Virtual Dynamic Keyboard for Communication in Intensive Care

Louisa Spahl

and Andreas Schrader

Institute of Telematics, University of L

ubeck, Germany

Keywords:

Active Communication, Intensive Care, Novel Interaction Devices, HCI, Virtual Keyboard.

Abstract:

Effective communication is of great importance for intensive care patients in the weaning process to express

their needs adequately. To support this process, the ACTIVATE patient application was developed, providing

a selection of typically used texts via a novel interaction device BIRDY, intended to be used in bed. However,

there are situations where patients would like to express more. Since the traditional layout of a static keyboard

does not ﬁt well with BIRDY gestures, we developed a virtual dynamic keyboard with letter prediction and

minimal input gesture needs. We tested different text corpora and forecasting models, implemented a prototype

based on the best candidate, and performed a preliminary user evaluation. The new virtual dynamic keyboard

is shown to be superior compared to static layouts.

1 INTRODUCTION

In Germany, 2,131,216 people were admitted to in-

tensive care units in 2017 and about 20% of them had

to be ventilated (Destatis, 2018). Natural language

communication with ventilated patients is usually dif-

ﬁcult or impossible with relatives and hospital staff

and intensiﬁes stress in patients (Kordts et al., 2018).

The breathing tube enters the patient’s windpipe via

the mouth or nose. It lies between the vocal cords

and prevents them from moving so that no sound can

be produced. Traditional methods, e.g., using writing

tablets, are cumbersome and error-prone, and cause a

high level of frustration. The presence of care staff is

also a prerequisite.

The BMBF project ACTIVATE

has developed a

special patient application for communication in in-

tensive care units. With the help of a ball-shaped

interaction device BIRDY, an application can be op-

erated independently by patients and realize syn-

chronous as well as asynchronous communication.

The ACTIVATE patient application has a repertoire of

typical sentences and actions for cognitively impaired

patients. But even if most communication needs are

covered in this way, the number of sentences is natu-

rally limited. The application therefore also offers a

virtual keyboard so that any text can be formulated by

cognitively ﬁt patients.

https://orcid.org/0009-0001-2084-7941

https://orcid.org/0000-0001-7926-0611

https://projekt-activate.de/

Due to the special situation of one-handed use in

bed with limited mobility, BIRDY only supports a

limited set of gestures, and leads to a cumbersome

and time-consuming text entry process.

We therefore developed a dynamic keyboard for

this special usage context to signiﬁcantly speed up

input and reduce patient frustration. The letters are

dynamically rearranged after each input to minimize

the number of gestures required. Three probability

models are compared for this purpose: The stochas-

tic Markov model (Jurafsky and Martin, 2023), the

LSTM (Long Short-Term Memory) model (Hochre-

iter and Schmidhuber, 1997), and the Deep Learning

Transformer model (Vaswani et al., 2017). Suitable

text corpora are identiﬁed for training. For the evalua-

tion of effectiveness of letter prediction and efﬁciency

of implementation, we performed machine simulation

to identify the best combination of corpora and model,

which was then integrated into the ACTIVATE ap-

plication and evaluated in a mixed-method laboratory

study with test persons.

The paper is structured as follows. Chapter 2

provides an introduction to the ACTIVATE project.

Chapter 3 describes comparable work, divided into

models and dynamic keyboards. This is followed in

Chapter 4 by a description of the prototype with con-

ception, corpora, user interface, and implementation.

The evaluation results are presented in chapter 5. The

last chapter summarizes the results, and gives an out-

look on further planned work.

Spahl, L. and Schrader, A.

Virtual Dynamic Keyboard for Communication in Intensive Care.

DOI: 10.5220/0013099500003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 41-52

ISBN: 978-989-758-731-3; ISSN: 2184-4305

2 PROJECT ACTIVATE

The ACTIVATE project, funded by the German Fed-

eral Ministry of Education and Research (BMBF),

consists of several applications designed to enable

communication between patients in intensive care and

hospital staff during the weaning process (weaning

from mechanical ventilation). Initially, only essen-

tial information about the current day (date, time,

weather) or location (hospital, city, nursing staff) is

given, more complex information about the therapy,

physical needs or pain is added in later phases.

Figure 1: ACTIVATE bed-side patient application con-

trolled via a ball-shaped novel interaction device.

The ACTIVATE system consists of (Kordts et al.,

2018) (Figure 1):

• an in-bed patient application (GUI)

• a novel interaction device BIRDY

• a related mobile app for care staff

Since the application is to be operated lying in bed

and the patient’s hands are impaired by infusion tubes

or swelling (e.g., from medication), conventional in-

put devices are not well suited. BIRDY (Ball-shaped

Interactive Rehabilitation Device) was therefore de-

veloped as a novel interactive device (Kordts et al.,

2018). BIRDY contains sensors and actuators and

has also been designed so that it can be used directly

on the patient’s body and supports cleanical disinfec-

tion. It has a rough surface to prevent slipping, is

light enough (99 g) to be operated without effort, and

has a size that can be easily grasped by one hand. A

charging station at the patient’s bedside can charge

two devices wirelessly at the same time. When ly-

ing down, the movement of the patient’s arms is re-

stricted. BIRDY is therefore operated with one hand

and only supports three gestures: left, right and select

(press) (Kopetz et al., 2019). The sensor values are

sent via Bluetooth to a single-board computer, which

interprets the gestures and forwards them to a PC at

the end of the bed, which in turn controls the user in-

terface of the patient application. Messages are also

sent to the staff app.

Typical texts of intensive care context have been

identiﬁed in workshops and made available in the ap-

plication via predeﬁned selection ﬁelds (Figure 1).

The actions in the ACTIVATE patient application are

arranged in a circle to work well with the limited in-

teractions of BIRDY. A virtual keyboard QWERTY

format (German version of the QWERTY layout) is

also provided for entering any text

3 STATE OF THE ART

First, we provide an overview of models for letter pre-

diction and then present comparable approaches for

virtual keyboards from literature.

3.1 Models

Neural networks have become increasingly popular in

recent years. For Natural Language Processing prob-

lems, there are now various architectures, like Long-

short term memory (LSTM) (Hochreiter and Schmid-

huber, 1997), Transformer (Vaswani et al., 2017), or

GPT-4 (Generative Pretrained Transformer) (OpenAI,

2023). All architectures have pros and cons (e.g., dis-

appearing gradients, only possible on small text cor-

pora, only trained in English), and the best approach

must be found for the given problem.

(Verwimp et al., 2017) developed an LSTM that

combines word and character embedding with repre-

sentations of the word or character as a number vec-

tors, respectively. In this approach, both the vector

representation of the word as well as that of the indi-

vidual characters are passed to the LSTM as input. A

maximum number of characters is set. The total size

of the embedding is kept constant, which is why the

word embedding shrinks and is partially replaced by

character embeddings, which reduces the number of

parameters and thus the size and complexity of the

model. Characters consist of a smaller vocabulary

than words, which is reﬂected in a smaller embedding

matrix. This special LSTM performs better than the

comparable word-level model.

(Mangal et al., 2019) investigated the speed of

LSTM, GRU (Gated Recurrent Unit) and bidirec-

tional RNN. They used the script of a TV series to

generate letters. They also implemented the models

with different numbers of layers. All three GRU mod-

els required the least time per iteration of individual

batches, closely followed by the LSTMs.

(Shakhovska et al., 2021) compared Markov

chain, LSTM, and a hybrid variant based on a group

of Ukrainian poems. The Markov chain required the

least training time and delivered the best results.

HEALTHINF 2025 - 18th International Conference on Health Informatics

Another word prediction network can be found in

the Google Gboard. (Hard et al., 2018) trained a Cou-

pled Input and Forget Gate (CIFG) network (Greff

et al., 2015) for this purpose, a variant of the LSTM

that predicts words in the virtual keyboard on the

smartphone. A network is trained on the server, and

ﬁne-tuned user-speciﬁcally on the smartphone. The

reason for this federated learning method is to im-

prove data protection.

(Suliman and Leith, 2023) questioned the concept

of federated learning and investigated whether it leads

to increased data protection. They came to the re-

sult that sentence input can be reconstructed from the

weights with a high degree of accuracy.

All works deal with the generation of words. Sci-

entiﬁc papers on models for generating individual let-

ters could not be found. In addition, the models men-

tioned were adapted according to their respective ar-

eas of application. This ranges from the creation of

television manuscripts to the generation of texts in

Ukrainian. For the use in intensive care context no

scientiﬁc work could be found.

3.2 Virtual Dynamic Keyboards

A virtual dynamic keyboard CanAssist for people

with disabilities was published in 2020

. The focus

was on text input for the existing eye-tracking sys-

tem. The letters were initially based on a dictionary

approach, but later on a statistical model.

(Pouplin et al., 2014) developed a virtual dynamic

keyboard for people with functional tetraplegia to ac-

celerate the speed of text input. They used either a

pointer mounted on the head as an input device, or

a scanning system that scans all relevant positions

on the keyboard one after the other, and the person

only needs a switch to conﬁrm the selection. This de-

pended on the person’s motor skills.

The software SibyLetter was used to predict the

next letters using 5-grams and then restructure the

keyboard accordingly (Schadle, 2004) (Wandmacher

et al., 2008). One corpus used was the newspaper “Le

Monde” with more than 100 million words. (Pouplin

et al., 2014) found that the speed of text input does

not increase with the virtual dynamic keyboard in a

one-month test with participants.

(Ljubic et al., 2014) described a dynamic key-

board with a modiﬁed form of input. Tilting move-

ments of a smartphone (right, left, up, and down)

leads to the selection of certain areas on the screen.

They compared the input speed for three types: (1)

https://www.canassist.ca/EN/main/programs/free-dow

nloads/dynamic-keyboard/dynamic-keyboard-developme

nt.html

One letter is highlighted on the standard keyboard,

(2) four letters are highlighted simultaneously and

re-arranged when selected, and (3) the keyboard is

halved after each selection until only four letters re-

main. All forms are based initially on the QWERTY

keyboard. The third input form was the fastest, but

showing only parts of keyboards might confuse users.

(Wojcik et al., 2018) describe a different keyboard

layout for people with motor impairments. The ﬁve

vowels (a, e, i, o, u) are arranged in a circle. All con-

sonants are not visibly located between these vowels

in alphabetical order (for example, b, c, and d are lo-

cated between a and e). They conducted a user study

with 20 users. Each user was asked to write 10 words

with their ﬁnger and 10 words with their ﬁst using

a virtual dynamic keyboard projected onto a touch-

sensitive mat (touchpad). Finger input was faster on

average, both for vowels and consonants.

(Kristensson and M

ullners, 2021) evaluated that

the best word prediction is for a word length of six

with three letters already entered.

(Agarwal et al., 2011) described a dynamic key-

board used for the secure input of passwords, without

spyware being able to intercept the keystrokes. In this

implementation, the letters on the virtual keyboard are

randomly shufﬂed (without probability prediction).

A patent by (Grifﬁn, 2013) describes a keyboard

with word predictions in which the words lie on the

same key as the next word to be entered. Visual high-

lighting is also described, such as displaying the let-

ter key in a larger size or a different color or gener-

ally display the key or the letter back-lit, underlined,

bolded, and/or italicized.

None of the works listed here used an input device

comparable to BIRDY with only three interaction op-

tions. The most similar are the four tilting movements

described in (Ljubic et al., 2014) and the scanning

system described in (Pouplin et al., 2014). The input

option is relevant for the design of the user interface

in order to optimize text input. The most similar user

interfaces compared to the circular menu of the ACTI-

VATE patient application are those of CanAssist and

(Wojcik et al., 2018).

However, the input devices are different and the

applications are adapted to them. (Pouplin et al.,

2014) found no improvement in the speed of the dy-

namic keyboard. However, the test subjects used this

keyboard less frequently than the familiar standard

keyboard. This leads to the conclusion that the pro-

cess may need a certain amount of practice and famil-

iarization time to get used to.

Virtual Dynamic Keyboard for Communication in Intensive Care

4 PROTOTYPE

For designing a useful dynamic keyboard, the most

probable next letters or words must be determined on

the basis of the previous inputs and then arranged at a

short distance to the current position.

For this, we compared different probability mod-

els and implemented the most successful model as a

prototype dynamic keyboard inside the ACTIVATE

user interface.

4.1 Concept

Extensive user surveys and workshops were already

carried out for the existing static ACTIVATE key-

board during the analysis phase of the project (Kordts

et al., 2018). Some of the requirements set out there

can also be adopted for the design of the dynamic

keyboard. For example, the system should be able to

be used independently as possible without help. Tu-

torials could provide assistance. The system should

be fault-tolerant and function automatically in order

to save care staff time. Unnecessary acoustic or vi-

sual disturbances should be avoided. The components

should be implemented interchangeably. Data com-

munication to a backend server should always be en-

crypted, user privacy should be protected as much as

possible, and the system should also be conﬁgurable

by staff.

Additional functional and non-functional require-

ments were also identiﬁed, which relate to the spe-

ciﬁc aspects of the dynamic arrangement of textual

elements on an interface.

At least four of the next most probable letters

should be determined and displayed in the immedi-

ate vicinity to the right and left of the cursor. These

should be clearly highlighted graphically. The system

should be based on the design of the ACTIVATE ap-

plication to prevent patients from getting used to it.

The display of the letters or words should be large

enough to be easily recognized from a distance of 2m

from the head of the bed. The backend should dis-

play the most likely next letters as well as words. The

keyboard should be able to respond correctly to in-

put via BIRDY. Other input devices should not be ex-

cluded. The voice output should be supported by a

corresponding read-aloud symbol. The existing tech-

nical implementation can be used for this purpose.

The next letters/words can only be output for the

German language. The system should make it possi-

ble to conduct hospital-speciﬁc conversations. Mod-

els should provide a result in less than a second. The

system may use the Internet and the models do not

have to be local on the device.

4.2 Corpora

The application is initially focused on German hos-

pitals, and therefore only German-language cor-

pora were considered, since foreign-language corpora

would not generate meaningful probabilities.

The speciﬁc use case of communication in the in-

tensive care units requires the support of typical sen-

tences in this environment. However, for the design

of the ACTIVATE application, the main communica-

tion needs of the patients have already been recorded

in numerous workshops and expert panels and made

available in the application by selection in circular

menus for various categories (e.g., “Ich bin durstig” (I

am thirsty) or ”Ich habe Schmerzen” (I am in pain)).

The keyboard should therefore support expressions

that are not yet covered by the existing system. These

can be conversations with doctors, nursing staff or

relatives. It is assumed that patients will not have

in-depth medical conversations directly after waking

up during the weaning process, and speciﬁc medical

terms may be irrelevant. It should also be noted that

medical corpora are extremely rare, are usually only

available in English and/or contain speciﬁc medical

terms without conversational context.

Corpus 1: SdeWaC The SdeWaC corpus

is an

extension of the deWaC corpus of the WaCky initia-

tive (Baroni et al., 2009). It is available free of charge

and contains more than 1.2 billion words from various

web sources. Only texts with the mime type text/html

and a size of 5 to 200 KB were retained.

Linguistically irrelevant texts such as warn-

ing messages or copyright declarations were also

deleted. The remaining texts were cleaned of HTML

and Javascript code, and duplicates (Baroni et al.,

2009). Format “One sentence per line” was se-

lected. An example is ”<year = “2007”><source

= ”10475”><error = ”0”>Mehr als 7,100 Arbeit-

nehmer sind Datenbank-Spezialisten.” When process-

ing the corpus for this application, the meta data, such

as <year>, <source> and <error> were deleted.

Corpus 2: GGPONC The GGPONC text corpus

is a specialist medical corpus. It was created in collab-

oration between the Hasso Plattner Institute for Dig-

ital Engineering gGmbH and the Friedrich Schiller

University in Jena (Borchert et al., 2020). The corpus

contains medical language based on the oncological

S3 guidelines. An example sentence from the cor-

pus is: ”Tabakkonsum ist ein wesentlicher Risikofak-

tor f

ur die Entwicklung des Mundh

ohlenkarzinoms.”

(Tobacco consumption is a signiﬁcant risk factor for

the development of oral cavity carcinoma.)

https://www.ims.uni-stuttgart.de/forschung/ressourc

en/korpora/sdewac/

HEALTHINF 2025 - 18th International Conference on Health Informatics

Corpus 3: Gutenberg-De The Gutenberg-DE

corpus

contains texts from the Gutenberg-De project.

A very small selection of copyright-free books was

made as a comparative corpus.

• Die Frau von dreißig Jahren by Honor

e de Balzac

translated by Hedwig Lachmann

• Student und Alkohol by Dr. Leopold Loewenfeld

• Das Haus by Lou Andreas-Salom

• Eiszeit und Klimawandel by Wilhelm B

olsche

The texts were chosen at random, but care was

taken to use modern language and spelling. Due to the

conditions of the German Copyright Act (UrhG, §64),

the texts are at least 70 years old. This means that the

new rules following the spelling reform of 1996 have

not been taken into account. As a result, the probabili-

ties of outdated special characters such as “ß” (instead

of “ss”) differ from current texts.

All three corpora were cleaned up: uppercase let-

ters are converted to lowercase, and all characters that

are not a-z

uß, digits 0-9, or space are removed.

The above example results in ”mehr als 7 100 arbeit-

nehmer sind datenbank spezialisten”. The semantic

meaning may change as a result of the ﬁltering. For

the statistical interpretation in the dynamic keyboard,

however, the semantic distortion is irrelevant, as spe-

cial characters are not supported anyhow.

4.3 User Interface

Based on the functional requirements, there are sev-

eral options for keyboard design.

The original idea of arranging the letters in a circle

(Fig. 2) was ruled out during the design process. The

circular menu in the ACTIVATE patient application

includes a limited set of elements for close communi-

cation and information display. A circular keyboard

could appear overloaded with a total of 30 letter keys,

the space bar, and additional function keys (such as

Back and Delete) and exceeds the screen size.

t e

b l

v p

ß n

m s

Zange

Zwiebel

Zeitung

Ich möchte

etwas Anderes

mitteilen

Haben Sie eine Z

Figure 2: Circular design layout of the dynamic keyboard.

Left: dedicated words on keys. Right: Grouping of letters.

https://www.projekt-gutenberg.org/info/texte/allwor

ka.html.

Alternatively, only a limited number of letter keys

could be displayed, which would mean that not all

words could be written. A possible solution could use

multiple letters per key as in CanAssist (Fig. 2). How-

ever, it requires at least two actions per letter, because

ﬁrst the group must be selected, and then the letter

must be selected after redistributing the letters con-

tained to a new circle.

Most people are familiar with a line-shaped key-

board. When entering words with the BIRDY input

device, it does not matter whether the letters are ar-

ranged in a circle or lines, as BIRDY does not support

up and down movements to switch between lines.

Figure 3: Linear design layout of the dynamic keyboard.

Letters and characters as well as action keys are

displayed in lines, just like on a traditional computer

keyboard. A left movement from the ﬁrst letter of the

ﬁrst line enables a jump to the last letter of the last

line and vice versa by a right movement. A keyboard

that is arranged in lines therefore has the same func-

tionality as a circular keyboard but can be displayed

more clearly on a screen.

The size of the letters has been adopted from the

static keyboard. The currently most likely letter is dis-

played larger and with a dark background in the mid-

dle. The next most likely letters are placed to the right

and left around it, and can therefore be reached with

just a few gestures.

All 30 letters ’a-z

uß’ must be displayed so that

all conceivable texts can be entered. The less likely a

letter is, the further away it is positioned. In Fig. 3

these are the letters ’ß’ at the top left and ’x’ at the

bottom right. There is also a delete button on the far

left. The distance should not play a major role in ev-

eryday use, as the correctness of the input is of sec-

ondary importance. At the right end there is a space,

the symbol for starting the read-aloud function, and a

back button.

In a circular keyboard, it would still be conceiv-

able to display complete words on keys, either by

omitting or by grouping letters (Fig. 2). In a line-

shaped keyboard, only four options remain:

1. directly after the next n most probable letters on

the keyboard

2. at the end after the least probable letters

3. at a different position, e.g., at the top or bottom

4. next to its ﬁrst letter

Virtual Dynamic Keyboard for Communication in Intensive Care

With all four options, the keyboard may look too

cluttered for the patient. In option 1, a visual break is

made and words or letters would change in a loose se-

quence. In option 2, the path to the words is probably

so long that it would not be worthwhile for patients to

navigate to them, especially for short words. Option 3

has a similar problem to option 2, but could be a better

alternative. Option 4, like option 1, also has the prob-

lem of a visual break. With each option, the keyboard

appears cluttered with additional words, and we have

decided against using complete words.

4.4 Implementation

The dynamic keyboard was implemented as a client-

server architecture with Docker containers.

The Markov model and the LSTM are trained on

the server. The Transformer model is a pre-trained ex-

ternal model that is accessed via an external interface.

For the probability models, we only use 30 lowercase

letters (a-z

oß), 10 digits and the space character,

i.e. a total of 41 characters. This means that the char-

acter set is signiﬁcantly smaller than the word set of

the German language, and the training of the models

requires considerably less time and memory.

The Markov chain is a stochastic process that is

trained quickly (in a few hours), and can answer

queries quickly. For this purpose, a histogram of N-

grams is determined for the respective corpus (using

the ngrams() function of the Python nltk9 package

We have implemented bigrams (N=2) and trigrams

(N=3) for this purpose. The calculation is aborted as

soon as saturation occurs, i.e. the probabilities of the

N-grams no longer change or change only slightly. If

certain trigram combinations did not occur in the se-

lected corpus, the letters were nevertheless assigned

to keys, but with a low probability. After estimation,

the dictionary is stored in a document database (Mon-

goDB

) to avoid re-training in subsequent calls.

The Transformer-based language model GPT

from the company OpenAI has been widely known

since ChatGPT

at the latest. We used a pre-trained

GPT2 network benjamin/gerpt2-large (Minixhofer,

2020) based on German texts. The network com-

pletes a partial sentence with several words, often

even an entire paragraph. Minixhofer uses the Au-

toModelForCausalLM model. The number of gen-

erated words can be set in the model using the

max new tokens parameter (we tested the values 1,

2 and 10). For the dynamic keyboard, the network’s

return was adapted so that only individual letters are

ttps://www.nltk.org/api/nltk.util.tml

https://www.mongodb.com/de-de

https://openai.com/

used. Transformer nets are not deterministic. To gen-

erate a letter probability from the sentence prediction,

the network was called several times.

The LSTM was implemented based on the Py-

Torch Tutorial

, which supports both word and letter

predictions. Despite the relatively high-performance

computer (GPU: NVIDIA RTX 3090 with 24 giga-

bytes of RAM, CPU: AMD Ryzon 9 5950X with

128 gigabytes of RAM) and splitting the corpus of

more than 5 GByte into ﬁles of 180-360 MByte, the

computing time for the training based on the cleaned

SdeWaC corpus took several months. Due to the rela-

tive similarity of the trigrams based on the Gutenberg-

DE corpus, further training was only carried out with

this data. The corpus has a size of 8,778 lines and a

total of 166,416 words and slightly more than 1 mil-

lion letters, and the training could be completed in a

few hours. The trained LSTM was persisted in a zip

ﬁle.

The server was implemented in Python (version

3.10) and offers a RESTful API with ﬁve endpoints,

which accesses the respective pre-trained data:

• give next letter using Transformer

• give next letter using Markov chain bigram

• give next letter using Markov chain trigram

• give next letter using LSTM

• write log ﬁle

In addition, code for the software-side evaluation

of the predictions based on test data sets has also been

implemented on the server.

The graphical user interface (client) is imple-

mented as an extension of the ACTIVATE patient

application with the help of the technologies used

there (HTML, CSS, vanilla JavaScript and JavaScript

framework Vue.js (Kopetz et al., 2021)). The client

is based on the vue-keyboard package

and asks the

server for a list of the next probable letters. The list

returned by the server is transformed so that the prob-

abilities of the letters decrease from the middle to the

ends of the list. The elements of the transformed list

are then distributed to three rows of keys on the virtual

keyboard.

5 EVALUATION

We performed several steps of comparison and tests

to evaluate our prototype implementation. First, we

https://closeheat.com/blog/pytorch-lstm-text-generat

ion-tutorial

https://github.com/MartyWallace/vue-keyboard,Vers

ion3.1.0

HEALTHINF 2025 - 18th International Conference on Health Informatics

compared the theoretical minimum of steps for the

different corpora to identify the ideal training data set.

Second, we compared the models to identify the best

forecasting system. Third, we performed a user test

to evaluate human interaction.

5.1 Corpora Comparison

In an online survey conducted beforehand, partici-

pants were asked to write down texts they could imag-

ine themselves saying as a patient in a weaning pro-

cess. 361 example sentences from 74 participants

were cleaned from redundancy and special characters

resulting in a test set of 257 sentences used to com-

pare with the three other different corpora introduced

above. The Markov model with trigrams is used for

the comparison. The next step is to check whether a

specialist medical corpus, a high-quality literary text

or a corpus collected from the Internet can change the

results and further reduce the number of steps. This

refers to movements to the right or left on the key-

board. The average number of steps calculated across

all test sentences when using trigrams can be seen in

Table 1. In one case, the average number is lower than

one. This is based on cases, where the predicted letter

placed in the middle is correct, and therefore, no steps

had to be performed.

Table 1: Average number of steps (left, right - pressing ig-

nored) when using trigrams across all test sets.

Korpus Number of steps ∅

SdeWaC 1.59594697

GGPONC 1.95818220

Gutenberg-De 1.59026818

Example sentences 0.95412386

The difference between the SdeWaC and the med-

ical corpus is relatively large. The medical corpus

is similarly poor with trigrams as the normal corpus

with bigrams. The selected medical corpus contains

specialized medical text on carcinomas based on the

oncological S3 guidelines. Words such as laryngeal

carcinoma, screening, genotyping, cytology, etc. are

not rare words in the corpus. However, these words

are probably not used by people in the weaning pro-

cess. Otherwise, they would already be available as

standard text in the ACTIVATE patient application.

The example sentences from the survey also mainly

contain everyday language. The medical terms used

in the survey are weaning process, ventilator, treat-

ment, pain and healing process. The last three terms

(or a modiﬁed form of these) are probably more com-

monly used normally in everyday language. For these

reasons, the SdeWaC corpus ﬁts the use case better.

Nevertheless, it should be mentioned that the sen-

tences from the survey are only a snapshot, and the

sentences may still differ from those from the real sit-

uation in the weaning process.

The average number of steps in the SdeWaC and

Gutenberg-De corpora are very similar. The corpora

have no textual overlap. The SdeWaC contains both

ﬁctional and non-ﬁctional texts. Although some texts

are informative texts from news and vacation desti-

nation websites, they are not recognizable specialist

literature websites like the texts in the GGPONC cor-

pus. The Gutenberg-De corpus is based on ﬁctional

texts (stories, novels). The non-ﬁctional texts there-

fore resemble the ﬁctional Gutenberg texts with lit-

tle technical language in terms of N-gram distribu-

tion. In the SdeWaC corpus, some technical terms

from medicine can be found, but this is the exception.

Therefore, SdeWac and GGPONC have different let-

ter probabilities, and SdeWaC and Gutenberg-De cor-

pus have similar numbers of steps. The average num-

ber of steps and the step frequency distribution are

best for the test sets.

5.2 Model Evaluation

In the following, we present comparison results of let-

ter forecast performance of each model.

5.2.1 Markov

The developed system is tested with sample sentences

from the survey. The best input scenario is assumed

for both the standard keyboard and the dynamic key-

board, i.e. simulating a user without errors, and not

using back keys. With the standard keyboard the start

position is at letter ’1’, with the dynamic keyboard the

initial setup is starting with letter ’d’ in the middle.

As expected, the standard keyboard is very inef-

ﬁcient. The average number of steps per letter tested

with the example sentences is 13.83 for the standard

keyboard, 2.07 for the Markov model bigram and 1.6

for the Markov model trigram (see Fig. 4) For both

the bigram and the trigram, the probability that the

next letter is 0 steps or one step away is higher than

50%. With the standard keyboard it is very arbitrary

and the probabilities from 0 to 13 steps add up to more

than 50%.

The measured time difference of the system time

before and after the function call is 0.03 seconds on

average (rounded), well below the maximum accept-

able delay between input and feedback of 500 ms

speciﬁed by (Wolff et al., 2011). The bi- and trigrams

for the Markov model were created from the SdeWaC

corpus and stored in the database in a Docker con-

tainer.

Virtual Dynamic Keyboard for Communication in Intensive Care

Figure 4: Probability of steps based on trigrams for dynamic (top, SdeWaC-corpus) versus static keyboard (bottom).

5.2.2 Transformer

The pre-trained GPT2 network is designed to gen-

erate several words up to sentences and paragraphs.

For this work, however, the network should predict

the next best possible letters. If no context is given

to the network, but it is to generate further words

with a small number of given letters, special char-

acters are often returned. These are initially irrele-

vant for the dynamic keyboard. With the parameters

bad words ids and max new tokens and the context

sentence, ”Ich liege im Krankenhaus und werde beat-

met. Jetzt m

ochte ich mit dem Arzt kommunizieren.“

(I’m in hospital and being ventilated. Now I want

to communicate with the doctor.) which was added

before each example sentence, acceptable sentences

could be generated. The ﬁrst generated letter is cho-

sen as one of the likely next letters. To generate more

than one letter, the same sentence must be given to

the model multiple times (iteration). After each pre-

dicted letter, the iteration counter decreases by one. If

the same letter is predicted multiple times, its prob-

ability increases, but the number of different letters

remains the same. To prevent an endless loop, the

maximum number of iterations is set to 20. With 30

letters in the german alphabet and one space char-

acter, the most distant letter is 15 steps away from

the pointer (focus in the middle). This is half of the

possible characters (a-z

uß and space). The average

number of steps for a sentence is 6.09 (rounded) for

the transformer without context and 6.06 (rounded)

for the transformer with context. The step frequency

distribution for the transformer does not show a uni-

form decrease in probability like the Markov model

trigram, as it is not based on a probability distribu-

tion but learned letter patterns. Additionally, we limit

generation to the ten most probable next letters to re-

duce processing time, resulting in a maximum of ﬁve

steps. If the results are compared with context and

without context, it can be seen that steps zero and one

are 2% more likely with context. From six steps at

the latest, all remaining characters that were not gen-

erated by the transformer are added with no particu-

lar order. This means that a high probability of steps

greater than ﬁve is possible, but can change with each

evaluation.

The average processing time required is 17.84 sec-

onds for the Transformer with context and 16.51 sec-

onds without the speciﬁed hospital context. The times

measured for the Transformer exceed the maximum

feedback value speciﬁed by (Wolff et al., 2011) by

far. Times for using the GUI would add to this.

5.2.3 LSTM

The evaluation of the LSTM resulted in an average

of 5.84 (rounded) steps for one sentence. One step is

the most frequent, followed by two and ﬁve steps. The

distribution does not resemble the bi- or trigram distri-

bution of the Markov model. Unlike the Transformer,

no special characters are generated because the train-

ing corpus does not contain them. The LSTM only

generates a maximum of the next ten possible letters.

All other existing letters in the alphabet (n=21) are

appended to the list of generated letters. This means

that all letters can still be entered using the keyboard.

Ten generated letters on the keyboard means that a

maximum of ﬁve steps are required to reach them.

HEALTHINF 2025 - 18th International Conference on Health Informatics

The step frequency distribution of the LSTM

shows that the sum of the probabilities for zero to ﬁve

steps is 57.98%, which is more than half of all step

probabilities (steps 0-15). However, the result of the

LSTM is still worse than that of the Markov model

(bi- and trigram). With the bigram, the ﬁrst two steps

(0,1,2) already have a probability of 71.4% and with

the trigrams even 80.03%. The measured processing

time is 1.33 seconds (rounded). The response time

is signiﬁcantly better than the transformer network’s,

but still too high to be acceptable.

Since the transformer model was rather slow and

created a lot of special characters that have to be ﬁl-

tered, and the LSTM model provided lower predic-

tion quality, we decided to integrate the determinis-

tic Markov model into the dynamic keyboard applica-

tion.

5.3 User Evaluation

In the laboratory test described above, perfect users

without any time spent on the gestures and without

errors in selecting the correct letters, were simulated.

For the usability test of the user interface and its use

by real people, we also conducted an initial labora-

tory study with test subjects. The aim was to test the

effectiveness of the new design in comparison with

the static keyboard, as well as the perception of the

changing letters on the keyboard among participants.

The qualitative study uses a mix of methods that

includes observing the interactions of the test sub-

jects, in particular their movements within the appli-

cation, as well as a structured questionnaire afterward.

The semi-randomized study took place in our lab-

oratory, where a hospital-like (intensive care) envi-

ronment was created. The ACTIVATE application

and BIRDY were provided at a hospital bedside. The

inclination of the head section and the positioning of

the screen were adapted to real conditions in the in-

tensive care unit. The test person lies in bed and holds

BIRDY in his hand while it lies on the bed.

Voluntary test subjects (n=9) were recruited for

the study from among the students and employees of

our university using a survey in our learning manage-

ment system. Upon request, student participants re-

ceived test points for the 30-minute test.

Inclusion criteria included a minimum age of 18

years and German language skills. Participants with

physical impairments of arms and legs, dyslexia, vi-

sual impairment or blindness, mental impairments,

and other illnesses that could impair the successful

completion of the task (e.g., epilepsy) were excluded.

After the purpose of the study has been explained

to the subjects and presented to them in writing, they

signed a declaration of consent and a data protection

declaration. The test person then lies down in the hos-

pital bed and is given BIRDY in the hand. BIRDY

is ﬁrst calibrated with the help of the study staff. In

particular, the magnetometer to determine the Earth’s

reference system is affected by constant interference

and must be adjusted. The accelerometer and the gy-

roscope are less affected.

The system was brieﬂy explained to the test per-

son and the basics are shown. This was followed by

the sentence input. To familiarize themselves with

the system, the test subjects select three sentences

provided by the ACTIVATE system: “Wann kommt

meine Familie?” (“When is my family coming?”),

”Was ist passiert?” (“What happened?”), and ”Durst”

(‘Thirst’) via the patient application.

Three further sentences were selected from the re-

sults of the survey, to avoid long input sequences:

“danke” (“thank you”), “wer sind sie” (“who are

you”), and ”ich liebe euch” (“I love you”). Please

note, that the sentences are not following correct Ger-

man grammar and spelling, since the keyboard does

not provide capital letters. This is obviously of minor

importance in the context of intensive care communi-

cation. The sentences are entered by the test subjects

using both the dynamic and the standard keyboard.

In the background, the program measures the sys-

tem time and records the number of BIRDY actions

(step to the right, left, and conﬁrmation by push ges-

ture). After successful completion, the test subjects

ﬁll out a questionnaire. The questionnaire is pseudo-

anonymous and does not ask for any personal data.

Participants can indicate wishes for further functions

or additional characters.

Nine test subjects registered for and participated

in the laboratory study. The age of participations was

between 21 and 36 years, the average was 26.5 years.

Six of the test subjects felt they belonged to the female

gender, and three test subjects to the male gender. All

test subjects successfully completed the ﬁrst task of

entering sentences with the ACTIVATE patient appli-

cation. This task was intended to familiarize them

with the system and the input device and is not dis-

cussed further here. In the second task, the test sub-

jects were asked to enter sentences using the dynamic

keyboard. This task was also successfully completed

by all participants.

In the third task, entering three sentences using

the standard keyboard, two test subjects had technical

difﬁculties with the BIRDY input device, so that this

could not be completed. For this reason, we decided

to discontinue the respective tests. This affected the

sentences ”wer sind sie” und ”ich liebe euch” for two

test subjects.

Virtual Dynamic Keyboard for Communication in Intensive Care

In the step evaluation, the steps were counted from

opening the respective keypad to starting the pronun-

ciation via the corresponding symbol. With the dy-

namic keyboard, the symbol was always 16 steps (15

right and 1 pressing movement) away from the last in-

put, because the pointer (highlighted letter) is always

in the center of the dynamic keyboard (see above).

The distance to the symbol on the standard keyboard

depends on the input of the last letter. It should also

be noted here that the test person can take a step to

the left from the ﬁrst keyboard symbol at the top left

(number 1) to land at “Zur

uck” (Back) at the bottom

right (3). The steps taken per participant, sentence and

keyboard are compared with the best possible steps.

For the standard keyboard in particular, a movement

from top left to bottom right and vice versa was con-

sidered if this path was the shorter one.

5.3.1 Results: Number of Steps

The shortest path for the ﬁrst sentence “danke”, from

opening the dynamic keyboard to the start of the out-

put requires 27 steps. Of these, 11 steps were required

to type the word/phrase and 16 steps to reach and acti-

vate the symbol. Three out of nine test subjects man-

aged it in 27 steps, three were slightly worse (+2, +4,

+6 steps) and three needed signiﬁcantly more steps.

For the second sentence “wer sind sie”, a mini-

mum of 36 (20+16) steps were required to enter the

text without errors. Four test subjects managed this.

Two were slightly above this (+2 steps) and three test

subjects needed more steps to enter the text. The third

sentence to be entered, “ich liebe euch”, required a

minimum of 57 (41+16) steps. One test person made

it, four were slightly above that (+2, +4) and four

missed the target by a long way. For two of the last

group, the problem was with the input device, which

no longer responded as desired. These were the test

subjects who dropped out later.

None of the nine test subjects managed to enter

all sentences correctly using the dynamic keyboard

without errors. This conﬁrms that it is important to

conduct a laboratory test with people to evaluate the

system instead of just using software to calculate the

smallest steps.

With the standard keyboard, no test person man-

aged to enter one of the three sentences with the

minimum steps. These were 74+16=90 for “danke”,

155+16=171 for “wer sind sie” and 221+18=239 for

”ich liebe euch” (the second summand is the path to

the read-aloud symbol). As already mentioned, the

steps to the symbol depend on the previous letter. In

the ﬁrst two sentences, by chance, there are also 16

steps (from the letter ’e’) and for the last sentence, it

is 18 steps (starting from ’h’). Only one test subject

managed to get close to the best possible number of

steps. This person managed 97 instead of 90 steps in

the ﬁrst set and 244 instead of 239 steps in the last set.

Two test subjects dropped out during the last two sets,

leaving only 7 relevant test subjects at the end,

Consequently, the dynamic keyboard offered the

test subjects considerable added value in terms of

compared to the standard keyboard. The dynamic

keyboard requires fewer steps per sentence.

5.3.2 Results: Required Time

The evaluation of the required time is somewhat more

complex, as there is no minimum reference time.

For different reasons, each person takes a different

amount of time to enter text into the keyboards.

In total, with the dynamic keyboard, it was pos-

sible to enter all 27 text entries. 15 sentences could

be entered in less than one minute. The ﬁrst two sen-

tences in particular (“danke”, “wer sind sie”) required

a shorter input time. With the standard keyboard, only

two entries could be entered in under a minute. When

calculating the average duration, the two test subjects

who aborted were not taken into account. The previ-

ous sentences already reﬂect the difﬁculties with the

BIRDY input device. The two test subjects waited in

between, hoping that BIRDY would recover after a

waiting period. The result can be seen in table 2.

Table 2: Average time needed to input the sentences once

with the dynamic keyboard and once with the standard key-

board (N=7).

Time format mm:ss,ms

Keyboard

Sentences Dynamic Standard

danke 00:42,096 01:21,347

wer sind sie 00:40,601 03:12,840

ich liebe euch 00:58,927 03:08,014

The small sample size does not allow a quantita-

tively signiﬁcant statement to be made. Nevertheless,

it can be seen that typing with the dynamic keyboard

is faster in all cases than with the standard keyboard.

5.3.3 Results: Questionary

A questionnaire with seven questions was presented

to the subjects after the tests. All test subjects were

comfortable with the dynamic keyboard and would

continue to use it instead of the standard keyboard.

Only one person had already used a different dynamic

keyboard before, for all others it was new.

Six out of nine test subjects would like to see more

characters on the keyboard. Numbers were mentioned

in particular, periods, commas, question and exclama-

HEALTHINF 2025 - 18th International Conference on Health Informatics

tion marks. positioning of additional characters were

preferred dynamically (n=3), statically at the edge of

the keyboard (n=2), or dynamically only at the poten-

tial end of a sentence (n=1).

It was also suggested that a number written in full

could be automatically converted into a sequence of

digits. But there is no added value in this case. The

text entered would only be shortened, which is irrel-

evant for pronunciation. There would only be an ad-

vantage if the input of a number is recognized early.

Especially it could perhaps be an advantage with long

numbers.

Eight out of nine test subjects see an additional

beneﬁt when using the dynamic keyboard compared

to the ACTIVATE patient application. The reasons

were similar for most of them. They stated that more

individualized utterances were possible, and less cog-

nitive effort or less concentration would be required.

All of them saw an additional beneﬁt of the dynamic

keyboard compared to the existing standard keyboard.

The dynamic keyboard is faster to type (provided the

word is recognized correctly) and more intuitive.

Vertical movements by BIRDY were desired. This

presumably related to the standard keyboard, as the

letters of the laboratory test sentences on the dynamic

keyboard were almost always on the same line (with

the exception of the read-aloud symbol). Vertical

movements were ruled out in the design of the device

in user studies as not comfortable for the position in

bed.

An idea was described to have another input de-

vice that takes over the function of the space bar, the

delete key, and the pronunciation function. Although

this could speed up input, it would lead to confu-

sion for patients. The input device would have to be

used in parallel with BIRDY in the other hand and the

patient would have to memorize which input device

is responsible for device is responsible for what and

which gestures exist if they are different.

Suggestions were also made to predict and display

words or to be able to memorize words and sentences.

Predicting words and integrating them into the key-

board has already been implemented in variants of

the prototype. In this case, a ﬁxed number of words

was dynamically added to the keyboard, depending on

where the initial letter of the word was located. This

is similar to the patent application by (Grifﬁn, 2013).

6 CONCLUSIONS AND

OUTLOOK

This work focuses on the development and evalua-

tion of a dynamic virtual keyboard for intensive care

units. Integrated into the ACTIVATE project’s com-

munication system, the keyboard uses the BIRDY in-

put device, designed for circular menus with sim-

ple left, right, and pressure gestures. For traditional

keyboard layouts, this control is cumbersome and re-

quires many movements.

The dynamic keyboard uses statistical models to

predict letters based on previous input, dynamically

rearranging their positions to minimize interactions.

Three models were implemented: Markov, LSTM,

and Transformer. The Markov model with trigrams

proved most efﬁcient, averaging 1.6 steps per letter

compared to 13.83 for a standard keyboard. In a lab-

oratory evaluation, users typed 73% faster with the

dynamic keyboard and preferred it over the standard

one.

The design of the GUI layout was not the focus.

And although graphic cues like highlighting letters

as well as optional voice output are already provided,

user-friendliness can further be optimized through an

ISO-compliant standard UCX process. Especially, an

integrated tutorial system would be beneﬁcial for be-

ginners.

Continuous model adaptation could enable user-

speciﬁc customization. While it may be straight-

forward to simply adjust the Markov-Model based on

the patient typing history over time and support real-

time adaptation, it may even degrade prediction qual-

ity due to patients’ errors in typing. Since manual cor-

rection is not a realistic option, another speech model

for automatic syntax correction would be required.

But most patients will not type much text anyhow,

therefore the potential advantage is limited.

Larger user tests are planned to conﬁrm results,

with a future clinical study to determine effectiveness

for real patients.

ACKNOWLEDGEMENTS

The authors would like to express their gratitude to

ACTIVATE research team members Boerge Kordts

and Jan Kopetz for supporting the integration of our

results into the ACTIVATE system architecture and

for very helpful concept discussions.

REFERENCES

Agarwal, M., Mehra, M., Pawar, R., and Shah, D. (2011).

Secure authentication using dynamic virtual keyboard

layout. In Proceedings of the International Con-

ference and Workshop on Emerging Trends in Tech-

nology, ICWET ’11, page 288–291. Association for

Computing Machinery.

Virtual Dynamic Keyboard for Communication in Intensive Care

Baroni, M., Bernardini, S., Ferraresi, A., and Zanchetta,

E. (2009). The wacky wide web: A collection of

very large linguistically processed web-crawled cor-

pora. Language Resources and Evaluation, 43:209–

226.

Borchert, F., Lohr, C., Modersohn, L., Langer, T., Foll-

mann, M., Sachs, J. P., Hahn, U., and Schapranow,

M.-P. (2020). GGPONC: A corpus of German med-

ical text with rich metadata based on clinical prac-

tice guidelines. In Proceedings of the 11th Interna-

tional Workshop on Health Text Mining and Informa-

tion Analysis, pages 38–48, Online. Association for

Computational Linguistics.

Destatis, S. B. (2018). Gesundheit. Grunddaten der

Krankenh

auser 2017. Fachserie 12 Reihe 6.1.1. Ar-

tikelnummer 2120611177004.

Greff, K., Srivastava, R., Koutn

ık, J., Steunebrink, B.,

and Schmidhuber, J. (2015). LSTM: A search space

odyssey. IEEE transactions on neural networks and

learning systems, 28.

Grifﬁn, J. T. (2013). Next letter prediction for virtual key-

board. https://worldwide.espacenet.com/patent/searc

h-/family/046875664/publication/US9134810B2?q

=pn%3DUS9134810B2.

Hard, A. S., Rao, K., Mathews, R., Beaufays, F., Augen-

stein, S., Eichner, H., Kiddon, C., and Ramage, D.

(2018). Federated learning for mobile keyboard pre-

diction. ArXiv, abs/1811.03604.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural computation, 9:1735–80.

Jurafsky, D. and Martin, J. H. (2023). N-gram Language

Models, chapter 3. Prentice Hall, 3rd edition. Draft.

Kopetz, J., Jochems, N., Henkel, A., Schley, A., Balzer,

K., Kordts, B., and Schrader, A. (2021). ACTI-

VATE. Sozio-technisches System zur Unterst

utzung

der Kommunikation von Intensivpatienten. BMBF

Forschungsvorhaben, Universit

at zu L

ubeck. Final re-

port.

Kopetz, J. P., Burgsm

uller, S., Vandereike, A.-K., Sengpiel,

M., Wessel, D., and Jochems, N. (2019). Finding user

preferences designing the innovative interaction de-

vice “birdy” for intensive care patients. In Bagnara,

S., Tartaglia, R., Albolino, S., Alexander, T., and Fu-

jita, Y., editors, Proceedings of the 20th Congress of

the International Ergonomics Association (IEA 2018),

pages 698–707, Cham. Springer International Pub-

lishing.

Kordts, B., Kopetz, J. P., Balzer, K., and Jochems, N.

(2018). Requirements for a system supporting patient

communication in intensive care in germany. In Boll,

S., Hein, A., Heuten, W., and Wolf-Ostermann, K., ed-

itors, Zukunft der Pﬂege : Tagungsband der 1. Clus-

terkonferenz 2018 - Innovative Technologien f

ur die

Pﬂege, pages 131–136. BIS-Verl. der Carl von Ossiet-

zky Universit

at Oldenburg.

Kristensson, P. O. and M

ullners, T. (2021). Design and

analysis of intelligent text entry systems with function

structure models and envelope analysis. In Proceed-

ings of the 2021 CHI Conference on Human Factors

in Computing Systems, CHI ’21. Association for Com-

puting Machinery.

Ljubic, S., Glavinic, V., and Kukec, M. (2014). Predicting

upper-bound text entry speeds for discrete-tilt-based

input on smartphones. Journal of Interaction Science,

Mangal, S., Joshi, P., and Modak, R. (2019). LSTM

vs. GRU vs. bidirectional RNN for script generation.

CoRR, abs/1908.04332. http://arxiv.org/abs/1908.043

32[Accessed:(25.09.2024)].

Minixhofer, B. (2020). GerPT2: German large and small

versions of GPT2. https://github.com/bminixhofer/g

erpt2 [Accessed: (25.09.2024)].

OpenAI (2023). Gpt-4 technical report. https://arxiv.org/ab

s/2303.08774 [Accessed: (25.09.2024)].

Pouplin, S., Robertson, J., Antoine, J.-Y., Blanchet, A.,

Kahloun, J., Volle, P., Bouteille, J., Lofaso, F., and

Bensmail, D. (2014). Effect of a dynamic keyboard

and word prediction systems on text input speed in

patients with functional tetraplegia. Journal of reha-

bilitation research and development, 51:467–480.

Schadle, I. (2004). Sibyl: Aac system using nlp techniques.

In Miesenberger, K., Klaus, J., Zagler, W. L., and

Burger, D., editors, Computers Helping People with

Special Needs, pages 1009–1015, Berlin, Heidelberg.

Springer.

Shakhovska, K., Dumyn, I., Kryvinska, N., and Kagita,

M. K. (2021). An approach for a next-word prediction

for ukrainian language. Wireless Communications and

Mobile Computing, 2021:1–9.

Suliman, M. and Leith, D. (2023). Two models are better

than one: Federated learning is not private for google

gboard next word prediction. In European Symposium

on Research in Computer Security, pages 105–122.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J.,

Jones, L., Gomez, A. N., Kaiser, L., and Polo-

sukhin, I. (2017). Attention is all you need. CoRR,

abs/1706.03762.

Verwimp, L., Pelemans, J., Van hamme, H., and Wambacq,

P. (2017). Character-word LSTM language models.

In Proceedings of the 15th Conference of the Euro-

pean Chapter of the Association for Computational

Linguistics: Volume 1, Long Papers, pages 417–427,

Valencia, Spain. Association for Computational Lin-

guistics.

Wandmacher, T., Antoine, J.-Y., Poirier, F., and D

eparte,

J.-P. (2008). Sibylle, an assistive communication sys-

tem adapting to the context and its user. ACM Trans.

Access. Comput., 1(1).

Wojcik, B., Morelli, T., and Hoeft, B. (2018). RingBoard

– A Dynamic Virtual Keyboard for Fist Based Text

Entry. Journal on Technology and Persons with Dis-

abilities, page 83.

Wolff, S., Kohrs, C., Scheich, H., and Brechmann,

A. (2011). Auswirkungen von prosodisch-

motivationalen und verz

ogerten R

uckmeldungen

auf die Lernleistung und Hirnaktivitat in einer

Mensch-Computer Interaktion. In INFORMATIK

2011, pages 238–238. GI e.V., Bonn.

HEALTHINF 2025 - 18th International Conference on Health Informatics