2.1. Data sources and search strategy

Electronic literature searches were performed using ACM Library, Cochrane Systematic Reviews Database, Google Scholar, PubMed, PsychINFO and Web of Science including publications up from January 2000 to March 2022. Two researchers reviewed the potential studies individually for eligibility. Was used a list of keywords to identify the studies, including relevant interventions, through an interactive process of search and refine (see Table 1). Each database was searched independently, according to a specific interaction research string: (“Mild Cognitive Impairment” OR “MCI”) AND (“Robot” OR “Robots” OR “Human robot interaction” OR “Humanoid robot” OR “Zoomorphic robot”).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Detailed search strategy.

https://doi.org/10.1371/journal.pone.0293324.t001

No limitations regarding study design or outcome measures were used. We included only English’s article. Eligible studies were those whose title or abstract specifically indicated the inclusion of MCI, use of social robots and protocol of cognitive and social interventions. There were no restrictions about the age and number of participants. In the first export were included only article with full-text available. The complete list was exported in EndNote to remove duplicates and then it was imported in Rayyan [29] for title and abstract screening.

2.2. Study selection criteria

The aim of this review was to evaluate the efficacy of social robot-based interventions with MCI patients to enhance cognitive and social abilities. The following selection criteria were applied to the articles found in databases: research studies, reviews and case reports were eligible for inclusion; chapters and not peer review studies were excluded. The abstracts of the identified publications were screened for relevance to the selection criteria. Specific inclusion criteria were: samples with MCI clinical condition; a cognitive and/or social training through a social robots. Papers that described unstructured and holistic intervention programs for patients with MCI were rejected.

2.3. Quality assessment

The ’Tool for assessing the risk of bias in randomized quantitative trials’ [32], which includes five domains related to the quality of the methodology in randomized trials, was used to assess the quality of the risk of bias (RoB2). Each domain was rated by two independent reviewers (GF, AT) who rated each domain by assigning it a risk of bias rating; the rating could be categorized as follows: low risk of bias; some concerns and high risk of bias. A judgement of ‘High’ risk of bias for any individual domain will lead to the result being at ‘High’ risk of bias overall, and a judgement of ‘Some concerns’ for any individual domain will lead to the result being at ‘Some concerns’, or ‘High’ risk, overall. With the purpose of calculating the risk of bias for the qualitative studies included in this systematic review, it was decided to use the tool of GRACE (Good Research for Comparative Effectiveness) that is an 11-item instrument designed to evaluate the quality of the data and the methods used either in the design and the analysis of the noninterventional and observational studies of comparative effectiveness [33]. The GRACE tool is composed by two sections: the first dedicated to the data (6 items) of the study regarding the treatment, outcomes, and population; the second dedicated to the methods (5 items) with more information about the population, the control group, possible follow up etc. [33].

2.4. Quantitative studies

Concerning quantitative studies and the dimension related to the randomization process (S1 and S2 Figs), some studies [17,34–36] have a low risk of bias; this suggests a random distribution of participants to the different intervention groups and the difference in outcome between the two groups is not an indicator of randomization problems. Two studies [18,24] show some doubts, while works [16,23] have a high risk of error related to the randomization process.

The second aspect explored was the deviation from the intended interventions; here only study [35] shows a low risk of bias while four works [15,18,34,36] show some concerns regarding the adopted methodology. The remained works [12,16,24] on the other hand, present a high risk of bias; this proportion is related to the awareness of the subjects and their caregivers about the intervention and to factors related to the experimental context that may have affected the outcome of the intervention.

The third dimension of Rob2 concerns bias due to the lack of outcome data; this could happen when participants withdraw from the study ’dropout’ and if they do not provide relevant data; if participants die, etc. In this case, the studies can be divided into low risk of bias: [35], (relating to the outcome of improvement in visual memory), [16–18,23,34,36] and high risk of bias: [35] related to the outcome of the improvement in executive functions, [35] in relation to increased cortical thicknesses, [18] in relation with change of prose memory, [12] concerning the increase in the frequency of communication in the robot group and the increased, in both groups, of interaction with staff, and [24].

Also, with regard to measurement of the outcome, one could categorize the work into studies with a low risk of bias: [12,17,24,34–36] and studies with a high risk of bias: [35], relating to the improvement in the executive functions in robot group, [12,16,18]. When referring to measurement error, this concerns misclassification (for dichotomous or categorical outcomes), the use of non-adapted tools for the measurement of the outcome being studied, the possibility that administrators may be aware of the intervention provided to subjects, etc.

Regarding the selection bias of the reported result, a dichotomous situation was found in studies [17,35] and [12] show some concern on the reported results, while a high risk of bias was observed for the studies [35] relating to the improvement in the executive functions in robot group, [12,16,18,24,34,36] with respect to the increase in frequency of positive expressions and reductions of loneliness feelings in the robot group. With regard to this dimension, several studies do not report pre-specified analysis plans that were finalized before unblinded outcome data were available for analysis and lack multiple eligible outcome measurements from which to derive data.

As a final result, a global calculation of the risk of bias relating to each job can be observed; in this case, some results lead to an uncertain risk of bias [17,34–36] and in overall results with a high risk of bias [35], regarding improvement in visual memory, [12,16,18,24].

2.5. Qualitative studies

With regard to data on qualitative studies, two independent assessors compiled the GRACE scale (S3 Fig) [33] and then compared the results and reached an inter-judge agreement. The Grace scale is subdivided into two macro sections: a first section on data and a second section on methods [33]. We will begin the analysis of the results of the papers included in the review from the first section which is concerned with investigating whether the study data are adequately recorded (D1-D2), whether they are clinical outcome (D3), whether they have been validated on similar populations in terms of diagnosis (D4), and whether the data have been measured on a comparison group (D5). What emerged from the evaluations was that all studies correctly recorded data except the one of [37] where, although both qualitative and quantitative measures were implemented in the study, the latter were recorded through methods that are closer to a qualitative and observational methodology. Moreover, the contents of qualitative survey measures such as comments and open-ended questions are not specified in the paper, therefore, there is insufficient information in the publication to allow us to say whether the treatment data has been adequately recorded. Moreover, most of the studies did not have a clinical outcome expect for the following works: [5,22,38]. Most of studies were pilot studies or focused on new types of populations, the only studies that have used a protocol or previously validated measures in other populations are [39–41]. Regarding the reproducibility of the results in a hypothetical control group, we find that the only one studies have this condition [38]. Finally, the recording of important covariates was not found in any of the included work. Overall, what emerges was that, limited to the data section, most of the studies mentioned presents what could be defined as a high risk of bias expect for one work [38], which could be interpreted as a study with a medium risk of bias. Considering the methods section, we found how for the first item (M1) all studies, except for [38], included new initiators instead a population already under treatment, whereas, for the second item (M2) the studies which used an historical comparator group were: [20,38,40–42]. For the third item (M3) the studies which take into account important confounding and effect-modifying variables were: [40,41].

Finally, none of the studies was free of “immortal time bias” (M4) and none of them has conducted meaningful analyses to test the key assumption which primary results were based (M5). In conclusion, regarding methods section, all the studies included presented what we could define as a high risk of bias.

These results suggest how the novelty of the used protocols in ’human-robot interaction’ studies, particularly in the rehabilitation context, are still in the development and validation phase all these aspects would be discussed and deepened later.

3.1. Participants

A first important finding relates to the clinical pictures that compose the study samples. As a matter of fact, only seven studies involved people with a diagnosis of MCI [16,18,25,35,36,38], while the remaining twelve included mixed samples (i.e., MCI and patients with mild/moderate/severe dementia) [15,17,20,21,23,24,34,37,39–42].This result highlights that the selected studies do not have homogeneous samples of MCI, despite the studies declared as aim the exploration of the effects of interventions with social robots in this clinical population. Underlining the lack of clarity of the studies with respect to sample recruitment, only in one of the nineteen studies reported the specific cognitive area of impairment (i.e., anamnestic MCI; [18]).

Thirteen studies administered cognitive or social training to MCI without control groups [15,16,18,20,22–25,34,37,39–41] while six studies recruited control groups [17,21,35,36,38,42] that, however, they did not compared MCI patients involved in robot-based interventions with other MCI patients without robot-based interventions. This finding shows that the lack of control groups may reduce a real evaluation of the positive effects on maintaining and/or improving cognitive and social skills in MCI.

Another relevant aspect is where the interventions with MCIs were conducted. Eleven studies were carried out in residences for the seniors or nursing homes [17,20–24,34,37,40–42], two studies offered a home-based protocol [25,36], and in six studies the interventions were conducted in specialized centers for cognitive disorders and dementia [15–18,35,38]. This is relevant because the outcome of an intervention may vary depending on contextual conditions, again underlining the generalization problem of these studies.

Finally, the interventions were conducted in different countries: four in Italy [15,16,18,38]; three in Korea [17,35,36], two in Australia [37,41], Japan [21,22] and USA [20,39]; one in Belgium [25], China [34], Ireland [42], Netherland [37], Singapore [23] and Spain [24]. Cultural differences also could have an important influence on the clinical conditions and the practices to approach to them. This is an additional element that prove the heterogeneity of this filed of research.

3.2 Type of robots and programming

Another interesting issue that characterizes most of the studies is the heterogeneity of social robots used. Fifteen studies used humanoid robots (BOMY, JACK, JAMES, MATILDA, MERO, NADINE, NAO, PEPPER, RYAN COMPANIONBOT, SILBOT and SOPHIE) [16–18,23–25,35,35–42] and three used zoomorphic robots (AIBO and PARO) [15,20,21]. The heterogeneity of the social robots employed in the various studies does not allow generalization of the data to social robots and remains open if a specific social robot is more effective with MCI than others. One study used both the humanoid robot NAO and the zoomorphic robot PARO [24]. One study compared the effectiveness of two different types of humanoid robots (SILBOT and MERO) [35], and one study used two humanoid robots simultaneously (JACK and SOPHIE, [40]). Although these few studies used different social robots and claimed in the objectives to compare the effectiveness of these, the analyses did not report any specific results on this issue, thus leaving unanswered which robot is more effective. Regarding cognitive interventions, all the studies used anthropomorphic social robots: one the BOMY [36], one both MERO and SILBOT [35], two the NAO [16,18], one the PEPPER [38], one the RYAN COMPANIONBOT [39] and one SILBOT [17]. This result highlights the importance of anthropomorphic features of robots for interventions that focus on the residual cognitive functions of MCI. However, there is no detailed consideration with respect to which anthropomorphic features are most functional for this clinical condition.

Regarding social interventions, eight studies proposed the use of anthropomorphic social robots: two the KABOCHAN [22,34], one the JAMES [25], one the MATILDA [41], one the NADINE [23], one the PEPPER [37], one the NAO [42], one both JACK and SOPHIE [40]. Four employed zoomorphic robots: one the AIBO [15], two with PARO [20,21], and one both the NAO and PARO [24]. The use of these two types of robots, anthropomorphic and zoomorphic, evidences a greater openness in interactions by MCIs toward robot design features in the emotional-relational sphere. This seems due to the caring behaviors that zoomorphic robots solicit in MCIs by supporting residual social skills.

All studies used the Wizard of Oz technique, and no autonomous interaction system was developed. This underlines how, to date, it is difficult to hypothesize interventions with social robots that can be actively applied by non-technicians in MCI care settings as well as within the homes of people with this clinical condition.

3.3. Assessment

Regarding the assessment of cognitive abilities of MCI, there is a wide variability of psychometric tests in literature and there are no standardized assessment protocols. In general, MCI neuropsychological assessment aims to identify the presence of specific cognitive impairments, quantify the severity of the disorder and, with respect to interventions, verify their effectiveness [19]. Regarding Theory of Mind abilities in MCI, they are measured through different classical tasks and tests (e.g., False Belief Tasks, Strange Stories, Reading the Mind in the Eyes) [13,43].

Regarding cognitive interventions, of the seven selected studies, only two studies used pre- and post-intervention neuropsychological assessment. One study used Cambridge Neuropsychological Test Automated Battery (CANTAB) [35], while the other one assessed episodic memory, short term memory and verbal fluency [18]. Of these two, only one assessed patients pre- and post-intervention with MRI [35]. All of the remaining five studies did not adopt a post-intervention assessment. Two studies specified the neuropsychological batteries used: specifically, the Seoul Neuropsychological Screening Battery [36], and the Mini-Mental State Examination—Dementia Screening (MMSE-DS) and Consortium to Establish a Registry for Alzheimer’s Disease Korean version [17]. One study adopted a pre-intervention neuropsychological complete battery and two functional scales (Activities of daily living and Instrumental activities of daily living [16]). Other two studies used clinical observations of the participants and clinical interviews with participants and caregivers. Of these two, one specified the observational protocol and clinical interviews (i.e., Saint Louis University Mental Status Examination to detect mild cognitive impairment; [39]) and the other one stated that they analyzed neurocognitive characteristics without specifying the tests, trials, and/or interviews adopted [38].

With respect to social interventions, of the twelve studies includes, only three studies conducted a pre and post-intervention neuropsychological assessment: using the Mini Mental State Examination (MMSE), Severe Mini Mental State Examination (sMMSE and Neuropsychiatric Inventory, NPI) [24]; emotional through analysis of interactions using Deep Neural Networks (DNN) techniques and Menorah Park Engagement Scale (MPES) [23]; individual sense of well-being International Quality of life Assessment (QOL SF-8) and residual skill level Tokyo Metropolitan Institute of Gerontology Index of Competence (TMIG-IC) [22]. Other three studies did only an initial screening of cognitive competence through the use of MMSE [15] and Montreal Cognitive Assessment (MoCa [34,42]). The remaining six studies did not explicitly state the assessment procedures because the objective primarily was to study the acceptance of social robots by MCI patients. Of these six studies, three conducted only an observational assessment [20,40,41], one assessed the level of dementia through a scale [18], one proposed a pre and post intervention questionnaire and semi-structure interview [25] and, finally, one didn’t declare a neuropsychological or ToM’s assessment [37].

Overall, quantitative assessment protocols of both cognitive and social interventions are extremely heterogeneous, lack standardized procedures, and, in addition, only a few studies present pre- and post-intervention evaluation. From a clinical perspective, these limitations prevent a full understanding of the actual benefits of the interventions (e.g., amnestic, language). At the same time, qualitative evaluations present similar issues since neither standardized procedure are applied nor observation protocols stated.

3.4. Type of training

4.1 Social robots in cognitive and social interventions

Regarding cognitive interventions, the robot was used only in individual sessions with multi-domain cognitive exercises, and this is in line with classic intervention protocols in which activities are proposed individually and, given the complexity of the clinical picture, focus on multiple cognitive dimensions. Conversely, in social interventions, robots were used in group sessions whose main goal was to solicit the interest of participants and be mediator/animator of social exchanges among them.

In general, in cognitive interventions the robot is used as a device to administer tasks, proposing activities that the participant needs to complete. These tasks by stimulating specific cognitive abilities enable their enhancement. Studies have mainly focused on two cognitive domains: memory and executive functions (i.e., working memory). With respect to memory, tasks concerned music quizzes (i.e., recognizing the title of a song and associating the song with the singer) and quizzes on selected texts (e.g., answering questions about them and remembering key words), while executive functions were not directly addressed by cognitive trainings but were considered indirectly in social trainings. Specifically, in social interventions, the game of bingo was implemented as it allows training working memory by remembering the numbers of the draws. With respect to social interventions, these were implemented with both anthropomorphic and zoomorphic robots, while cognitive interventions only anthropomorphic robots were adopted. Zoomorphic robots due to their animal features were not considered tools to support cognitive functions probably because if they spoke, they would lose their resemblance to the animal. This limitation from the cognitive side, on the other hand, represents a strength for social interventions because their characteristics solicit caretaking of the robot by participants supporting relational components. More specifically, zoomorphic robots are particularly relevant in more severe dementia conditions while patients with MCIs who have different residual abilities, the use of anthropomorphic robots may be equally effective. This bias could be due to advanced cognitive impairment that reduces initiative in interactions and, therefore, it is easier to have an interaction on the sensory level occurring with zoomorphic robots that can be picked up and stroked.

Although most of the studies claimed to examine the effects of interventions with social robots, they actually examined the acceptability of the robots by the participants. Although the studies presented patients with different levels of cognitive decline, in all studies there were no episodes of rejection toward social robots regardless of the type of robot used. Even in cases characterized by a more severe clinical picture, both anthropomorphic and zoomorphic robots were easily integrated into daily and rehabilitative activities and generated curiosity fostering greater involvement in the therapeutic sessions. Despite these very positive results related to the acceptability of robots by the elderly, it is not possible to generalize the results related to cognitive and social interventions.

4.2 Clinical issues

From a clinical perspective, the studies report several methodological problems, the main ones are: clinical heterogeneity of samples; absence of standardized, pre/post-intervention and brain assessment protocols through functional techniques; and absence of control samples.

With respect to clinical samples, studies are mostly presented with mixed patient groups (MCI and mild/moderate/severe dementia), and those with only patients with MCI do not report specifics regarding the most impaired areas, except for one study with amnestic patients (aMCI). Prospectively, studies should build their samples by considering only MCI clinical populations by specifying the clinical subpopulation (i.e., amnestic, non-amnestic, and multidomain). This would allow the creation of customized interventions that could have more precise effects based on the specific needs of different clinical subpopulations. In fact, the studies analyzed in this review indiscriminately used cognitive exercises on memory without considering that MCI patients might have memory skills that are still functioning, but instead need specific exercises, for example, on executive functions. The problems in detecting clinical subpopulations of MCI also relates to the heterogeneity of the assessments used and the absence of functional brain assessments. Most importantly, not all studies report the assessment procedures adopted, and those studies that have reported them do not present structured protocols integrated with functional brain techniques. Future studies should detail more precisely the assessment tools adopted and consider assessments at different levels, including neurofunctional ones. The absence of control groups is an important methodological limitation because it results in a reduction in the generalizability of the positive results of the cognitive and social interventions. Future studies should consider including homogeneous MCI samples as control groups or compare different clinical populations grouped homogeneously with respect to the characteristics of the MCI sample involved (e.g., mild dementias with specific memory issues).

4.3 Robotics issues

Regarding the robotic area, there are two critical aspects to highlight: the variety of robots used and the use of the Wizard of Oz technique for all interventions. In recent years, it has emerged clearly in the literature on human-robot interaction how the type of robot strongly affects the perception and interactions with humans [43,44]. A few studies have tried to address this issue by comparing different robots (i.e., different anthropomorphic robots or an anthropomorphic robot vs. a zoomorphic robot), but the results do not present clear results because specific analyses were not carried out and, therefore, it remains open whether some social robots may be more effective than others in performing cognitive and social interventions. Another issue is the use of Wizard-of-Oz technique in interactions, highlighting a reduction in the scalability of these interventions. As a matter of fact, it would be difficult in specialized care institutions and people’s homes to use robots if the programs must necessarily be run by technicians.