In close collaboration with numerous caregivers, we chose to focus our study on the expertise of caregivers rather than relying on discourse analyzed or produced automatically by AI models, such as large language models (LLMs).

Indeed, even though progress has been made in learning the language of caregivers [56], we believe that some in-depth studies are necessary to fine-tune the training of LLMs. Additionally, in the procedure we used for collecting data, the caregiver is central in assessing the care sessions while taking into account verbal, nonverbal, and contextual elements, which would be very challenging for current LLMs. In other words, by centering our analysis on the nuanced understanding of caregivers, we aim to capture the complexity of caregiver communication, which may not currently be fully appreciated or accurately interpreted by automated systems alone.

While LLMs offer promising potential in generating interaction strategies, the unique complexity of caregiver communication requires a deeper and more human understanding. A framework provided by caregivers is essential for fully understanding these interactions. Only a caregiver, through her experience, can recognize the intention to reassure a resident in countless subtle ways that cannot be captured by predefined rules or algorithms. These complex skills, shaped by the caregiver’s deep contextual understanding, are crucial for interpreting and effectively responding to care situations, which even advanced AI models such as LLMs may currently struggle to reproduce.

Data collection

To begin our research on interactions with elderly individuals in nursing homes, we first needed to collect data related to caregiving situations. This task proved challenging, particularly given recent scandals that have been widely publicized in the media. Access to healthcare facilities is often restricted for external observers. Consequently, we recognize that our data, collected from a single establishment, may carry biases and may not fully reflect broader trends across the sector.

The two types of data used in our analysis are logically and temporally connected. The first dataset, collected in 2019, served to identify key variables that informed the structure of our database, which was subsequently populated with data gathered in 2023.

Data collection took place at EHPAD Stéphane Kubiak in Oignies, a nursing home operated by the French group ‘La Vie Active’. This EHPAD has around 100 residents, 38 of whom are in a special Alzheimer’s unit. This EHPAD corresponds to the national average for the number of people with Alzheimer’s in retirement homes, which is 40% according to the “Direction de la Recherche, des Études, de l’Évaluation et des Statistiques” or French Directorate of Research, Studies, Evaluation and Statistics [2]. At the time of the two data collection steps, the average age of residents was over 80, which corresponds to the national average for people accommodated in these facilities [2]. While most of the residents are female, as is the case in the majority of nursing homes in France [2], other sensitive data, such as ethnic origin, are not accessible due to personal data protection regulations, such as the “Règlement Général sur la Protection des Données” or General Data Protection Regulation (RGPD) in Europe.

The 2019 corpus consists of 34 fully transcribed audio and video recordings capturing a broad range of interactions between caregivers and residents. These recordings highlight a diverse set of caregiving activities, such as assisting with meals, helping residents out of bed, and preparing them for sleep, with each scenario involving a different combination of caregivers and residents. The corpus showcases a remarkable variety of care situations, featuring around 40 distinct resident profiles, ranging from individuals with Alzheimer-type conditions to those with varying levels of dependency. This variety extends to the caregivers as well, ensuring a rich representation of different approaches and interactions in the caregiving process.

Table 2 Transcription convention

In the remainder of this paper, we will analyze these interactions, which take place during what we refer to as care sessions, care situations, care interventions, or care acts, depending on the context, and were collected within the nursing home setting. To cite specific interaction items, we will employ ICOR-like [5] conventions as outlined in Tables 1 and 2. Note that, for the sake of readability, we have preferred to use icons rather than identifiers, and to add a few punctuation marks when inserting the transcription extracts. This allows us to highlight actions and comments, which clarifies and contextualizes the discourse.

Key features of language-based interactions

Through contextualized transcription, we were able to conceptualize the lexicon as a set of features. In the context of elderly care in nursing homes, each feature represents a specific criterion, identified through a combination of technical and medical literature, including informal geriatric care guidelines, and extensive discussions with caregivers. This approach was necessary due to the lack of formal training in elderly care, as highlighted in reports such as the 2022 Cour des comptes analysis on medical care in nursing homes [20]

The first step in our approach is to identify the features, understood as influencing factors, that are most likely to affect interactions. Some of these features are linked to the resident, while others pertain to the caregiver. Understanding these features helps anticipate whether a resident’s response will be appropriate or inappropriate, thus enabling caregivers to better tailor their interactions. This approach forms the basis of the AI-driven method we introduce and explain in detail later in the paper.

In this section, we identify the key features that characterize the fundamental concepts of language-based interactions between caregivers and nursing home residents, according to three main dimensions:

Care context

Infantilizing language

Language disorders

Among the variables most obviously considered influencing factors and likely to have an impact on interactions are those linked to the context of care like pain or aggression. Those situations can be transversal to all care situations regardless of the age of the patient/resident. For instance, the discomfort, such as pain, expressed by a resident could be observed in the corpus as follows:

and

and

Any kind of aggression (verbal or physical) requires the careful involvement of the caregiver. This is illustrated below in the context of moving an elderly person from one room to another (warning: this extract contains coarse language which may shock but reflects the reality of care practice):

These factors will provoke a response on the part of the caregiver: familiarity [63], closed-ended questions [1], or repetition for example. We will illustrate this caregiver reaction with two additional examples. First, we begin with the concept of “renarcissism” [60], which focuses on helping individuals regain their self-confidence. This is reflected in the audio and video corpus through comments such as follows:

The second example highlights reassurance, demonstrated in extracts where caregivers provide comfort to those in their care:

and

and

and

However, some interaction parameters are specific to the care of the elderly, such as the use of elderspeak.

We focus on the infantilizing language that may be used rather frequently in nursing homes and its potential impact on residents. Some parallels have already been drawn between a public made up of children and one made up of elderly people [71]. In particular, elderly persons are often infantilized by the use of what we call patronizing speech, over-accommodation, baby talk, or elderspeak. This form of language interaction results in a slower rate of speech [68], a form of familiarity and closenessFootnote 1 [66], excessive repetition [68], exaggerated intonation [71], the use of collective pronouns (we, us) [67], and a limited vocabulary [39]. According to many studies, using elderspeak with older people does not improve care, especially for those suffering from dementia.

On the contrary, it causes resistiveness [80]. Consequently, the misplacement of a type of communication (initially intended for children) to an inappropriate target (the elderly) does have an impact on care interventions.

Moreover, caregivers should be aware that the aim of verbal interaction is not the same for children and elderly people. In the context of a dialog with the elderly, it is well known that a crucial issue is to slow down cognitive decline [29], by recalling elements that have been lost by the elderly person, as illustrated in this authentic extract from our corpus:

Here, a caregiver mentions a visit from the resident’s wife (your wife came this weekend?) even though the resident does not seem to remember. Then, he insists twice.

Following on from the remarks above, one can emphasize the differences when interacting with an audience composed of either children or elderly people. Indeed, when interacting with a child or toddler, the aim is to solicit the cognitive and linguistic development [18, 26, 61], whereas, with the elderly people, the aim is to preserve life-long acquired knowledge. Actually, for elderly people, it is essential to recall the basic concepts of everyday life, in particular when the person suffers from cognitive disorders. For example, in the following extract, the caregiver reminds the resident of her ability to eat on her own:

We believe that language disorders, particularly those stemming from cognitive impairments, must be taken seriously (consider the various pathologies that can affect the elderly). Alzheimer’s disease, which affects 40% of residents, is one of the most common illnesses in nursing homes in France [2]. This progressive neurodegenerative disease is often accompanied by language disorders, including perseveration [57], which refers to the inappropriate repetition of a verbal or gestural response. This can be observed in the following extract where the first syllable of the word “serious” is repeated by the resident, and then by the caregiver. This behavior stresses the resident’s difficulty in expressing herself and seems to lead to rejection.

The various language disorders associated with this disease give rise to numerous unforeseen events in the interaction with the elderly person, such as aborted words [75] or a change in tone [16]. Another illustration is given below, where shouted words are written in capital letters (and preceded by “shouting” between brackets) and repetitions concern the words “no” and “go”:

Identifying these features serves two functions. First, it highlights the importance of language skills in specialized settings for professionalizing practices. In the field of personal assistance, there is often confusion between intuitive help and professionalized help, leading to the misconception that assistance is accessible to all. However, this is far from true. The argument here is that caregivers can improve their communication in response to events by drawing on their experience and knowledge of best practices. Second, translating the lexicon into data science features enables the use of AI tools. These tools, especially in unsupervised machine learning, help identify the underlying causes of specific situations.

The method used in this study can be summarized in Fig. 1.

To summarize, our contributions are as follows:

An interdisciplinary collaborative work carried out (over a period of three years) by people from both the Health sector (caregivers) and the Academic sector (researchers in Linguistics and Computer Science).

An extensive campaign conducted for collecting data (audio/video recording of care sessions, questionnaire completed by caregivers) in nursing homes.

The identification of key features of language-based care interactions.

The development of AI models (random forest, restricted Boltzmann machine) for better understanding, predicting, simulating, and explaining care interactions.

The practical exploitation (general knowledge of the importance of some features, and, forthcoming, a mobile application) of these models.

The demonstration that new challenges lie ahead for training: Caregiver training should include in the future a strong focus on the language skills mobilized during care interactions with elderly residents.

Fig. 1

Human-centered analysis of care interactions. The caregiver is at the center of our study; their interactions initiate our reflection, and their experience allows us to adjust data collection and refine our models. The results of our research support the training and interactions of caregivers with residents

The database

The database was created in August 2023 based on the analysis of the corpus. A teacher responsible for training future caregivers and assistance professionals conducted this second round of data collection. Ideally, the labeling process would involve multiple experts to cross-check and validate classifications, reducing personal bias and ensuring a more robust and reliable dataset. However, the teacher’s specialized expertise lends significant credibility to the task. Since the labeling was performed by an educator who trains future caregivers, this expertise helps mitigate bias and enhances the accuracy of the process. His pedagogical experience and in-depth knowledge of best practices in elderly care provide a solid foundation for precise and informed labeling, further improving the reliability of the data. It is also important to note that data collection should not be conducted by a caregiver unfamiliar with this population, which adds the challenge of finding professionals with the appropriate profile for such tasks.

This data collection was funded by a research prize awarded at the end of 2022Footnote 2 and a research quality bonus obtained between our two research laboratoriesFootnote 3. These resources enabled the comprehensive analysis of the resident-caregiver interactions within the facility.

These residents often present unique challenges, requiring caregivers to have specific training and understanding of neurodegenerative conditions.

Gender distribution is another important sociological factor to consider. As noted, most residents are female, which is consistent with broader trends in nursing homes across France, where women are more likely to live in these facilities than men [2]. This can be attributed to women’s longer life expectancy compared to men, meaning they are more likely to reach advanced ages where institutional care becomes necessary.

Regarding the data collection process, the use of video and audio examples to build an initial corpus was essential in identifying relevant features. The questionnaire, which was accessible via a tablet, allowed for the systematic observation of more than 503 interactions over a month. These interactions, involving around 100 residents and 50 caregivers, were observed across all sectors of the nursing home, including the Alzheimer’s unit. This sample size is substantial for the research methods used and provides a robust dataset for analyzing interactions in a caregiving context.

In this sense, the facility not only reflects national trends in aging and care but also serves as a microcosm of the broader dynamics within French nursing homes, offering insight into how care is delivered in environments that are increasingly shaped by the needs of an aging population. The data collected can provide valuable insights into the ways residents and caregivers interact, potentially influencing future care practices.

To delve deeper into these interactions, it is important to consider both the intrinsic features within the interactions themselves and the external factors that may shape them. For instance, the caregiver’s initial training, specialized courses on elderly care, and overall experience are likely to have a significant impact on the nature and quality of these interactions.

This section (background) of the questionnaire pertains to the aforementioned aspects:

Particular attention has also been paid to various contextual factors, including the type of unit or service, the time of day, the type of assistance or care required, and the location.

Also, in relation to the context, a caregiver who is accustomed to caring for a specific individual will have a different relationship with that person. This will have a clear impact on the interaction and care.

Unfortunately, nursing homes often experience high staff turnover [50], which can reduce the quality of care provided to residents [19].

Similarly, the degree of dependency or a pathology may influence the interactions. For instance, neurodegenerative diseases such as Alzheimer’s can make interactions (almost) impossible.

Finally, we have included the intrinsic parameters (variables) that are specific to the interactions. This includes triggers for the elderly personFootnote 4 and reactionary gestures for the caregiver.Footnote 5

This also includes factors that could be seen as infantilizing, such as using inappropriate vocabularyFootnote 6 [42],

or depersonalizing such as substituting collective pronouns (for example, we are not happy this morning [9, 71]).

The final variable in the questionnaire relates to the success of the care, which can be classified as successful, partially successful, or unsuccessful. Caregiver interpretation of real-life interactions is crucial, as the quality of the data directly impacts the effectiveness of training an AI model. This is why, during data collection, we prioritized the caregiver’s expertise and minimized reliance on automatic interpretation. Furthermore, the results of trained AI models must be interpreted by humans. This interpretation is ultimately entrusted to a human specialist, which inevitably introduces a degree of subjectivity.

The dataset

The final stage involved transforming the three non-binary responses—regarding treatment success, the caregiver’s overall experience, and their experience in nursing homes—into binary variables to serve as inputs for our AI model. This decision was driven by the efficiency of some of our models when applied to binary data. Additionally, we aimed to develop a mobile application for caregivers, which would not have been feasible without this transformation. This conversion required interpreting the three corresponding questions based on criteria derived from our discussions with caregivers. Regarding caregiver experience in nursing homes, we classified any experience of less than one year as “low.” For this question, the questionnaire includes the following elements:

The corresponding binary variable for “solid experience” in this case has two possible values: 0 for “weak experience,” which applies to the first two answer choices, and 1 for “solid experience,” which applies to the remaining answers. Therefore, a caregiver with exactly one year of experience is categorized as “experienced,” though a different threshold could have classified them as inexperienced.

We chose to define a treatment as successful when it was fully completed. A completed treatment is coded with a value of 1, and 0 otherwise.

Moreover, while we opted for a binary approach, we acknowledge that different thresholds could have been applied, such as considering experience below 1.5 years as weak. Similarly, partial care completion could have been classified as successful. These alternative interpretations highlight the flexibility of the approach and suggest that multi-level encoding could be explored in future studies. However, for this initial stage, binary encoding provides a clear and effective starting point, especially given the subjective nature of some features like caregiver experience and care completion.

In the care sector, opinions on what constitutes adequate experience or successful care can vary significantly. Adjusting key thresholds, such as the minimum experience period or the degree of care completion, can lead to different interpretations of the same dataset. This flexibility might result in the creation of multiple datasets with varying binary variables. In this paper, however, we have limited our study to a single dataset. It is important to note that our discussions with caregivers strongly influenced the establishment of these key thresholds, giving us confidence in the relevance and applicability of our work. Moreover, we observed that all “reasonable” thresholds led to fairly similar results. Thus, we adjusted the experience threshold up to two years, and treatment success could include partial success. In all cases, the interpretations remained largely consistent. We selected the thresholds that most closely aligned with the caregivers’ perspectives.

This has resulted in 50 binary variables as follows:

Training in Interaction PRO

Soothing Verbal Communication PRO

Reassuring Posture or Gesture PRO

Approval PRO

Reformulation PRO

Repetition PRO

Depersonalization PRO

Use of the First Name PRO

Renarcissization PRO

Unsuitable Vocabulary PRO

Mention of Family PRO

Mention of Residents Past PRO

Mention of Everyday Objects PRO

Humor PRO

Laughter PRO

Modulation of Voice PRO

Closed-ended Questions PRO

Slowed Speech Rate PRO

Aggressiveness RES

Pain RES

Sadness RES

Use of Non-Verbal RES

Verbal Refusal RES

Nonverbal Refusal RES

Abnormal Repetition RES

Incorrect Words RES

Foreign Language RES

Voice Modulation RES

Speaking Little RES

Deafness RES

Not Adapted Location

Pathology-Impacted Interaction

Usual Caregiver

Possible Interaction

Successful Care

Assistance for Getting up

Assistance for Going to Bed

Assistance for Taking a Nap

Assistance for Walking

Assistance for Transfer

Assistance for Mobility

Hygiene Care (partial assistance)

Hygiene Care (full assistance)

Comfort Care

Well-Being Care

Assistance (Partial) with Meals

Assistance (Full) with Meals

Technical Care

General Professional Experience

Nursing Home Experience

The terms RES and PRO are used to differentiate between features associated with residents and caregivers, respectively.

Machine learning for care

The efficacy of machine learning techniques in the domain of medicine has been substantiated. In particular, they are capable of outperforming traditional statistical techniques. [73]. The application of AI in the field of medicine has yielded numerous successful outcomes over the course of its development [34, 38, 40, 43, 44, 62, 74, 74,