Recursive self-feedback enables target language production in aphasia with pathological language mixing

Abstract

Stroke in bilingual adults can lead to aphasia with pathological language mixing (PLM), a rare neurological condition with limited treatment options. Recursive self-feedback (RSF) offers a self-directed approach, allowing individuals with PLM to iteratively self-monitor and correct their language mixing independently. We demonstrated the effects of RSF in a 59-year-old female bilingual speaker of Spanish and English (Patient P3) with severe Broca’s aphasia and PLM, 67-months poststroke. Over 14 intensive RSF sessions (2 hours/day, 5 days/week, for 3 weeks), P3 used a mobile app to self-correct her mixing errors and increase the proportion of English words in her spontaneous speech. Post-treatment assessments showed improvements across both treated and untreated narrative prompts, with an 11-point increase in her Aphasia Quotient on the English version of the Western Aphasia Battery-Revised. These findings demonstrate that RSF enables target language production, yielding clinically meaningful language recovery in P3 without external cueing, modelling or feedback. RSF is a promising procedure for aphasia therapy, opening new pathways for targeting language access, control and recovery in aphasia and PLM.

BACKGROUND

Stroke-related damage to language and cognitive-control brain networks in bilingual adults can disrupt suppression of a nontarget language and cause deficits in selection and production of a target language, leading to aphasia with pathological language mixing (PLM)—a rare neurological condition (Abutalebi et al., 2000; Mooijman et al., 2025). PLM is characterized by unintentional and involuntary language mixing in an utterance due to impaired inhibitory control mechanisms in the bilingual brain (Abutalebi et al., 2000). PLM is distinct from code-switching which is the intentional alternation of two or more languages in an utterance or during conversation (e.g., Cheng & Butler, 1989). The nature and source of PLM remain a topic of debate. One perspective attributes PLM to impaired cognitive control mechanisms, which disrupt appropriate language selection and inhibition (e.g., Abutalebi et al., 2000; Abutalebi & Green, 2007; Kong et al., 2014). An alternative view proposes that in many, albeit not all, instances language mixing functions as a compensatory strategy for impaired lexical access, whereby bilinguals with aphasia insert words from the non-target language to circumvent lexical access failures—failures that may be further intensified by the severity of aphasia (e.g., Bihovsky et al., 2023; Goral et al., 2019; Lerman et al., 2019).

Regardless of the source of PLM, there is a critical need for interventions that directly target disrupted language control and impaired lexical access mechanisms in aphasia. Recursive self-feedback (RSF) is a self-directed procedure that enables individuals to monitor and refine their own language output independently over time without relying on external cues, models, or feedback (Imaezue et al., 2023, 2024). Here, our idea is if PLM reflects weakened cognitive control (e.g., Kong et al., 2014), RSF may help strengthen internal monitoring and selection systems. Alternatively, if PLM is a function of impaired lexical access (e.g., Goral et al., 2019), RSF still offers a structured path toward more intentional, goal-directed target language improvement (Imaezue et al., 2023). In both frameworks, RSF addresses the core challenge of PLM—disrupted language access or control—by enhancing metacognitive awareness and internal feedback processes essential for correcting PLM in everyday communication.

RSF is grounded in recursive functional learning—a learning procedure through which use-dependent plasticity is driven by self-feedback loops that optimize engaged linguistic and cognitive subsystems during task performance (Imaezue et al., 2023). RSF may promote language control and recovery in PLM by refining performance through primary and compensatory feedback processes, potentially restoring function in the impaired cognitive-linguistic control subsystems. RSF has been shown to improve both scripted-sentence production (Imaezue et al., 2023, 2024) and spontaneous speech (Imaezue, 2024a) in individuals with chronic mild to moderate nonfluent aphasia. This study extends prior work by examining whether RSF facilitates target language production in aphasia with PLM.

METHOD

Study Design

We used a single-subject experimental design, collecting two baseline measures of elicited spontaneous speech at Timepoints 1 (T1) and 2 (T2), one week apart. Baseline performance on the English and Spanish versions of the Western Aphasia Battery-Revised (WAB-R: Kertesz, 2006) was assessed at T1 and T2, respectively, before treatment. Task-specific post-treatment assessments of spontaneous speech were conducted within a week after treatment at Timepoint 3 (T3), while the English WAB-R was administered two weeks post-treatment at Timepoint 4 (T4).

Participant

Patient P3 is a 59-year-old right-handed female who is a bilingual speaker of Spanish and English. She is 67-months poststroke with chronic aphasia in both languages, with PLM as per her clinical report. P3 was born and raised in a Spanish-speaking community, with Spanish as her native language (L1). She started learning English (L2) at age 19, immigrated to the United States in 2003 at age 37 and attained high proficiency at age 38. Prior to the stroke, she used English as her dominant language consistently across multiple domains, including at home and in her workplace. At the time of the study, she resided in the United States in an English-dominant household. Taken together, these factors suggest that although Spanish was her early-acquired first language, English, her late-acquired second language, later became her dominant language. She maintained high proficiency in both languages in adulthood prior to her stroke (See Supplementary Materials)

In 2019, P3 sustained a left middle cerebral artery aneurysm with subarachnoid hemorrhage which impaired her ability to communicate in both languages, as confirmed by her clinical report. Since she moved to the United States, she has lived in an English-speaking household but complained about experiencing PLM poststroke, unintentionally combining English and Spanish words in her utterances and during conversations (see Supplementary Material). For instance, P3 consistently and unintentionally mixed both languages in her utterances, even when explicitly instructed to respond in only one language or when interacting with research team members who informed her that they did not speak Spanish. We provide a transcript of P3’s baseline response to the prompt, ‘What do you do to keep your place clean and tidy?’ at T1 to illustrate the nature of her PLM prior to treatment. Note, the words in parentheses are English translations of the Spanish words.

“Okay. A very person clean. I persona (person) clean freak en (in) chicken the kitchen el (the) /den/ eh bueno (good) del (of the) passe (pass) de (of) nada (nothing) the house clean. Pero (but) eso (that) d es muy (is very) horrible. I clean poquito (a little) poquito (a little) algo (something) perfecto (perfect). Esto (this) me gusta (I like) clean. Ya (that’s it).”

P3 aimed to minimize her PLM and increase the use of English words in her utterances. This goal shaped the clinical target of the study: improving the proportion of English words in her utterances.

P3 provided signed written consent forms approved by the Institutional Review Board of the University of South Florida (USF IRB ID: STUDY006201) before participating in the study. This study is part of an ongoing clinical trial on spontaneous speech with RSF registered with ClinicalTrials.gov (NCT06323629). P3 did not receive financial compensation for her participation.

Language and Cognitive Assessments

Background information, including demographics and medical history, was obtained via a self-report questionnaire. Baseline aphasia severity and type were assessed using the WAB-R in both English and Spanish, according to which P3 was diagnosed with severe Broca’s aphasia in both languages (Table 1). Using a translanguaging approach (Goral et al., 2024), P3’s Spanish WAB-R Aphasia Quotient (AQ) was substantially higher when responses in both languages were included (L1 AQ + L2 responses = 49.2) compared to Spanish-only responses (L1 AQ = 33.7). A similar pattern emerged on the English WAB-R, with a higher score when responses in both languages were counted (L2 AQ + L1 responses = 37.7) than when only English responses were considered (L2 AQ = 32.3). This suggests that P3’s language mixing was bidirectional, with intrusions occurring from each language into the other. Post-treatment generalization to English was measured by P3’s English responses to the English WAB-R. Assessments were conducted by author CD, a certified bilingual (Spanish-English) speech-language pathologist.

Table 1.

Participant Responses on the Western Aphasia Battery-Revised (WAB-R) Subtests and Cognitive Linguistic Quick Test Plus (CLQT+) during Pre- and Post-Treatment Assessments

WAB-R and CLQT+ Subscales	Baseline Score/Max Score	Post-treatment Score/Max Score
English Version: WAB-R
Fluency and Grammar	2/10	4/10
Spontaneous Speech	2/10	5/10
Auditory comprehension	5/10	5/10
Repetition	1.8/10	2/10
Naming	4.5/10	5.7/10
Aphasia Quotient	32.3/100	43.2/100
Spanish Version: WAB-R
Fluency and grammar	3/10	N/A
Spontaneous speech	2/10	N/A
Auditory comprehension	5.85/10	N/A
Repetition	1.8/10	N/A
Naming	4.5/10	N/A
Aphasia Quotient	33.7/100	N/A
English Version: CLQT+
Attention	194/215	N/A
Memory	72/185*	N/A
Executive Function	31/40	N/A
Language	14.5/37*	N/A
Visuospatial skills	98/105	N/A
Clock drawing	11/13	N/A
Composite Severity Range	2.8/4.0	N/A

Note. WAB-R = Western Aphasia Battery-Revised; CLQT+ = Cognitive Linguistic Quick Test Plus; N/A = Not administered; Asterisks (*) = impaired cognitive-linguistic subsystem as measured by CLQT+. No post-treatment measures were conducted for Spanish Western Aphasia Battery given that English was the target language

Pre- and post-stroke language proficiencies were evaluated using the Language Experience and Proficiency Questionnaire (LEAP-Q; Marian et al., 2007), revealing high pre-stroke proficiency in both languages and significant post-stroke impairment (see Supplementary Material). The Bilingual Switching Questionnaire (BSWQ; Rodriguez-Fornells et al., 2012) indicated frequent unintentional language switching post-stroke. Patient P3 completed both self-reports—LEAP-Q and BSW—with help from her spouse. During assessments, P3 complained about her language mixing and indicated her goal of reducing the intrusion of Spanish words when speaking English. Overall, results from the assessments, observations and clinical report suggest that Patient P3’s PLM may not be an impairment of lexical access but of language control.

Cognitive abilities were assessed using the English version of the Cognitive Linguistic Quick Test-Plus (CLQT+; Helm-Estabrooks, 2017). CLQT+ results indicated overall mild cognitive impairment, with executive functions and attention within normal limits but significant memory deficits (Table 1).

Treatment

The treatment involved spontaneous speech with RSF (Imaezue, 2024a). P3 received the treatment at home through a custom mobile app and a smartphone (Figure 1A). The app fully automated the treatment sequence, switching between text-displayed prompts, speech recording at 44.1kHz sampling rate and immediate speech playback. The app’s interface was simple and intuitive, allowing for a seamless user experience (Figure 1A).

Figure 1:

Panel A shows treatment device (smartphone) and app user interface. Panel B shows schematic of spontaneous speech with recursive self-feedback.

During the spontaneous speech with RSF treatment, P3 followed a structured process, as follows. First, she read and listened to a narrative prompt in English (e.g., What do you like to do to celebrate Thanksgiving?). Then, she pressed the app’s “Record” button to capture her initial one-minute response (Figure 1A), ensuring a sufficient speech sample (Imaezue, 2024a). After completing her recording, the app immediately played back her speech. Without external reference, P3 was instructed to actively listen to her previous speech attempt, evaluate unintended use of Spanish (primary treatment target) and other performance errors (e.g., incorrect word selection or word omission in English), and then minimize the errors in a subsequent attempt by responding in English alone. This process was reapplied recursively five times per prompt across all prompts, with each attempt building upon the most recent self-generated feedback (Figure 1B).

P3 was trained on the treatment app using a practice version and reported no difficulties with its use. The intervention was delivered at home via a smartphone and a noise-canceling headphone, both provided at no cost. P3 engaged in app-based practice for two hours per day, five days per week, over approximately three weeks, totaling ~28 hours across 14 days. She practiced producing 10 narrative prompts selected and adapted from a pool of 30 prompts (Imaezue, 2024a). The app automatically recorded engagement and speech samples with metadata, which were uploaded to our lab’s cloud infrastructure for real-time fidelity monitoring. Compliance was supported through biweekly check-ins.

P3 reported receiving word-retrieval therapy during the acute phase of her stroke (<6 months post-onset) but the language targeted at that time was not specified. In addition, she reported having not received any treatment targeting spontaneous speech. Also, she did not participate in any additional speech-language therapy during the course of this study.

Assessment of Treatment Outcome Measures

Spontaneous speech in English was assessed remotely via a mobile app using 10 treated and 10 untreated narrative prompts that were administered at T1 (baseline), T2 (baseline) and T3 (post-treatment). The prompts 20 (treated + untreated), for eliciting spontaneous speech, were randomly selected and assigned to the treated and untreated conditions (see Supplementary Materials for the list). P3 had one minute per prompt, matching treatment conditions, and responded only once per prompt without self- or external feedback. The app automatically switched between prompts after each recording. P3 was trained on its usage with a practice version. Baseline and post-treatment English WAB-R evaluations were also conducted remotely via videoconferencing. All data were transcribed and coded for language (English and Spanish), correct information units (CIUs) and word type by author CD and she was assisted by research assistant CS who is also a bilingual English/Spanish speaker.

Using the Quantitative Production Analysis framework (Saffran et al., 1989), we analyzed treatment-induced changes in P3’s speech across the following bilingual metrics: (1) Proportion of total English words—English words (excluding repeated prompts) relative to total words in both languages; (2) Proportion of English CIUs—contextually relevant, meaningful English words (Nicholas & Brookshire, 1993) relative to total CIUs in both languages; (3) Proportion of English open-class words—English nouns, verbs, adjectives, and adverbs relative to total open-class words; and (4) Proportion of English closed-class words—English function words (e.g., pronouns, articles, prepositions, conjunctions) relative to total closed-class words across both languages. Computationally, each metric was calculated using the proportion of English target words relative to the total of English and Spanish variants per response to a narrative prompt (e.g., Proportion of English CIUs = English CIUs / [English CIUs + Spanish CIUs] × 100). The task-specific metrics incorporate both linguistic outcomes (verbal outputs in English) and nonlinguistic ones i.e., cognitive suppression and inhibition of the nontarget language (Spanish). Increase in these task specific metrics at T3 relative to T1 indicated improvement in language control and performance. Furthermore, we compared P3’s pre- and post-treatment English WAB-R AQs to establish treatment-induced generalization to holistic language recovery in English. A five points treatment-induced increase in English WAB-R’s AQ is typically used as a standard to indicate clinically significant improvement in aphasia severity (Rose et al., 2021).

Statistical Analyses

Descriptive statistics were used to summarize the results. We established post-test improvement relative to the baselines using the within-case standardized mean difference (SMD) analysis (Beeson & Robey, 2006). SMD estimates were calculated with a 95% confidence interval (CI) and interpreted using a standardized framework: 0.2 indicates a small effect, 0.5 indicates a medium effect, and > 0.8 indicates a large effect (Takeshima et al., 2014). The data was analyzed using RStudio 2024.09.0+375 Cranberry Hibiscus.

RESULTS

Overall, P3 showed improvement across all task-specific metrics for both trained and untrained narrative prompts. Figure 2 (panels A-H) illustrates P3’s performance on these measures. For the trained prompts, P3 demonstrated a strong increase in the proportion of total English words (effect size = 1.18, 95% CI). This improvement generalized to the untrained prompts, where she exhibited a moderate effect (effect size = 0.74, 95% CI). Similarly, in the trained prompts, P3 showed strong treatment-induced gains in the proportion of English CIUs (effect size = 1.30, 95% CI), English open-class words (effect size = 0.87, 95% CI), and English closed-class words (effect size = 1.77, 95% CI). For the untrained prompts, generalization effects were moderate for the proportion of English CIUs (effect size = 0.74, 95% CI) but weaker for English open-class words (effect size = 0.22, 95% CI) and English closed-class words (effect size = 0.37, 95% CI).

Figure 2:

Summary and effect sizes of task-specific outcomes for trained and untrained prompts. Panels A–D show the mean proportion of total English words, English CIUs, English open class words and English closed words across responses to narrative prompts, respectively. Errors bars represent standard deviation. Panels E–H show the same measures for untreated prompts. Note. CIUs = correct information units. Single asterisk (*) = small treatment-induced effect. Double asterisks (**) = moderate treatment-induced effect. Triple asterisks (***) = strong treatment-induced effect. The effect size is a metric for post-treatment performance relative to baseline 1 alone.

Furthermore, P3 demonstrated treatment-induced clinically significant holistic aphasia recovery in English. Table 1 shows that P3 demonstrated severe Broca’s aphasia at baseline with similar AQ scores in English (32.3/100) and Spanish (33.7/100). Following RSF, P3’s English AQ had an 11-point increase after treatment (43.3/100) with the most gains evident in the domain of spontaneous speech (3-point increase).

DISCUSSION

This study provides the first evidence that RSF may enable bilingual individuals with PLM to control and improve their use of the target language. Our findings show that RSF enabled a bilingual patient with chronic severe Broca’s aphasia and PLM to achieve significant gains on treated narrative prompts, with generalization to untreated prompts and clinically meaningful recovery in the target language—measured using the English WAB-R. These findings are consistent with our previous findings on the beneficial effects of RSF beyond treated tasks in nonfluent aphasia (Imaezue, 2024b; Imaezue et al., 2023, 2024). While the study did not examine the effects of RSF on L2 suppression during L1 production—given the patient’s clinical goal, the findings suggest that RSF may have supported language control by facilitating L1 suppression during L2 production. This finding suggests that RSF may strengthen both the cognitive and linguistic processes underlying language control and recovery, respectively, as predicted by recursive functional learning (Imaezue et al., 2023; Imaezue & Goral, 2024). In doing so, it both challenges and complements traditional models of aphasia treatment and PLM recovery, which rely heavily on external cueing, modeling, or clinician-provided feedback (e.g., Ansaldo et al., 2009; Lee et al., 2016).

Although Patient P3 exhibited severe long-term memory impairment, she demonstrated preserved attention and executive functioning within normal limits—cognitive capacities critical for error monitoring and correction. These findings suggest that RSF may rely on short-term or working memory to hold recent verbal output ‘online’ for immediate comparison and language refinement. Hence, the dynamics of RSF reflect a procedure that is optimized by relatively preserved short-term or working memory, attentional control and executive functions.

Furthermore, RSF may leverage these preserved cognitive processes during offline self-feedback to compensate for impaired real-time language control and feedback mechanisms in bilingual aphasia with PLM (Behroozmand et al., 2018). By re-evaluating and adjusting her own speech immediately after production without external input, Patient P3 may have engaged language and cognitive control networks—including prefrontal, parietal, temporal, and subcortical regions (Sani et al., 2021; Xia et al., 2024). This engagement may promote adaptive neuroplasticity, enabling not only suppression of the non-target language but also retrieval of the target language (Abutalebi & Green, 2007; Imaezue et al., 2023). These findings underscore the potential of self-directed neural recalibration, where recovery relies not only on language-specific systems but also on cognitive-executive mechanisms (Imaezue et al., 2023; Jylkkä et al., 2018; Lambon-Ralph & Fillingham, 2007). These findings provide new avenues for future neuroimaging studies to explore. While this study focuses on severe Broca’s aphasia with PLM, RSF’s self-directed approach may hold promise for a wider range of bilingual aphasia, from mild to moderate forms of bilingual aphasia with PLM (Imaezue et al., 2023, 2024). Given its accessibility and adaptability, RSF could be integrated into digital therapies that enable bilinguals to regain language control over time beyond traditional clinical settings.

While the inclusion of a comparison treatment is often used to benchmark intervention effects, our preliminary study was designed to examine the feasibility and impact of RSF as a self-directed procedure for mitigating PLM in bilingual aphasia. Future studies should compare the effect of RSF with other comparison interventions on PLM. Furthermore, this study targeted the participant’s clinical need to increase her English verbal outputs. We recommend future studies conduct bidirectional post-treatment evaluations of language control across both languages. We further recommend that future studies conduct fine-grained analyses of linguistic changes as a function of treatment and patient related variables like age of language acquisition, level of proficiency, education level etc. (e.g., Roberts et al., 2024; Schmid et al., 2018).

CONCLUSIONS

RSF is conceptually simple and language-agnostic, leveraging universal learning mechanisms to drive language control and recovery without external cueing, modeling, or feedback. This highlights its scalability and promise for transforming aphasia care. More broadly, these results point to a shift toward autonomous recovery models and open a new window for exploring the behavioral and neurocognitive mechanisms that support self-recovery in bilingual individuals with aphasia. Future research should further validate these findings and explore broader applications in rehabilitation of communication disorders.

Highlights.

• A bilingual adult with severe Broca’s aphasia and pathological language mixing improved her target language production using recursive self-feedback (RSF).

• RSF produced clinically significant and generalized improvements in target language production without external cueing, modeling or feedback.

• RSF offers a scalable, language-agnostic approach to autonomous language control and recovery, supporting accessible, patient-driven aphasia therapy.