract for motor outcomes after stroke (Feng et al., 2015; Ito et al., 2022; Park and Ohn, 2022). In recent times, the John Hopkins University (JHU) white-matter atlas, which comprises volumetric white matter tracts, has found application. Using it as a kind of normative connectome, damaged white matter pathways were identified that could cause a specific deficit (unilateral spatial neglect) after subcortical stroke (Cha et al., 2022). Similarly, the JHU atlas could also be applied to neuronal substrates associated with subcortical aphasia in stroke patients (Kim et al., 2021). All these methods involve similar approaches that deal with the clarification or prediction of functional impairment after stroke. Our study finds its place in this context, demonstrating that the presented method can significantly improve the prediction of stroke severity.

Looking ahead, the method has the potential to better support the clinical management of stroke patients by providing a more accurate and individualized approach to predicting outcome and guiding treatment. Notably, the prerequisite for MRI is a FLAIR image, from which the corresponding T1 can be generated (Iglesias et al., 2023). This streamlined imaging protocol could reduce image acquisition time, potentially enabling an earlier start of therapy without any relevant loss of time.

In general, changes in white matter integrity have been correlated with alterations in cortical and subcortical structural connectivity in a range of other neurological conditions. The Network Modification (NeMo) Tool, for instance, has been applied to patient data across a spectrum of diseases, including Alzheimer’s disease, fronto-temporal dementia, normal pressure hydrocephalus, and mild traumatic brain injury, demonstrating the potential applicability of these analytical tools in elucidating and mapping disease-related changes in brain connectivity (Kuceyeski et al., 2013). These precedents underscore the relevance and adaptability of connectivity-based approaches and tools in investigating a wide array of neurological conditions, providing a comprehensive context for understanding the brain’s complex network structure and its implications in disease. The ability of these methodologies to provide nuanced insights into the brain’s structural and functional wiring underlines their potential in advancing our understanding of neurological diseases and refining diagnostic and therapeutic strategies.

It is imperative, however, to acknowledge the limitations of the present work. The use of a normative structural connectome has potential practical limitations. During preprocessing, the creation of manually drawn lesion masks (gold standard) is time-consuming, and the normalization of the masked T1 image and the application of the LST might fail, e.g., due to insufficient lesion segmentation or unsatisfactory lesion filling. It remains to be investigated whether these procedures can be seamlessly integrated into routine clinical practice. Additionally, the objective of our study was not to achieve maximum predictive precision but to explore the feasibility of the probabilistic connectome approach. While we focused on predicting the patient’s current clinical state rather than long-term outcomes, it is worthwhile to discuss the implications of this choice. Predicting the current state is one of the first steps in developing methods to predict future outcomes and can prove valuable in specific scenarios, such as when direct clinical information is unavailable (e.g., due to sedation). However, predicting long-term outcomes is arguably more clinically valuable, requiring an understanding of how acute cerebral disconnections reflect the current state and might forecast future outcomes. Another limitation of our study pertains to the inclusion criteria, focussing on motor-impaired patients, potentially influencing the interpretation of results. By concentrating on this patient group, the precision of lesion size correlation with clinical status might be artificially enhanced compared to a more diverse patient cohort. This specificity in patient selection might also exaggerate the correlation between lesion size and clinical status, as evidenced by discussions regarding the limitations of lesion size as a predictor (Price et al., 2017). It is generally recognised that clinical impairment does not solely stem from the structural lesion but arises from disruptions in the intricate network of brain functions. The conceptualization of brain function as a network operation has propelled multiple studies to utilize network information for outcome prediction. This approach has significantly outperformed models based solely on lesion size and location (Koch et al., 2016, 2021; Kuceyeski et al., 2016). The efficacy of existing acute stroke treatments has also been corroborated through these studies. For instance, systemic thrombolysis has been associated with reduced global network disruption and preserved structural connectivity (Schlemm et al., 2022). It is important to note that our study did not delve into the comprehensive analysis of the entire complex network. This broader exploration, encompassing network measures (Sporns and Zwi, 2004; Bassett and Sporns, 2017; Caliandro et al., 2017) and other techniques like network-based statistics (Zalesky et al., 2010), was not the primary focus of our research. Furthermore, our method was tested on a homogeneous subpopulation comprising single subcortical ischemic stroke lesions. This is in contrast to the great heterogeneity of the overall population of stroke patients to be examined. Consequently, questions arise regarding the extent to which our approach can be extrapolated to cortical lesions and how these can be compared with subcortical lesions. An additional limitation concerns the mirroring of lesion projections in the processing of MRI scans. While this ensures uniform analysis of effects, it disregards the interhemispheric variability inherent in the human brain.

The limitations can be overcome by combining this structural connectome-based approach with longitudinal clinical data. It is also advisable to compare our results with functional connectivity data obtained from stroke patients. Further studies with larger sample sizes should aim to encompass more diverse cohorts, encompassing both cortical and subcortical ischemic strokes. This enhancement of previous knowledge can provide a way for clinicians to apply and intuitively understand additional potential predictors, such as disconnection metrics, in stroke diagnostics and prediction models.

Our study demonstrates the viability of leveraging probabilistic connectome data to forecast the clinical status of stroke patients effectively. The key advantage of this methodology is its reliance on data that are already routinely gathered in acute stroke care settings. This enables the generation of predictions during the critical phase when initial decisions about reperfusion strategies are being made. As a result, our approach holds the potential to identify patients at a heightened risk for adverse outcomes early on, thereby facilitating timely interventions aimed at enhancing their recovery prospects. Moreover, our method offers the possibility of refining rehabilitation strategies by pinpointing the specific brain regions affected by stroke. Consequently, our findings have important implications for the clinical management of stroke patients.

Further research endeavours should prioritize the validation of our method in larger and more diverse datasets. In addition, exploring the utility of this approach in predicting various dimensions of stroke outcomes, including functional independence and quality of life, is warranted. In future, integrating this information with clinical features and biomarkers in a machine learning approach could lead to the breakthrough of predictive outcome modeling into the realm of clinical decision making.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by the Ethics Committee, University Hospital Jena, Germany. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and institutional requirements.

KW: Writing–review and editing, Writing–original draft, Visualization, Validation, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. CK: Writing–review and editing, Writing–original draft, Visualization, Validation, Supervision, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. SB: Writing–review and editing, Supervision, Software, Data curation. FW: Writing–review and editing, Data curation. MS: Writing–review and editing. DG: Writing–review and editing, Data curation. TM: Writing–review and editing, Data curation. FG: Writing–review and editing, Data curation. UT: Writing–review and editing, Data curation. CG: Writing–review and editing, Writing–original draft, Visualization, Validation, Supervision, Software, Resources, Methodology, Investigation, Formal analysis, Data curation, Conceptualization.