Precision rehabilitation for aphasia by patient age,sex,aphasia severity,and time since stroke? A prespecified,systematic review-based,individual participant data,network,subgroup meta-analysis

Effectiveness of rehabilitation interventions

Examining the effectiveness of complex rehabilitation interventions is challenging. ⁴ Aphasia interventions, for example, vary by regime (frequency, intensity, duration, and dosage), provider, delivery model, and therapeutic approach. Therapists’ clinical decision making is typically informed by findings from high-quality, group-level comparisons within randomized controlled trials (RCTs) and estimates of the average effect based on meta-synthesis of aggregate group data, which in turn inform clinical guidelines.^5,6 High-level aphasia rehabilitation evidence has, to date, been primarily based on aggregate data from high-quality RCTs and meta-synthesis,^5,7 thus limiting the degree to which it might inform precision rehabilitation decisions. ⁸

Aggregate data limitations

Aggregate data based on relatively large samples and adequate statistical power provide crucial evidence of the effectiveness of rehabilitation interventions across recruited populations. The ability of an isolated trial to determine subgroups’ differential treatment response, however, is limited.^8,9 Limitations to aggregate data meta-syntheses include the risk of outcome reporting bias, ¹⁰ restricted inclusion of language data, and the need to synthesize various clinical language assessments into a standardized mean difference^11,12, making clinical interpretation and application of emerging evidence a challenge.^9,13 In addition, meta-syntheses of aggregate data risk masking the responsiveness of different subgroups to specific interventions, known as ecological bias. ¹⁴ Aphasia meta-syntheses have presented findings of interest (and concern); for example, one meta-analysis highlighted the benefits of intensive SLT early post-onset though these findings were confounded by higher participant dropouts (n = 35/114) compared to lower-intensity SLT participants (n = 17/102; p = 0.01). ⁵ In contrast, among the RCTs which recruited participants years post-onset, neither dropouts nor significant language gains were evident. ⁵ Further examination of the potential benefits, risks, and evident variability in tolerance and responsiveness to the different SLT regimens may be worthwhile but could not be advanced until individual participants’ data (IPD) on demographics were available.

Individual participant data meta-analyses

Where strong theoretical reasons and clinical plausibility suggest differential responses to an intervention (e.g., aphasia rehabilitation), large IPD meta-analysis is the gold standard. ¹⁴ Our earlier IPD meta-analysis found that trial participants’ greatest overall and auditory comprehension language gains were associated with > 20–50 h SLT dosage, delivered 2-to-4 h each week, and between 3 and 5 + days. Greatest auditory comprehension gains were associated with 9 + SLT-h weekly over 4-to-5 days. ¹⁵ Meta-analyses using IPD also support the exploration of differential treatment response across participants,^8,16 reducing the risk of ecological bias, facilitating the inclusion of previously unreported dropouts, outcome measurement, and follow-up data while increasing statistical power.^9,10,17 In a highly heterogeneous population such as people with aphasia, investigations of subgroups’ (based on age, aphasia severity, chronicity, or sex) treatment responsiveness might be explored with greater statistical power and adjustments for confounders than at trial level.^9,14,18 Information on differential responsiveness of clinically relevant subgroups may inform the development of precision aphasia rehabilitation approaches and future RCT-based treatment evaluations

Search strategy and selection criteria

We created the international RELEASE IPD database of aphasia research to support several planned analyses,^15,19,20 systematically identifying and reviewing published and unpublished datasets with ⩾ 10 IPD on aphasia, language outcome, and time since stroke (reported in full elsewhere).^15,19,21 Briefly, several electronic databases, including MEDLINE and EMBASE, were searched (inception-September 2015 for eligible datasets). Anticipating lengthy data cleaning and analysis procedures, we also searched trial registrations for eligible trials completing beyond the electronic search date.^15,21 Non-English datasets were translated. Language data derived from stroke or screening measures were excluded. Non-randomized trials, case series, and clinical registries were included in the database and supported previous analyses ²⁰ but were excluded from this subgroup analysis.

Full text reports were reviewed independently by two reviewers; a third resolved disagreements. Potentially eligible datasets were invited to contribute IPD. One reminder was sent to non-respondents, followed by attempts to contact co-authors. We confirmed dataset eligibility with respondents prior to contribution. A protocol guided data searching, identification, extraction, and analyses (PROSPERO CRD42018110947). ²¹ Included datasets reported relevant ethical and gatekeeper approvals.

Data extraction and preparation

For this prespecified subgroup analysis, we extracted data on demography (including sex, age, and language used), stroke (time post-onset, hemisphere, and aphasia severity), SLT intervention, and language outcome (overall-language ability, auditory comprehension, and functional communication). Language recovery was defined as the change in absolute language score from baseline to first post-intervention follow-up. We checked and collated language domain measurements, as agreed a priori by the RELEASE collaborators. Whenever possible, baseline and subsequent timepoint data extraction was confirmed and additional unreported data sought from the primary researchers. ²¹

For each outcome, the most frequently used measure by dataset was identified as the anchor measure. All remaining measures of that outcome (minority measures) were transformed to match the anchor measure’s range and format, thus retaining the clinical relevance of the anchor measures’ change score. ¹⁹ Anchor measures comprised the Western aphasia battery-aphasia quotient (WAB-AQ) ²² for overall-language ability; the Aachen Aphasia test-token test (AAT-TT) scored positively for auditory comprehension; ²³ and the Aachen Aphasia test-spontaneous speech communication (AAT-SSC) rating domain score for functional communication. ²³

SLT interventions targeting language recovery were categorized by regimen (frequency, intensity, and dosage; Supplemental Material A). All data were checked by an independent researcher and, where possible, with available documentation and primary researchers. We recorded unavailable data as “unreported.” Aphasia with a non-stroke etiology, unreported time post-onset, and duplicate IPD were excluded. In the absence of SLT intervention records at IPD level, we applied group-level SLT descriptions to the IPD accordingly. Final data formatting decisions were made through collaborator discussion. Categorical data formats (e.g., 5-to-6 weeks) were recorded as means (5.5 weeks). Language co-interventions (e.g., pharmacological) were documented and IPD were included up to the point of any crossover.

IPD network meta-analysis

Drawing only RCT IPD from the wider RELEASE database (21), we conducted a network meta-analysis of SLT interventions delivered by language outcomes (15). Where traditional meta-analyses consider pair-wise trial comparisons (e.g., treatment 1 vs 2), network meta-analysis considers three or more interventions simultaneously by making direct (treatment 1 vs 2; treatment 2 vs 3) and indirect comparisons (treatment 1 vs 3), thus yielding more precise estimates than paired direct/indirect estimates and making it possible to compare the effectiveness of interventions. We used datasets as random effects and demographics and interventions as fixed effects. Data analysis for this article used SAS™ software (9.4 using PROCMIXED). Using a statistical inferencing approach, we sought to highlight important research questions, considerations for future trial design, and clinical implementation. Language recovery was defined as the mean of the absolute change from baseline to the first follow-up on the transformed standardized measure. Effect sizes were estimated and reported (95% CI). Our minimum sample size for each analysis was 20 IPD (2 RCTs).

We included prespecified potential confounders in the base model (age, sex, aphasia severity, and time post-onset) and simultaneously examined the impact of IPD and language variables on the intervention effect. Our one-stage network meta-analysis examined IPD and SLT intervention regimen variations by age, time since onset, aphasia severity at baseline and sex subgroups, and associated estimates of mean language gains from baseline (Supplemental Material A). Continuous regimen variables were grouped for comparison (e.g., 3 vs 4 SLT-days/week). We dichotomized IPD based on key demographic and clinical data; males and females, median age (working age ⩽ 65 years and older > 65 years; rehabilitation timing after aphasia onset (early ⩽ 3 months and late > 3 months); moderate–severe and mild–moderate groups based on the overall median language modality score.

We considered clustering by dataset distinguishing IPD from dataset-based interactions.^15,21 Where > 20% of a dataset variable was missing, we excluded it from that network analysis. Patterns of loss were checked; we compared missing data to demographic and other variables using the t-test or Mann–Whitney U test. In the absence of evidence of influence, we studied the data missing at random. We excluded any data not missing at random. For risk of bias and heterogeneity checking, see Supplemental Material B.

Age and language rehabilitation

Early versus late rehabilitation

Aphasia severity and language rehabilitation

Sex and language rehabilitation

Clinical implications

Greatest overall-language and auditory comprehension gains across subgroups were associated with higher dosage levels > 20–50 SLT-h or above) than current clinical provision reported for 4–16 h.^27–30 Higher dosage rehabilitation within existing resources may be challenging, requiring alternative delivery models to achieve the requisite dose, such as telerehabilitation, self-management, trained family members, prescribed home-practice tasks, and group therapy. Our findings also highlighted frequency–intensity–dosage ranges below which language gains from baseline were not observed. Some plausible, clinically relevant subgroups may benefit from precision rehabilitation approaches based on age, aphasia severity, chronicity, and sex.

Research implications

High-quality targeted SLT RCTs should be conducted to evaluate the acceptability, clinical, and cost effectiveness of precision aphasia rehabilitation approaches. Participants should be selected by age-group, time since stroke, and severity or stratifying intervention by subgroup. Despite no language restrictions, our predominantly English-speaking participant data highlighted an underrepresentation of non-English aphasia research. Minimal SLT-frequency–intensity–dosage levels should be applied to intervention development and evaluation, in addition to stratification by age, time post-stroke, and severity. Continued collaborative approaches including research data sharing will support the reduction of research waste and further insights into precision stroke rehabilitation, including SLT for aphasia.