Blood–brain barrier profile pretreatment is associated with hemorrhagic transformation after endovascular reperfusion

Abstract

Background

While advances in endovascular thrombectomy (EVT) have led to high reperfusion rates, most patients treated with EVT do not avoid disability. Post‐reperfusion hemorrhagic transformation (HT) is a potential target for improving outcomes. This study examined pretreatment blood–brain barrier (BBB) disruption in tissue that would subsequently become part of the final infarct to evaluate its role in post‐EVT HT.

Methods

This post hoc analysis of the FRAME study, which enrolled patients with anterior large vessel occlusion who received EVT within 6 hours of onset, included patients if they had successful pretreatment MRI perfusion weighted imaging (PWI) and underwent successful EVT. BBB disruption was measured as the percent signal change due to gadolinium leakage on the PWI source images prior to thrombectomy. Mean permeability derangement (MPD) was defined as the average of all voxels in the stroke core that are two standard deviations above normal. The primary outcome was hemorrhagic transformation with parenchymal hematoma (PH).

Results

In total, 164 patients were included; mean age was 71 and 48% were female. PH occurred in 57 patients. Median MPD was 13.5% for patients with PH versus 3.6% for patients without (p < 0.0001). Elevated MPD was independently associated with PH with a 20% increased risk of PH for each 5% increase in MPD (OR 1.206; 95% CI 1.037:1.405; p = 0.0147, adjusted for NIHSS and procedure duration).

Conclusions

Even in patients who are successfully recanalized in an early time window, pretreatment BBB disruption in regions that go on to infarct is associated with an increased risk of post‐EVT HT.

Introduction

Over the past 10 years, the treatment of acute ischemic stroke (AIS) of the anterior circulation due to large vessel occlusion (LVO) has advanced from early¹ to late window,^2,3 and now to large core.⁴ These advances have been driven by improved devices for delivering endovascular thrombectomy (EVT) such that recanalization of the occluded vessel is achieved in greater than 80% of patients.⁵ However, despite high rates of technical success with EVT, less than half of patients retain functional independence. Thus, although most patients with LVO may now be eligible for EVT, most of those who are treated will not avoid disability.

Persistent deficits despite successful recanalization can be due to multiple factors. The occurrence of hemorrhagic transformation (HT) after endovascular reperfusion has been associated with poor outcome in both early⁶ and late⁷ time windows. HT is only considered symptomatic if there is an accompanying clinical deterioration; additionally, the risk of symptomatic HT is not a significant factor in patient selection for EVT as the beneficial effects of reperfusion consistently outweigh the risk of HT. Despite this, the presence of HT appears to negatively impact recovery after stroke. In the FRAME study, less than a third of the patients suffering a parenchymal hematoma (PH) had a functional recovery (modified Rankins score 0–2) at 3 months compared with an 80% recovery rate for the patients who had no HT.⁶

It is known that the risk of HT after EVT increases when there is evidence of blood–brain barrier (BBB) disruption on pretreatment imaging.^8,9 Disruption of the BBB can be detected using the same perfusion imaging used for penumbral imaging.¹⁰ More data are needed about how imaging the BBB could predict posttreatment PH so that strategies to mitigate the negative effects of HT can be developed.

The purpose of this post hoc proof-of-concept analysis of the FRAME study⁵ was to test the hypothesis that, in tissue which is successfully reperfused, regions that progress to infarct and subsequently develop PH will have demonstrated increased BBB disruption on the pretreatment imaging. Additionally, the accuracy of a prespecified threshold for BBB disruption for predicting PH was tested.

Methods

Our analysis was conducted according to the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) criteria for observational studies. The research was approved by the French Ethical Committee (Comité de Protection des Personnes Sud-Ouest et Outre Mer III). Each patient or her/his legal representative signed a written informed consent. The data supporting the study findings are available upon reasonable request.

This was a post hoc analysis of data from the FRAME study which has been described in detail elsewhere.⁵ Briefly, the FRAME study enrolled patients from two large stroke centers in France (both of which use MRI as first-line imaging for stroke triage) who presented with anterior circulation LVO and underwent perfusion weighted imaging (PWI) prior to thrombectomy. The perfusion imaging collected was not available to the treatment team and thus did not play a role in patient selection. The current study included all patients from the FRAME study who had MRI-based perfusion imaging which could be successfully processed to create blood–brain permeability images¹⁰ and achieved TICI2b or better reperfusion during EVT.

MRI protocol

Dynamic susceptibility contrast (DSC) perfusion weighted imaging (PWI) was collected on MRI scanners during a single injection of a weight-based dose of gadolinium at two centers with the following parameters: Centre Hospitalier Universitaire de Toulouse – 3 T Siemens Skyra scanner, 60 volumes, TE: 35 ms, TR: 1.8 s, flip angle: 80 degrees, resolution: 2.3 × 2.3 × 6 mm; Universite de Bordeaux – 1.5 T Siemens Aera scanner, 60 volumes, TE: 45 ms, TR: 1.8 s, flip angle: 80 degrees, resolution: 2.6 × 2.6 × 6 mm.

Blood–brain permeability imaging

The BBB analysis was performed by one author (R.L.) who was blinded to all other clinical and radiographic variables. PWI is created using dynamic susceptibility contrast MRI in which an intravascular contrast bolus is tracked during its first pass through the cerebral circulation. Although the temporal change in the recorded signal is largely due to intravascular gadolinium, in the setting of BBB disruption, leakage of gadolinium into the brain parenchyma causes the recorded signal to move in the opposite direction, creating a divergent signal due to extravascular contrast that can be isolated from the intravascular contribution and quantified to reflect BBB disruption.^10–12 The amount of signal lost due to contrast leakage is reflected as a percent change compared with normal tissue. Normal tissue is derived as an average of voxels outside of the stroke that demonstrate a reliable intravascular signal change. Thus blood–brain permeability imaging is a voxel-by-voxel quantification of signal lost due to contrast leakage.

The final infarct core, including areas of hemorrhagic conversion, was manually delineated on 24-hour post-EVT diffusion weighted imaging (DWI) and moved to the pretreatment PWI source images after co-registration. The resulting region of interest (ROI) was overlain on the blood–brain permeability images and the mean permeability derangement (MPD) was calculated. MPD is an average of all voxels in the ROI that are greater than two standard deviations above normal. This approach is used to identify regions of severe BBB disruption that are most likely to reflect BBB rupture. Previous studies have suggested that a MPD of >20% carries an increased risk of PH^8,9 which was tested prospectively in the current study.

Outcome measures

The primary outcome for the analysis was PH defined using ECASS criteria¹³ as PH type 1 or type 2 on imaging performed 24 h post-EVT. Hemorrhage grading performed by the core lab as part of the original FRAME analysis, which was qualitative and primarily based on T2*-weighted MRI (except 13 subjects (8%) for whom it was based on CT), was used in the current study. We acknowledge that MRI is more sensitive to HT to than CT imaging and will likely result in more severe grading. Our intent was to capture any space occupying HT lesion, even if relatively small, as this may have an effect on outcome.

Statistical analysis

A descriptive analysis was performed by PH status (PH1 + PH2 versus No PH) using quantity with percentage for qualitative variables and median with interquartile range (Q1; Q3) for quantitative variables. The association between MPD and PH status was examined using a univariable logistic regression model with PH status as the dependent variable and MPD as the continuous independent variable. To find potential factors associated with PH status, a multivariable logistic regression model was performed with PH status as the dependent variable and independent variables of age, history of diabetes, the use of thrombolysis, vessel occlusion site, NIHSS, initial blood pressure on arrival, onset-to-imaging time, length of procedure, baseline core volume, baseline Tmax 6 sec volume, and MPD. The final model was obtained with a stepwise forward selection method, whereby independent variables were entered into the model (beginning with the variable with the lowest p value) and were retained only if they remained associated at p < 0.05 with the dependent variable in the multivariable model. The stepwise selection process terminates if no further effect can be added to the model or if the current model is identical to a previously visited model. The discrimination of the multivariable model to predict PH was assessed using the area under the receiving operating curve (AUC) and its 95% CI. The AUC of the multivariable model including MPD was compared with the AUC of the same model but excluding MPD, using the nonparametric approach of DeLong, DeLong, and Clarke-Pearson. To find the optimal cut point for MPD that best discriminated the PH status (defined by maximizing correct classifications), a receiver operating characteristic (ROC) curve analysis was performed. Using the identified cut point of MPD (as the independent variable), a univariable logistic regression model was performed for PH status (as the dependent variable). All tests were two-sided and considered significant at α level of 0.05. Statistical analyses were conducted using SAS® Software version 9.4.

Results

Of the 218 patients in the FRAME study, 193 had successfully generated BBB maps, of which 164 achieved reperfusion and were included in this analysis (Fig. 1). The mean age was 71 and 48% were female. PH occurred in 57 (35%) patients. Table 1 shows the population characteristics separated by PH occurrence. The median MPD was 13.5% for patients with PH and 3.6% for patients without (p < 0.0001). Figure 2 shows examples of BBB maps from pretreatment imaging and the follow-up imaging for three patients with HT. Elevated MPD was significantly associated with PH (p < 0.0001) with a 25% increased likelihood of developing a PH for each 5% increase in MPD (OR 1.25; 95% CI 1.09:1.45; p = 0.0018). Increasing MPD was also associated with poor functional outcome defined as a modified Rankin score >2 at 3 months (p = 0.0002) and internal carotid artery as the occlusion site (p = 0.038).

[IMAGE OMITTED. SEE PDF]

Table 1 Population characteristics according to hemorrhagic transformation.

	No PH	PH1 + PH2	All patients	p-value*
n = 107	n = 57	n = 164
Age mean (±SD)	70 (15)	74 (12)	71 (14)	0.1126
Female sex n (%)	51 (48%)	27 (47%)	78 (48%)	0.9713
Risk factors
Hypertension n (%)	57 (53%)	41 (72%)	98 (60%)	0.0203
Hyperlipidemia n (%)	31 (29%)	17 (30%)	48 (29%)	0.9090
Atrial fibrillation (n (%)	23 (21%)	16 (28%)	39 (24%)	0.3463
Diabetes n (%)	13 (12%)	14 (25%)	27 (16%)	0.0412
Baseline NIHSS median (IQR)	15 (9–19)	19 (16–23)	16.5 (12–21)	0.0001
IV thrombolysis n (%)	80 (75%)	35 (61%)	115 (70%)	0.0750
Systolic blood pressure, mmHg median (IQR)	149 (130–160)	158 (140–178)	150 (135–170)	0.0273
Blood glucose, mmol/L median (IQR)	6.4 (5.7–7.4)	7.0 (6.4–8.4)	6.6 (5.8–7.6)	0.0037
Occlusion site
ICA n (%)	19 (18%)	23 (40%)	42 (26%)	0.0016
M1 n (%)	57 (53%)	27 (47%)	84 (51%)
M2 n (%)	31 (29%)	7 (12%)	38 (23%)
Hemisphere
Left n (%)	57 (53%)	28 (49%)	85 (52%)	0.6127
Imaging
Onset to imaging, min; median (IQR)	143 (99–221)	164 (118–263)	153 (105–231)	0.0692
Baseline ischemic core volume, mL median (IQR)	13.4 (5.7–48.2)	18.9 (13.9–64.3)	17.1 (8.1–58.3)	0.0062
Baseline Tmax 6 sec volume, mL median (IQR)	91.1 (58.9–129.8)	110.3 (86.7–151.1)	99 (67.9–137.9)	0.0091
MPD median (IQR)	3.6 (0–14.8)	13.5 (6.7–22.4)	7.3 (0–17.3)	<0.0001

[IMAGE OMITTED. SEE PDF]

Multivariable logistic regression did not find an association with PH for age, history of diabetes, thrombolysis, vessel occlusion site, blood pressure, onset-to-imaging time, baseline core volume, or baseline Tmax 6 sec volume. The variables found to be independently associated with PH were NIHSS (OR 1.72; 95% CI 1.27:2.40; p = 0.0007, per 5-point increase), procedure duration (OR 1.15; 95% CI 1.02:1.30; p = 0.0235, per 10 min increase), and MPD (OR 1.206; 95% CI 1.037:1.405; p = 0.0147, per 5% increase).

The ROC curve for predicting PH using MPD is shown in Figure 3. The area under the curve is 0.695 (95% CI 0.613:0.777). The ROC analysis confirmed the prespecified MPD threshold of 20%, identifying 19.71% as the optimal cut point to maximize correct classification, with a sensitivity of 31.6% and a specificity of 88.8%. An MPD greater than 20% more than tripled the likelihood of a PH occurring (OR 3.37; 95% CI 1.49:7.85; p = 0.004). The AUC of the ROC curve for the final multivariable model including MPD was 0.73 (95% CI 0.65–0.81), whereas the AUC of the same multivariable model but excluding MPD was 0.71 (95% CI 0.63–0.79), which were not significantly different (p = 0.38).

[IMAGE OMITTED. SEE PDF]

Discussion

Our results are the first to demonstrate that, for patients with LVO-related AIS who underwent successful EVT, evidence of severe BBB disruption on pretreatment imaging is associated with the occurrence of PH on post-treatment imaging. This study adds to our understanding of the association between BBB disruption and PH by demonstrating that even in a population which receives optimal care, with successful recanalization in an early time window, the risk of PH remains closely linked to the pretreatment BBB profile. Of note, the PH rate of 35% reported in this study reflects all PHs regardless of if they were symptomatic, as it is not expected that BBB disruption would capture this clinical distinction. The symptomatic ICH rate from the parent FRAME study was 4%.⁵

It is known from animal models that the BBB permeability progressively increases with time in the setting of sustained ischemia.¹⁴ We have demonstrated previously that in humans, mild disruption of the BBB detected prior to reperfusion can be reversible in the setting of early reperfusion.¹⁵ However, when BBB dysfunction progresses to BBB rupture, recanalization of the affected vessel carries an increased risk of PH, as demonstrated by the increased risk of bleeding with time.¹⁶

Typically, for the purposes of clinical trials, PH in the setting of reperfusion therapy is only considered relevant if it is accompanied by a concomitant clinical deterioration. This makes sense when assessing the risk/benefit profile of a treatment, but it does not capture what impact PH may have on long-term stroke outcome more broadly. PH partly results from the thrombo-inflammatory processes initiated during the acute phase of AIS,¹⁷ and contributes to post-stroke inflammation leading to increased edema¹⁸ and worse outcomes.^19,20 Thus the development of treatment strategies to mitigate or prevent the effects of PH would be desirable. Testing of such strategies would likely benefit from targeted patient selection using BBB profiles.

There are several situations when knowing the risk of hemorrhagic transformation at the time of an acute intervention can be helpful when considering add-on therapies that may themselves have a risk of PH. For instance, there is evidence that additional intra-arterial thrombolysis administered after revascularization of an LVO may improve outcome.²¹ It was recently shown that the use of antiplatelets and anticoagulants during EVT can be associated with an increased risk of HT.²² However, in some situations, such as emergent stenting, antithrombotics are often required to maintain stent patency. Other antithrombotic and thrombolytic strategies (such as cangrelor or tirofiban) are being tested to improve microvascular reperfusion and prevent erratic emboli with promising preliminary results. Expansion of their use to large and late LVO-related AIS may carry a higher risk of PH. Additionally, when performing immediate post-EVT imaging, often while still on the table in the angio suite, it can be difficult to distinguish contrast staining from PH, thus a prior knowledge of the risk of PH could be helpful.

In the current study, we used a method for identifying regions of BBB rupture that has been applied in two prior studies, one looking at tPA treated patients,⁸ and one looking at patients treated with EVT.⁹ The variable mean permeability derangement (MPD), which is the average of all voxels in the affected region that are two standard deviations above normal, identifies regions of severe BBB disruption. In both of the prior studies, a MPD of approximately 20% appeared to identify patients at risk for PH. This number was again identified in the current analysis as a likely cut point for estimating risk. However, with its modest performance on ROC analysis, this metric should likely be combined with other factors when developing models for predicting HT. In addition, it is not expected that a pre treatment BBB assessment would be able to predict technical factors related to the EVT procedure itself which may also influence the risk of HT. As the use of posttreatment imaging to guide the pretreatment analysis would not be available for a predictive model, additional studies aimed at determining the best pretreatment imaging to use are needed.

There are several limitations to this study. First, the ROI used was based on follow-up imaging. This approach was used to ensure that the ROI focused on the BBB properties of the most relevant tissue, specifically the tissue which would go on to infarct. However, such an ROI would not be available in the pretreatment setting. Future studies will need to determine what is the best ROI that can be derived from pretreatment imaging. Second, this was an MRI-based study, which may not be applicable to centers that do not perform hyperacute MRI. CT perfusion can also be used to detect BBB disruption, and while the signal changes are much smaller than seen on MRI, a CT perfusion-based model would likely have broader application. Third, although MPD was independently associated with PH in our multivariable model, the accuracy of the multivariable model (as evaluated by the AUC of the ROC curve) did not significantly differ with or without MPD. This may be explained by our moderate sample size. Fourth, some patients were excluded due to technical issues with processing the BBB images which may introduce a bias excluding patients who were unable to tolerate an MRI scan. Fifth, the number of passes during thrombectomy was not part of the parent study thus not controlled for in this analysis; however, procedure duration, which is strongly linked to number of passes, was part of the model. Last, this study focused on the early time window and the results may apply differently to a late window or a large core population. Strengths of this study include the use of multicenter clinical trial data, the prespecified threshold based on prior work, and the use of a blinded, automated BBB analysis.

Conclusions

The BBB profile of patients presenting with LVO has critical implications for their risk of HT after EVT. Despite the risk of HT being an important part of medical decision-making, analysis of the BBB is not currently being used in clinical practice. Future studies aimed to counter the negative impact of HT after EVT may benefit from assessing the risk of HT with BBB imaging.

Acknowledgments

FRAME (French Acute Multimodal Imaging Study to Select Patients for Mechanical Thrombectomy) was supported by a research grant from the French Ministry of Health, Clinical Research Hospital Program 2015 (PHRCI-15-076). The study sponsor, CHU de Toulouse, has no role in study design, collection, analysis, and interpretation of data; writing of this manuscript; or the decision to submit for publication. The corresponding author had full access to all of the data in the study and had the final responsibility for the decision to submit the report for publication. Author R.L. is supported by AHA grant 23IPA1043237 and NIH grant R01NS123386.

Conflict of Interest

Nothing to report.

Author Contributions

Richard Leigh and Jean-Marc Olivot: conception and design of the study, acquisition and analysis of data, and drafting a significant portion of the manuscript or figures. Pierre Seners: acquisition and analysis of data and drafting a significant portion of the manuscript or figures. Vanessa Rousseau and Soren Christensen: acquisition and analysis of data. Jean-François Albucher, Amel Drif, Christophe Cognard, Adrien Guenego, Alain Viguier, Agnes Sommet, Nicolas Raposo, Lionel Calviere, Anne-Christine Januel, Michael Mlynash, Fabrice Bonneville, Brice Gaudilliere, Claire Thalamas, Igor Sibon, Thomas Tourdias, Mikael Mazighi, Jeremy J Heit, and Benjamin Maier: conception and design of the study and acquisition, and analysis of data. Gregory W. Albers: conception and design of the study.

Data Availability Statement

The data supporting the study findings are available upon reasonable request.

References

Goyal M, Menon BK, van Zwam WH, et al. Endovascular thrombectomy after large‐vessel ischaemic stroke: a meta‐analysis of individual patient data from five randomised trials. Lancet. 2016;387:1723‐1731. doi:10.1016/S0140‐6736(16)00163‐X

Nogueira RG, Jadhav AP, Haussen DC, et al. Thrombectomy 6 to 24 hours after stroke with a mismatch between deficit and infarct. N Engl J Med. 2018;378:11‐21. doi:10.1056/NEJMoa1706442

Albers GW, Marks MP, Kemp S, et al. Thrombectomy for stroke at 6 to 16 hours with selection by perfusion imaging. N Engl J Med. 2018;378:708‐718. doi:10.1056/NEJMoa1713973

Sarraj A, Hassan AE, Abraham MG, et al. Trial of endovascular Thrombectomy for large ischemic strokes. N Engl J Med. 2023;388:1259‐1271. doi:10.1056/NEJMoa2214403

Olivot JM, Albucher JF, Guenego A, et al. Mismatch profile influences outcome after mechanical Thrombectomy. Stroke. 2021;52:232‐240. doi:10.1161/STROKEAHA.120.031929

Olivot JM, Heit JJ, Mazighi M, et al. What predicts poor outcome after successful thrombectomy in early time window? J Neurointerv Surg. 2022;14:1051‐1055. doi:10.1136/neurintsurg‐2021‐017946

Heit JJ, Mlynash M, Christensen S, et al. What predicts poor outcome after successful thrombectomy in late time windows? J Neurointerv Surg. 2021;13:421‐425. doi:10.1136/neurintsurg‐2020‐016125

Leigh R, Jen SS, Hillis AE, Krakauer JW, Barker PB. Stir, Investigators VI. Pretreatment blood‐brain barrier damage and post‐treatment intracranial hemorrhage in patients receiving intravenous tissue‐type plasminogen activator. Stroke. 2014;45:2030‐2035. doi:10.1161/STROKEAHA.114.005249

Leigh R, Christensen S, Campbell BC, et al. Pretreatment blood‐brain barrier disruption and post‐endovascular intracranial hemorrhage. Neurology. 2016;87:263‐269. doi:10.1212/WNL.0000000000002862

Leigh R, Jen SS, Varma DD, Hillis AE, Barker PB. Arrival time correction for dynamic susceptibility contrast MR permeability imaging in stroke patients. PLoS One. 2012;7: [eLocator: e52656]. doi:10.1371/journal.pone.0052656

Boxerman JL, Schmainda KM, Weisskoff RM. Relative cerebral blood volume maps corrected for contrast agent extravasation significantly correlate with glioma tumor grade, whereas uncorrected maps do not. AJNR Am J Neuroradiol. 2006;27:859‐867.

Zaharchuk G. Theoretical basis of hemodynamic MR imaging techniques to measure cerebral blood volume, cerebral blood flow, and permeability. AJNR Am J Neuroradiol. 2007;28:1850‐1858. doi:10.3174/ajnr.A0831

Hacke W, Kaste M, Fieschi C, et al. Randomised double‐blind placebo‐controlled trial of thrombolytic therapy with intravenous alteplase in acute ischaemic stroke (ECASS II). Second European‐Australasian Acute Stroke Study Investigators. Lancet. 1998;352:1245‐1251. doi: S0140673698080209 [pii].

Sandoval KE, Witt KA. Blood‐brain barrier tight junction permeability and ischemic stroke. Neurobiol Dis. 2008;32:200‐219. doi:10.1016/j.nbd.2008.08.005

Simpkins AN, Dias C, Leigh R. National Institutes of Health natural history of stroke I. Identification of reversible disruption of the human blood‐brain barrier following acute ischemia. Stroke. 2016;47:2405‐2408. doi:10.1161/STROKEAHA.116.013805

Olivot JM, Finitsis S, Lapergue B, et al. Parenchymal hemorrhage rate is associated with time to reperfusion and outcome. Ann Neurol. 2022;92:882‐887. doi:10.1002/ana.26478

Maier B, Tsai AS, Einhaus JF, et al. Neuroimaging is the new "spatial omic": multi‐omic approaches to neuro‐inflammation and immuno‐thrombosis in acute ischemic stroke. Semin Immunopathol. 2023;45:125‐143. doi:10.1007/s00281‐023‐00984‐6

Ng FC, Churilov L, Yassi N, et al. Microvascular dysfunction in blood‐brain barrier disruption and Hypoperfusion within the infarct Posttreatment are associated with cerebral edema. Stroke. 2022;53:1597‐1605. doi:10.1161/STROKEAHA.121.036104

Nadareishvili Z, Simpkins AN, Hitomi E, Reyes D, Leigh R. Post‐stroke blood‐brain barrier disruption and poor functional outcome in patients receiving thrombolytic therapy. Cerebrovasc Dis. 2019;47:135‐142. doi:10.1159/000499666

Leigh R, Hitomi E, Hutchison RM, Elkins J. Post‐stroke blood‐brain barrier disruption predicts poor outcome in patients enrolled in the ACTION study. Journal of Neuroimaging. 2021;31:751‐757. doi:10.1111/jon.12862

Renu A, Millan M, San Roman L, et al. Effect of intra‐arterial Alteplase vs placebo following successful Thrombectomy on functional outcomes in patients with large vessel occlusion acute ischemic stroke: the CHOICE randomized clinical trial. JAMA. 2022;327:826‐835. doi:10.1001/jama.2022.1645

van der Steen W, van de Graaf RA, Chalos V, et al. Safety and efficacy of aspirin, unfractionated heparin, both, or neither during endovascular stroke treatment (MR CLEAN‐MED): an open‐label, multicentre, randomised controlled trial. Lancet. 2022;399:1059‐1069. doi:10.1016/S0140‐6736(22)00014‐9