Comparing language outcomes in monolingual and bilingual stroke patients

Abstract

Post-stroke prognoses are usually inductive, generalizing trends learned from one group of patients, whose outcomes are known, to make predictions for new patients. Research into the recovery of language function is almost exclusively focused on monolingual stroke patients, but bilingualism is the norm in many parts of the world. If bilingual language recruits qualitatively different networks in the brain, prognostic models developed for monolinguals might not generalize well to bilingual stroke patients. Here, we sought to establish how applicable post-stroke prognostic models, trained with monolingual patient data, are to bilingual stroke patients who had been ordinarily resident in the UK for many years. We used an algorithm to extract binary lesion images for each stroke patient, and assessed their language with a standard tool. We used feature selection and cross-validation to find 'good' prognostic models for each of 22 different language skills, using monolingual data only (174 patients; 112 males and 62 females; age at stroke: mean = 53.0 years, standard deviation = 12.2 years, range = 17.2-80.1 years; time post-stroke: mean = 55.6 months, standard deviation = 62.6 months, range = 3.1-431.9 months), then made predictions for both monolinguals and bilinguals (33 patients; 18 males and 15 females; age at stroke: mean = 49.0 years, standard deviation = 13.2 years, range = 23.1-77.0 years; time post-stroke: mean = 49.2 months, standard deviation = 55.8 months, range = 3.9-219.9 months) separately, after training with monolingual data only. We measured group differences by comparing prediction error distributions, and used a Bayesian test to search for group differences in terms of lesion-deficit associations in the brain. Our models distinguish better outcomes from worse outcomes equally well within each group, but tended to be over-optimistic when predicting bilingual language outcomes: our bilingual patients tended to have poorer language skills than expected, based on trends learned from monolingual data alone, and this was significant (P < 0.05, corrected for multiple comparisons) in 13/22 language tasks. Both patient groups appeared to be sensitive to damage in the same sets of regions, though the bilinguals were more sensitive than the monolinguals. media-1vid1 10.1093/brain/awv020_video_abstract awv020_video_abstract.

Keywords: aphasia; bilingualism; language; prognosis; stroke.

Figure 1

Patient data. (A) Frequency maps of the two patient groups’ lesions. Two lesion frequency maps in standard (MNI) space, with sagittal and coronal slices centred at x = −21 mm, y = −2 mm, z = 21 mm: the map for the monolingual group is on the top and the map for the bilingual group is at the bottom. (B) Histogram of the differences between the bilingual patients’ language scores in their (non-English) native language and in English. L1 scores were available in 7 of the 22 language assessments considered in the original analyses. The legend indicates both the names of the tasks and the numbers of scores available for comparison in each task. To support comparison across the language tasks, all differences (native language score minus English language score) were standardized to the same range: negative differences indicate that the patient’s language score was better when tested in English than when tested in their own (non-English) native language.

Figure 2

Frequency of the regions implicated by our best prognostic models for all (22) language tasks. Because our patient population was restricted to those with left hemisphere stroke only, we only considered regions in the left hemisphere of the brain.

Figure 3

Predictions and prediction errors, by patient group, in tasks where significant group differences were observed. Top: Scatter plots of the predicted versus actual scores in each of the 13(/22) tasks where significant differences were observed in Table 3; predicted and actual scores are equal along the red line in each case (i.e. perfect predictions would fall along this line). Note that the predictions for the bilingual group (top right) tend to fall above the red line, which means that predicted scores tend to be higher than actual scores in these tasks. Bottom: Histograms of the prediction errors for predictions made in each of the same 13 tasks; the distribution for the monolingual group is centred close to zero (mean = −0.018), whereas the distribution for the bilingual group is positive (mean = 4.07).

Figure 4

Frequency map of regions where both patient groups have strong associations between lesion load and one of the 13 critical language tasks. The frequency of each region (max = 11) refers to the number of tasks (/13) where both patient groups displayed strong evidence of a correlation between task score and lesion load in that region.