Non-invasive brain stimulation offers a potential therapeutic advance for the treatment of PPA. Transcranial direct current stimulation (tDCS) is one particular form of brain stimulation with numerous advantages (e.g., cost, portability, comfort) in terms of its clinical application to PPA. In conventional tDCS, a low amperage current typically ranging between 1 and 2 mA is applied to the scalp using probabilistic anatomical landmarking (10/20 EEG system) to guide the strategic placement of two or more electrode sponges (i.e., montages). The head (including skull, cerebrospinal fluid, air-filled sinuses and brain) effectively closes the circuit between anodal and cathodal electrode leads. Depending on the location of the electrodes, that is the montage, electrons flowing from cathode to anode will alter cell membrane potentials of specific groups of neurons (Nitsche and Paulus, 2000). Effects of tDCS are evident across a variety of timescales ranging from an hour or more to months after stimulation. Although the dynamics of neuroplastic change induced by tDCS remain poorly understood, it appears that dose (duration and current intensity) is a powerful moderating variable (Reinhart et al., 2016). In general, it has been argued that a longer stimulation protocol distributed over multiple sessions with higher current intensity is more likely to produce durable treatment effects (Nitsche and Paulus, 2000) (also see Reinhart et al., 2016 for other factors determining the tDCS effects). It is believed that anodal stimulation produces excitability in a swath of cortex proximal underlying the position of anodal lead(s) (Nitsche and Paulus, 2000; Cuypers et al., 2013).

The ability to selectively target and optimize membrane potentials within local populations of neurons is particularly relevant to the treatment of circumscribed brain injuries. In post-stroke aphasia, for example, many investigators believe that optimal recovery of language occurs on the context of peri/ipsilesional functional reorganization (Benjamin et al., 2014; Reilly et al., 2014). That is, recovery is best when the tissue surrounding a lesion assumes function of the damaged region. tDCS may prove useful in artificially stimulating tissue surrounding a lesion (Fridriksson et al., 2011; Richardson et al., 2015) or in either upregulating or suppressing the functioning of homologous regions of the contralateral hemisphere (Vines et al., 2011; Turkeltaub, 2015). In each of these cases, stimulation via a strategically placed distribution of electrodes may aid in improving language perception and/or production (Hamilton et al., 2011).

Few neurostimulation studies of AD have been reported to date, and this lack of precedence is a rate limiting factor in the optimization of electrode montage and dose. One potential target is the left temporoparietal region (Fernandez et al., 2002) that has previously shown hypoactivity (e.g., Schliebs and Arendt, 2011). Ferrucci et al. (2008) targeted the temporoparietal region and reported a positive effect of a one-session anodal stimulation on word recognition task (i.e., if a word was presented previously), while cathodal condition worsened the performance and the sham condition had no effect. Alternatively, anodal stimulation in the left DLPFC and temporal regions (i.e., site T7 in the 10/20 EEG system (Boggio et al., 2009) or sites T3 and T4 (Boggio et al., 2012)) also produced gains in a visual memory task. In terms of language function, previous work using non-invasive brain stimulation has shown that repetitive TMS or anodal tDCS stimulation to the DLPFC can potentially improve action and/or object naming performance in AD (Cotelli et al., 2008; Fertonani et al., 2010). Thus, though more investigation is needed, neurostimulation has the potential to confer benefits in the context of diffuse neuropathologies (e.g., AD) (also see Hansen, 2012).

Although tDCS does show promise as an independent therapy, a more compelling and potentially powerful application of this technology is as an adjuvant to behavioral treatment (Crinion, 2016). That is, tDCS paired with a behavioral treatment offers the promise of synergistic effects that would not otherwise be apparent through stimulation or behavioral treatment alone (Monti et al., 2013; Price et al., 2015; Tippett et al., 2015).

Clinical criteria for PPA currently discriminate between three distinct variants (Gorno-Tempini et al., 2011). When cortical atrophy predominantly impacts the left posterior inferior frontal cortex, patients experience non-fluent/agrammatic PPA (nfvPPA), a syndrome characterized by slowed and imprecise speech output, along with agrammatism (Grossman, 2012). PPA also includes two variants with pathology that predominantly impacts the temporal lobes (i.e., logopenic variant PPA (lvPPA), and semantic variant PPA (svPPA)). Patients with lvPPA typically experience atrophy of posterior temporal and inferior parietal lobe structures and in turn present with a constellation of behavioral symptoms that include profound problems with sentence repetition. lvPPA naming difficulty has also been attributed to impaired lexical retrieval relative to frank semantic loss (Gorno-Tempini et al., 2008). In contrast, patients with svPPA experience profound impairments in word and object knowledge associated with the degeneration of ventral and anterolateral portions of the temporal lobes (Hodges and Patterson, 2007; Patterson et al., 2007; Bonner et al., 2016; Cousins et al., 2016). There seems to be relative preservation of knowledge of personal objects over the equivalent objects that are not used daily by the patient (e.g., a patient’s coffee cup versus any coffee cup) (Snowden et al., 1994; Bozeat et al., 2002; Giovannetti et al., 2006). It has been argued that the root cause of naming impairment in svPPA is the dimming of semantic memory, the substrate for word meaning.

Patients with Alzheimer’s Disease not explicitly linked to PPA also typically manifest naming impairments. However, the etiology of anomia in AD remains contested. Some have argued based on the presence of retained priming and cueing effects in AD (e.g., it has a tail, barks, and you walk it with a leash… it’s a_______) that its naming impairment reflects impaired access to knowledge (Reilly et al., 2011; Libon et al., 2013). Others have proposed that naming impairments in AD have a more fundamental root cause in degraded semantic knowledge (Hodges et al., 1992, 1996) or result from an interaction between degraded semantic knowledge and executive and perceptual processes that are crucial for lexical retrieval (Reilly et al., 2011). In the canonical presentation of AD, people experience plaque and tangle pathology that impacts a diffuse range of cortical subcortical regions (Holtzman et al., 2011). Previous neuroanatomical correlation studies do, however, demonstrate that plaque burden in AD is not uniformly distributed throughout the brain and that volumetric loss within particular regions (e.g., anterior temporal cortex (Frings et al., 2011; Domoto-Reilly et al., 2012) and lateral temporal cortex (Grossman et al., 2004)) is predictive of the severity of naming impairment. In neurotypical adults, it has been hypothesized these regions of the temporal lobe play crucial roles in conceptual representation (Reilly et al., 2016).

Speech and language deficits manifest differently across each of progressive aphasia and dementia variants (Reilly et al., 2010; Libon et al., 2013). Such variability precludes the application of a one-size-fits-all approach to language intervention and its formal assessment (e.g., Montembeault et al., 2016). For example, treatments targeting motor speech fluency are not readily applicable for patients with svPPA whose output is typically fluent. Likewise, re-training the conceptual attributes of a set of target words may prove ineffective for patients with nfvPPA who do not typically experience frank semantic impairments throughout the early stages of the disease. In the current study, we focused on treatment of progressive anomia associated with a semantically based naming impairment – progressive anomia – in a cohort of patients with fluent variant of PPA (svPPA or lvPPA) or AD.

Reilly (2016) outlined both a treatment approach and item-selection protocol for semantic anomia (i.e., semantically based naming impairment) in svPPA. Briefly, this approach involves the maintenance of a small, carefully crafted micro-lexicon consisting of 100 words. The target items include a range of functional semantic categories, including familiar people, hygiene implements, foods, clothes, places, household items, and activities. When patients begin treatment, they are randomly assigned a set of training items from fixed lists, and items from these lists that are not assigned act as controls. For example, the assignment of clothes items for explicit training may include, “shirt, socks, glasses” while “pants, watch, hat” act as untrained stimuli. In this way, we impose item rather than participant level controls. In the first stage of the treatment, patients are assigned their target items in conjunction with the primary caregiver. We then travel to the patient’s home and take digital photographs of the trained and untrained items and array all the pictures within a book organized by semantic category.

Once a suitable item pool is established, the treatment involves acquisition of baseline neuropsychological measures followed by repeated training of the target lexicon using a modification of semantic feature analysis (SFA). SFA traditionally involves asking patients either to generate or verify a matrix of conceptual attributes (e.g., what is it? where is it found? what is it used for?) for a given stimulus (Boyle, 2004; Hashimoto and Frome, 2011). The goal of SFA, when applied to post-stroke aphasia, involves re-establishing weakened links between semantic knowledge that has come untethered to lexical representations. In the current treatment, we adapted this paradigm to an error-reduced format by reading the features aloud to the patient in blocks of five items and subsequently asking the patients to name the target item and generate the features he/she heard minutes before (Reilly, 2016). This treatment is then regularly repeated as disease severity worsens with the goal of maintaining a core functional vocabulary.

Here we examined the effectiveness of anodal tDCS administered to the left temporoparietal region as an adjuvant to this semantic treatment approach in a sample of patients with progressive anomia. The left temporoparietal region had previously reported as a pivotal hub region for semantic processing (Price et al., 2016; Reilly et al., 2016). We hypothesize that targeted stimulation of this region coupled with a behavioral semantic challenge will produce synergistic and potentially lasting gains that result in improved naming.

Participants (N = 5) included four individuals (3 males, ages 55–74, mean age = 66.6, SD = 8.56) diagnosed with a fluent variant of PPA (svPPA or lvPPA) and one individual with severe anomia associated with early onset AD. Diagnoses were made by experienced behavioral neurologists (i.e., Grossman, Coslett, Hamilton) at the University of Pennsylvania in accord with published diagnostic criteria (Gorno-Tempini et al., 2011; McKhann et al., 2011). All participants were right-handed, native speakers of English with no history of seizures, implanted medical devices, or previous brain injuries (e.g., stroke). All participants gave informed consent and the study was approved by the Institutional Review Board of the University of Pennsylvania. Table 1 reflects neuropsychological and demographic characteristics of the patient sample.

We tailored a target lexicon consisting of approximately 100 words to each patient (see Reilly, 2016 for detailed selection criteria). Briefly, target items were quasi-randomly selected from fixed lists representing the following six semantic categories: places, foods, clothes, household items, activities, and hygiene implements. Items in the fixed lists were normed by a group of healthy older adults (N = 15, mean age = 67.7) who were not involved in the study. Low frequency or familiar items were excluded and thus the fixed list contained 30 comparable items in each category (mean frequency = 4.84, mean familiarity = 4.86 on a 7-point Likert scale).

At baseline, we assigned training items by first selecting 15 target words from lists that consisted of 30 items per category (e.g., 15 articles of clothing). Unassigned target words from each semantic category served as item-level controls throughout the treatment. In some cases, patients were assigned items that had little functional utility (e.g., subject CM never wears ties), and for these cases we permitted the patient and caregiver together to swap in new items from the original list.² After assignment, we compared the word frequency and familiarity between trained and untrained items for each subject. The subjective frequency and familiarity ratings (i.e., the degree to which one comes into contact with a word’s concept) were not very different between trained and untrained items for subject CM and RC. Although statistical significance showed difference for the other three patients, the difference was less than 0.6 units on a 7-point Likert scale indicating small effects. We also included a category of familiar people (N = 15) with the constraint that patients and caregivers should work together to include people who are frequently encountered, highly familiar, and have high personal relevance (e.g., a spouse or caregiver but not a grandchild seen once a year). For the ‘people’ category, it was therefore impossible to randomly assign a set of target items. For the people training condition, we included a set of famous faces (N = 15) as item-level controls.

Once a suitable item pool was established for each patient, we traveled to that patient’s home and obtained digital photographs of as many of the trained and untrained items as possible. We then generated a communication book/binder for each participant by arraying 500 pixel² photographs in groups of 4 per-page clustered by semantic category (e.g., people, foods, hygiene implements, etc.).

Semantic feature analysis typically involves asking the patient to either generate or verify a set of features for a given target item (Boyle, 2004; Hashimoto and Frome, 2011). We modified these task demands to capitalize on the benefit of an error-reduced learning approach (Frattali, 2004; Jokel and Anderson, 2012). We did so by having the treating clinician first announce the name of the target item and identify five of its constituent semantic features. We employed distinct semantic feature matrices for people and objects. For objects, features included use (i.e., what do you do with it?), physical appearance (i.e., what does it look like?), location (i.e., where do you find this?), and association (i.e., what is one thing that you often see near/with this?). For people, the matrices included relation (i.e., how do you know the person?), physical appearance, activity (i.e., what do you do together?), and fact (i.e., what’s an interesting fact about the person?).

For the instruction component, the clinician announced the name and features for blocks of five items. Immediately afterward, we cued the patient to name, generate features, and produce a novel sentence for the items that he/she just reviewed. Patients were instructed to avoid empty statements (e.g., an apple is a thing.). Using this technique of explicit instruction coupled with self-generation, we repeatedly cycled through the item pool.

Patients completed ten days of tDCS paired with behavioral therapy spaced over two consecutive weeks (i.e., Monday–Friday). Each tDCS session began by seating the patient and probabilistically landmarking the left temporoparietal region on the scalp using the 10/20 EEG coordinate system (Nitsche et al., 2008). We placed the anode at the left P3, and the cathode was centered over the forehead. The 1 cm² rubber electrode leads were placed within 5 cm² sponges that were saturated with saline and fixed in place on the scalp using rubber straps. Once the electrodes were placed, we administered tDCS using a Magstim Eldith 1 Channel DC Stimulator. Patients received 20 minutes of 1.5 mA with ramp up/down periods of 30 s.

We employed a pre/post treatment design pairing the behavioral intervention with online anodal tDCS. Prior to initiating the treatment, we administered a battery of neuropsychological assessments and gauged baseline naming for both the trained and untrained items. We conducted post-testing immediately after the treatment (i.e., at 2 weeks) with follow-up testing at 6 months. During each treatment session, we initiated behavioral treatment within 5 min of the onset of tDCS stimulation, and the behavioral treatment was administered for a duration of approximately 30 min per session (extending 5–10 min past tDCS rampdown to finish the behavioral intervention in a session).

Treatment consisted of a clinician (author AB) announcing the name of each target word followed by an explicit description of its semantic features. The patient was then cued to name each word, self-generate its features and generate a novel sentence containing the target word. Responses were considered correct if the target name was successfully produced. Patients were trained on 21 target items each day, 3 from each category, repeated twice through 10 sessions.

We analyzed accuracy data using the R statistical software package lme4 (Bates et al., 2015) via logistic mixed effects model. We treated individual patients and target items as random factors (Baayen et al., 2008; Jaeger, 2008) and entered the following as fixed factors in the model: Time (three levels, pre-stimulation, post-stimulation, 6-month follow-up), Item type (two levels, trained, untrained), and item category (seven levels, activity, clothes, food, household, hygiene, people, place). Individual participant data were analyzed via McNemar’s tests.

Figure 1 represents a plot of accuracy data across semantic categories for the trained and untrained items at three different time points (baseline, immediately post tDCS, and at 6-month follow-up). The best fitting model revealed three significant factors: Time, Item type, and Semantic Category.

The baseline naming accuracy was 69% for trained items and 40% for untrained items (p < 0.001). Although trained and untrained item lists were comparable in general item frequency and familiarity, the higher performance for the trained items at baseline could be related to the nature of functional utility to patients themselves. After 2 weeks of behavioral + tDCS treatment, patients showed accuracies of 77% for trained and 41% for untrained items. Six months post treatment, patients showed accuracies of 68% for trained items and 27% for untrained items. These accuracy differences resulted in a significant effect of training (b = 1.54, SE = 0.18, p < 0.001). The magnitude of the treatment effect diminished at 6-month follow-up (b = –0.97, SE = 0.18, p < 0.001). Patients persisted in naming trained items more accurately than untrained items at post stimulation (b = 0.56, SE = 0.23, p = 0.02) as well as at 6-month follow-up (b = 0.73, SE = 0.24, p = 0.002).

In addition to the main effect of treatment and time, patients showed a significant interaction between these two variables. This interaction was such that for trained items there was significant improvement in naming immediately post tDCS (b = 0.63, SE = 0.18, p < 0.001) but that these improvements fell to near baseline levels at the 6-month follow-up (69% vs. 68%). Untrained items, however, showed a flat pattern of response accuracy from pre- to post-tDCS (40% vs. 41%) over a 2 week span. Moreover, the untrained items showed significant drop-off at 6-month follow-up relative to the pre-treatment baseline (b = –0.96, SE = 0.16, p < 0.001). Figure 2 (left panel) illustrates the relevant main effects and interactions.

The magnitude of the treatment effects differed across semantic categories at different time points. To better understand the interaction, we conducted individual mixed effect models to examine the effects of Time for each semantic category. Trained items were more likely to show positive improvement immediately after tDCS. The most prominent improvements were seen in naming accuracy for inanimate object categories such as Clothes (51% → 70%, p = 0.007), Household Items (64% → 74%, p = 0.07), and Places (75% → 84%, trend, p = 0.094). However, untrained items did not show significant improvement post tDCS, indicating that treatment gains did not diffuse within semantic categories (e.g., improvements in naming of ‘socks’ did not see parallel improvements in naming ‘shirts’). At the 6 month follow-up, patients showed marked loss in the naming of several semantic categories. Among trained items, patients performed worse at 6-months for Hygiene (p = 0.04) and People (p = 0.003). Among untrained items, patients showed significant decrements in naming accuracy at 6-months for most of the categories: Activities (p = 0.002), Clothes (p < 0.001), Food (p = 0.01), Household (trend, p = 0.098), Place (p = 0.05), and People (p = 0.05). Table 2 reflects the mean accuracy for each semantic category.

We assessed naming accuracy at three discrete time points (pre-, post-, 6-months) within subjects. Table 3 reflects individual accuracies at each of these time points. This design yields three potential pairwise comparisons (T₁–T₂, T₁–T₃, and T₂–T₃). We derived accuracy change scores within participants using the following formula: (Mean_Post–Mean_Pre)/Mean_Pre. In the derivation of these difference scores, “pre” represents the baseline for each comparison. That is, observation T₁ is “pre” for the contrast of T₁–T₂ and T₁–T₃, whereas observation T₂ is “pre” for the contrast of T₂–T₃. The extent of improvement between time points was compared using non-parametric, paired sample McNemar’s tests.

Two patients (i.e., VM, RC) showed significant improvement in naming accuracy for trained items post-stimulation (T₁–T₂) with changes in accuracy ranging from 38 to 53%. Subject JB improved her naming accuracy to 100% at post stimulation but together with subject CM they maintained near ceiling performance (>95%) at post-stimulation and they demonstrated small naming accuracy decrement at 6 month follow-up (∼ 10%). Another participant (subject MC – early onset AD) showed no significant effects of treatment in a contrast of pre- and post-stimulation. Subject VM showed greater naming accuracy decrement at 6 month followup (T₁–T₃ = 25%) but only subject RC continued to show positive improvement (T₁–T₃ = 73%). Whereas the group as a whole showed no improvement for the untrained items at post-stimulation (T₁–T₂), subject MC (T₁–T₃ = –49%), JB (T₁–T₃ = –55%), and RC (T₂–T₃ = –36%) showed significant drop-off at 6 month follow-up. See Table 3 and Figure 1 (right panel) for the individual participant performance.