Introduction
Predicting response to treatment strategies for stroke patients prior to therapy can aid the clinical decision making process and improve patient outcome. Algorithms combining multiple MRI parameters accurately predict tissue infarction in patients imaged <12 h from symptom onset who did not receive thrombolysis 1 . This study extends application of these algorithms to patients imaged at the hyperacute stage who were treated with thrombolytic therapy.
Methods
Acute stroke patients who received diffusion-weighted (DWI) and perfusion-weighted (PWI) MRI within 6 h of symptom onset and a follow-up after 5–8 days (F/U) 2 were retrospectively analyzed. Patients either received standard medical treatment (Group 1; n=12), or thrombolysis (Group 2; n=29). All thrombolysed patients were imaged prior to drug administration. Apparent diffusion coefficient (ADC), T2-weighted images, isotropic DWI, CBF, CBV, mean transit time (MTT) and transit delay maps were calculated, coregistered, normalized with respect to contralateral normal white matter values and used to train a predictive algorithm which outputs infarction risk on a voxel-wise basis 1 . Model parameters were trained using regions of infarcted and non-infarcted tissue outlined on the F/U. Model 1 was trained from all Group 1 data. Model 2 was developed using Group 2 data and applied using jackknifing. Predicted lesion volumes (PLV) were defined as tissue with > 50% infarction risk. PLV were compared to the measured lesion volumes (MLV) that had been used for training. Patients exhibiting complete reperfusion on F/U according to modified TIMI criteria 2 were classified as reperfusers and others as non-reperfusers.
Results
Absolute differences between PLV and MLV were larger for Model 1 (105±57 cm3) than Model 2 (78±49 cm3) (p<.001). The calculated infarction risk was greater for Model 1 than Model 2 (p<.001). Model 2 infarction risk was lower for reperfusers than non-reperfusers (p=.04). Fig 1 shows an example where, using the same input data, significantly lower infarction risk is predicted if the patient were to receive thrombolysis.
Example input and output for a patient imaged 4.5 h after symptom onset.
Discussion
The results show that infarction risk on a voxel-wise basis is predicted to be reduced by thrombolysis, demonstrating the potential of these algorithms for prospectively identifying effect of a therapeutic intervention. In addition, statistical algorithms may provide an objective measure for identifying patients most likely to respond favorably to intervention. Furthermore, the spatial heterogeneity of predicted risk values likely reflects varying degrees of existing tissue injury and salvageability within the PLV.