Stroke is a leading cause of morbidity and mortality in the United States, and despite the development of thrombolysis and endovascular therapy (EVT) for patients with acute ischemic stroke (AIS), up to half of patients still experience poor clinical outcomes (1–6). Recent randomized trials have shown us that our traditional paradigm to predict who will return to functional independence after AIS is incomplete, and our understanding will now need to go beyond “time is brain” and existing estimates of infarct size on pre-treatment imaging (5–12).

Most stroke centers utilize non-contrast computed tomography (NCCT) to triage patients with ischemic stroke, and a common marker to estimate early infarct severity is the Alberta Stroke Program Early CT Score (ASPECTS) (13, 14). Modern trials have shown us that reperfusion therapies can benefit some patients regardless of how low the ASPECTS is, and there are also a large proportion of patients who do poorly despite having a good ASPECTS (8–10, 15–17). Furthermore, ASPECTS has wide inter-rater variability, and unfortunately in many settings there is limited access to neuroradiology expertise (18, 19). CT perfusion (CTP) imaging has also been used for treatment selection; however, CTP is resource intensive, only effective in a narrow set of circumstances, and can otherwise be plagued with overestimation of the infarct core and inability to identify many infarcts (20–22).

Therefore, we aimed in this study to evaluate net water uptake (NWU) as a new NCCT biomarker that could be automated and highly reliable to predict post-stroke outcomes. NWU is a measurement of brain injury and edema based on the degree of hypoattenuation in the stroke area compared to contralateral normal tissue (23). The equation to calculate net water uptake is Net Water Uptake (%) = 1 - (Density_ipsilateral / Density_contralateral) × 100. NWU is a tissue-level measurement and more granular than traditional imaging scores which are evaluated at the subject level. Early studies have shown NWU reliably predicts malignant cerebral edema and poor outcomes after AIS (24–28). Here, we used a prospectively collected registry cohort to evaluate our recently developed automated algorithm that calculates NWU in the primary regions of the anterior circulation territory on NCCT after image registration and segmentation (see Figure 1). We hypothesized that automated NWU will have non-inferior performance compared to neuroradiologist-assessed ASPECTS when predicting post-stroke functional outcomes.

Among 402 patients with AIS, median age was 69 [IQR 57–80], 49.3% were female, median NIHSS was 11 [IQR 5–19], and median pre-morbid mRS was 0 [IQR 0–1] (see Table 1). In addition, 67.2% had a large vessel occlusion, median ASPECTS was 9 [IQR 7–10], 44.3% received tPA, 39.1% received endovascular therapy, and median 90-day mRS was 3 [IQR 1–5]. All 402 subjects had successful automated NCCT image processing, and the median time to perform NWU calculations was 87 s [IQR 77–95].

In multivariable logistic regression, lower ASPECTS and higher average NWU was associated with greater likelihood of poor functional outcome measured by 90-day mRS (OR 0.84 [CI 0.74, 0.95] and OR 1.14 [CI 1.02, 1.26], respectively). See Table 2 for full results. In ROC analysis, the ASPECTS-based model performed the same as the NWU-based model when classifying 90-day mRS outcome, with AUROC 0.732 and 0.749, respectively, (p = 0.513 with Delong test, see Figure 2).

In the secondary analysis, the ASPECTS and NWU models showed varying levels of performance to predict 90-day mRS outcome among different clinically relevant subgroups (see Table 3). The NWU-based model had excellent performance when classifying 90-day mRS outcome for patients with large infarct core at presentation defined as ASPECTS 0–5 (AUROC 0.863). In addition, the models seemed to perform better for the subgroup presenting in the very early time window (less than 3 h from last known well) compared to the late time window. Overall, the differences in performance between subgroups that were seen by the ASPECTS-based model were mirrored by the NWU-based model.

Among the secondary outcomes, average NWU showed a significant association in multivariable logistic regression with several individual neurologic deficits including language impairment, visual impairment, severe dysphagia, arm and leg motor weakness, and LOC deficit at hospital discharge (see Supplemental Table 1). The NWU and ASPECTS-based models performed similarly, and this performance was consistently excellent or very good based on AUROC (see Figure 2 and Supplementary Figure 1). For example, when predicting language impairment at discharge, the NWU-based model showed an AUROC of 0.787 versus the ASPECTS-based model AUROC of 0.776 (p = 0.639 with Delong test).

When examining the association between individual brain regions and specific neurologic deficits, logistic regression showed that most of the brain regions showed consistently significant predictive power for the precise deficits except for the caudate and M4 region (see Figure 3 and Supplemental Figure 2). Particular deficits were not isolated to certain brain regions, but rather NWU in almost any region showed significant association to each precise outcome. Lastly, four case examples are shown in Figure 4 to demonstrate some common clinical scenarios and the resulting ASPECTS and NWU findings.

In this cohort study of over 400 patients with anterior circulation acute ischemic stroke, we demonstrate the utility of a novel imaging biomarker, net water uptake, that quantifies hypoattenuation in brain regions on NCCT in a fully automated manner. We found that the performance of this marker, which can be calculated within 2 min, is equivalent to expert neuroradiologist-defined ASPECT scoring, across the entire cohort as well as select crucial subgroups.

One of the first clinical studies of NWU investigated its use as a “tissue-clock” to predict ischemic stroke time window based on non-contrast CT (23). Studies have also demonstrated that NWU is useful for predicting midline shift and the presence of malignant cerebral edema to potentially inform the need for decompressive hemicraniectomy (24, 27, 28). Some of these early studies have relied on advanced imaging for segmentation of infarct core prior to calculating NWU in the predicted core or region-of-interest, so we sought to develop an agnostic approach that calculated NWU in 10 standard regions using NCCT alone (23, 24, 27, 28, 38). In addition, some studies have relied on commercial software from biotechnology companies which limits widespread use compared to this open-access approach (26–28). Lastly, instead of predicting imaging outcomes, our study focused on the prediction of long-term clinical outcomes that can help guide treatment decision-making and patient expectations. This study demonstrated that automated NWU is highly reliable as an independent predictor of these clinical outcomes and performs equally to the ASPECT score which requires a subjective visual assessment by a trained neuroradiologist or neurologist. NWU provides a granular measurement of tissue injury and edema which adds new quantitative information beyond the 0 to 10 scale of the ASPECTS. We suspect this imaging marker is quantifying localized edema from irreversible ischemia and early immunological response (38–41). Not only could NWU be studied in future trials to improve treatment selection for thrombolysis and EVT, but it could also prove to be useful to stratify patients for new investigative treatments such as neuroprotectants and immunomodulatory agents.

In the secondary analysis, we observed that NWU and ASPECTS generally have a stronger association with 90-day outcome for patients in the very early time window and also those who are presenting with a large infarct core. This study was not powered to identify differences in biomarker performance among these subpopulations, so future studies will be focused on evaluating the utility of NWU to triage these difficult cases. For example, recent randomized controlled trials showed a benefit of endovascular therapy for patients with ASPECTS in the range of 0 to 5, suggesting the scale can no longer be relied upon for EVT decision-making (8–10). Automated NWU could not only fill this gap, but it could be accomplished with freely available software and only non-contrast CT.

Clinical scales like the mRS and the NIHSS are not always reflective of how patients rate their own disability after stroke, are biased toward motor disability, and are not fully reliable to gage long-term disability (42–47). We anticipate that a more personalized approach is required in the future of stroke care, and in this study NWU has also shown strong association with individual neurologic outcomes that impact a patient’s quality of life and daily activities. The NWU from individual regions of the brain were shown to have a strong association with precise neurologic deficits such as language impairment, visual impairment, and severe dysphagia. Furthermore, these secondary outcomes were assessed at time of hospital discharge instead of long-term follow up, because 90-day outcomes are influenced by many factors that are not directly stroke related including insurance status, resource availability, and systemic disparities. For example, recent studies have demonstrated that lower rates of acute and post-acute treatments were observed in Black patients with stroke compared to their White counterparts (48–51). Our study demonstrated a similar finding in that there was a significant difference in rate of good 90-day mRS among Non-Hispanic Black patients compared to White patients (Table 1), but there was no significant difference in disability between races at the time of hospital discharge (26% vs. 20% respectively, p = 0.117). Although this study was not focused on this research question, the finding suggests that there may be disparities in post-hospital stroke care that affect long-term recovery and should be further investigated in future studies.

This study has limitations. Because of the observational and retrospective nature of this cohort study, the lack of randomization can introduce confounding factors that are not fully accounted for in the baseline comparison in Table 1. The primary analysis did not include all possible co-variables, so there is potential for residual confounding. On the other hand, the population is representative of commonly seen cohorts with AIS in both real-world practice and prospective randomized trials (1, 4, 7). Also, the study cohort only sampled from a single geographic region, a large metropolitan area in the southern United States (Houston, TX), so the findings will need to be replicated in a geographically diverse future study. Additionally, the study population contained both patients with and without LVO, yet the novel imaging marker was still strongly associated with the outcomes in subgroups. As an automated and easy-to-use NCCT triage tool, NWU could be widely applicable among patients with AIS, however future trials can investigate its potential benefit for treatment decision-making among sub-populations and for specific reperfusion or neuroprotectant therapies. When compared to ASPECTS, automated NWU is invulnerable to subjectivity and inter-rater variability, provides a new degree of nuance to NCCT evaluation, and could be used even in settings where vascular neurology and neuroradiology expertise are not available.

In summary, we found that a fully automated NWU assessment provided quantitative evaluation of ischemia equivalent to expert neuroradiologist-assessed ASPECT scoring when predicting clinical outcomes. Because of its ease of acquisition and quantified outputs that are not subject to inter-rater disagreements, NWU may serve as a useful tool for clinical practice and upcoming clinical trials.