Review of electrooculography-based human-computer interaction: recent technologies, challenges and future trends

Electrode arrangement

In the field of EOG-based human-computer interaction, the arrangement and configuration of electrodes are greatly influenced by the research objectives. Based on the motivations of various studies, the electrode configurations in EOG-HCI research can be broadly categorized into three types: standard arrangement, task-oriented arrangement, and wearable-oriented arrangement. We have conducted a statistical analysis of the advantages, disadvantages, and the number of literature using each type of electrode configuration. The results are presented in Table 2.

Figure 4 shows the EOG standard electrode arrangement, with a and b representing horizontal EOG electrodes which produce the horizontal EOG signal (HEOG) and c and d representing vertical EOG electrodes which produce the horizontal EOG signal (VEOG).

Figure 4. Standard Electrode Arrangement for EOG.

In task-oriented electrode arrangement, Li et al. used only the EOG signal related to blinking to perform commands on a graphical user interface (GUI)^[5,18-20,32]. For this task, they measured only the vertical EOG signal using a single electrode. Kim et al. designed an interaction task where users answered questions by using left-right horizontal eye movements^[37]. Therefore, the interaction system involved a pair of electrodes placed approximately 2.5 cm outside the corner of each eye.

In wearable-oriented electrode arrangement, Ryu et al. adopted a glasses-based electrode positioning scheme that adapts to the structure of the glasses frame^[13]. Heo et al. designed a wearable EOG system in the form of a headband, placing four electrodes on the forehead with vertical channel electrodes arranged vertically and horizontal channel electrodes arranged horizontally, reducing the number of electrodes by sharing a common positive electrode^[32].

Acquisition hardware devices

EOG signals are non-periodic, with a frequency range typically between 0 to 50 Hz^{[30,35,39-41]}. According to the Nyquist criterion, the sampling frequency must be at least twice the maximum frequency component of the signal to avoid aliasing. Moreover, EOG signals are susceptible to interference from muscle activity and movement noise, necessitating amplification and filtering. For interactive hardware devices, user safety and comfort are also important considerations.

In general, EOG acquisition hardware can be categorized into custom-designed devices and commercial devices. In this review, 19 studies opted to develop custom-designed devices tailored and optimized to meet specific research requirements and experimental scenarios. In contrast, other studies utilized commercial devices such as NuAmps(Compumedics Neuroscan, Inc., Abbotsford, Australia), NeuroSky MindWave Mobile Headset (NeuroSky, San Jose, CA, USA), Neuroscan (Neuroscan Ltd., TX, USA), Enobio-32 EEG system (Neuroelectrics, Barcelona, Spain), BIOPAC MP150 (BIOPAC Systems, Goleta, CA, USA), EOG Mindo devices(Mindo, National Chiao Tung University Brain Research Center, Taiwan, China), gUSBamp amplifier (g.tec Medical Engineering GmbH, Austria) and the ActiveTwo biopotential recording system(Biosemi, Amsterdam, Netherlands). Self-designed EOG acquisition hardware involves systems carefully designed by researchers to align with their research goals. Many researchers integrated circuits into wearable devices to achieve greater comfort and portability. For instance, Perez Reynoso et al. designed an analog processing stage supplemented with a digital filtering module^[4]. The designed acquisition system was embedded in portable glasses with a PCB board, ensuring user safety and comfort. Milanizadeh et al. employed the ADS1299 biopotential amplifier to record EOG signals, utilizing upgraded safety glasses as the electrode support for easy fixation on users' faces^[23]. Ianez et al. integrated all electronic and mechanical components into a pair of EOG glasses, aiming for an aesthetically appealing design^[27]. Barea et al. used a commercially available model (Vuzix Wrap 230, Vuzix, Rochester, NY, USA) of glasses with head-mounted displays capable of fixing dry electrodes for EOG signal acquisition and serving as a medium for projecting user interfaces^[28]. Heo et al. designed a wearable EOG system based on a headband, placing Ag/AgCl electrodes inside the headband^[32]. Lin et al. proposed a wearable HCI-based system that can detect ten types of eye movements^[34]. The system integrates the EOG signal acquisition device and the foam type acquisition dry electrodes into the glasses.

To assess hardware performance, Usakli et al. used Common-Mode Rejection Ratio (CMRR) during testing to measure the circuit's suppression capability of common-mode signals^[38]. CMRR is a parameter used to evaluate the ability of a circuit or signal processing system to suppress common-mode signals. The calculation of CMRR is typically represented in decibels (dB) and is defined as follows:

Where Vc represents the amplitude of the differential mode signal, and Vcm represents the amplitude of the common-mode signal. A higher value of CMRR indicates a stronger noise suppression capability of the system against common-mode signals.

In eye movement signal acquisition devices, a high CMRR helps reduce interference from power line frequencies and motion artifacts, preserving signal accuracy and stability, thus providing more reliable input for subsequent signal processing and interactive tasks.

Another commonly used hardware evaluation metric is the Signal-to-Noise Ratio (SNR), which is used to measure the relative strength or clarity between the measured signal and the noise in the system. It reflects the accuracy and reliability of the signal obtained by the EOG acquisition system. The calculation of SNR is typically represented in decibels (dB) and is defined as follows:

Where Psignal and Pnoise are the power of the signal and noise, respectively. Signal and noise power are calculated as the square of the amplitudes in the frequency domain after performing Fourier transform. According to Pleshkov et al., based on the research by Lopez et al.^[42], the effective frequency range of EOG signals is within [0, 30] Hz, while the bandwidth occupied by noise signals falls within [30, 122.5]^[43]. In addition to improving hardware design and implementing appropriate filters, Milanizadeh et al. proposed that the SNR can be improved by subtracting the signals of electrode pairs (up/down and left/right) from each other in a standard electrode arrangement^[23].

Self-designed devices typically consist of front-end amplifiers, hardware filters, second-stage amplifiers, analog-to-digital converters, and microprocessors, complemented by Bluetooth and battery modules. The system configuration and signal quality requirements determine the different settings for amplifier gain, filter cutoff frequencies, and sampling frequency.

In the context of interactive design, factors such as user comfort, safety, and device cost are crucial considerations for researchers. Some studies have used dry or semi-dry electrodes^[26] to improve comfort. Furthermore, wet electrodes, due to the use of conductive gel, initially provide signal quality superior to that obtained by dry electrodes within a short time. However, wet electrodes also face the issue of signal quality degradation as the conductive gel gradually dries out. In contrast, dry electrodes, which do not require the use of conductive gel, yield more suitable for long-term signal acquisition scenarios. Additionally, wearable designs consider human ergonomics^[44], offering users greater freedom, comfort, and safety. System safety is also a priority, adhering to International Electrotechnical Commission standards^[45] and incorporating current-limiting measures and protective components to ensure user and device safety.

Preprocessing

Before achieving an efficient and reliable EOG-HCI system, preprocessing steps are crucial. The main goals of preprocessing are to remove noise and baseline drift from the EOG signals, enabling more accurate and efficient feature extraction and interaction command recognition.

The EOG signal bandwidth is 0-50 Hz. Within this frequency range, EOG signals are often affected by various forms of interference, such as 50/60Hz power line noise and muscle activity^[46]. Additionally, baseline drift is another significant factor that affects the quality of EOG signals. Baseline drift, unrelated to eye movements, can cause large amplitude offsets in the signals, leading to the loss of characteristic features or waveforms associated with normal eye movements^[47].

During the hardware acquisition stage, hardware-based filtering is commonly employed for signal preprocessing. However, further preprocessing on the computer is still necessary to reduce noise and eliminate baseline drift, preparing the signals for subsequent signal processing tasks^[48].

To address noise and baseline issues in EOG signals, the most direct approach is to use bandpass filters for signal filtering. In the literature review, A total of 14 studies directly applied bandpass filters to the signals, with lower cutoff frequencies ranging from 0.1 to 0.5 Hz and upper cutoff frequencies ranging from 10 to 62.5 Hz. After passing through the bandpass filter, the signals were regarded as the signal with noise and baseline removed. To further ensure the removal of 50/60Hz power line noise, additional 50/60 Hz notch filters were used on top of the bandpass filters in studies^[12,15]. Additionally, moving average filters and median filters were commonly used for further signal smoothing^{[16,20,24,27,34]}.

Given the non-stationary nature of EOG signals, wavelet transform has also become a preferred choice for many researchers. López et al. used thresholding to select appropriate thresholds for wavelet transform coefficients to remove or attenuate low-weight spectral components^[35]. They tested different mother wavelets and thresholding rules and found that using bior5.5 (biorthogonal) wavelet with the maximum-minimum thresholding method yielded the best signal-to-noise ratio. Choudhari et al. used the least squares method to fit and remove baseline effects, followed by denoising the signals using the db-10 mother wavelet and a fourth-order Butterworth filter with a frequency range of 1 to 10 Hz^[7]. In^[30], Perez-Reynoso et al. proposed a digital notch filter using Fast Fourier Transform(FFT) to eliminate unknown noise frequencies, achieving an undisturbed noise-free EOG signal by analyzing the EOG spectrum in real-time and selecting cutoff frequencies.

Ryu et al. introduced a novel method for removing baseline drift and noise from differential EOG signals, termed the Differential of the Sine-Based Fitting Curve (DOSbFC) method^[13]. This method involved calculating the DOSbFC signal by computing the voltage values in increments of 1° of eye gaze direction and the time required to stop eye movement after eye movement completion, using a formula to select and average the most similar potentials within a sliding window. Lee et al. used a linear function to fit and remove baseline drift and combined the amplitude-increasing waveform components in the horizontal and vertical channels with the EOG's maximum amplitude threshold to eliminate muscle artifacts^[24].

In addition to noise and baseline, blink signals are also considered artifacts in some interactive tasks and need to be removed during preprocessing. Chang et al. detected blinks based on a blink threshold and removed corresponding values^[36]. They also calculated a mean value for baseline every 100 ms, using linear interpolation with adjacent signals to fill in the missing data due to blink removal and performed linear regression to eliminate low-frequency drift.

Effective and rational preprocessing methods will significantly improve the performance and user experience of eye-movement-based interaction systems, promoting their widespread application in practical scenarios.

Features and algorithms

In HCI research, algorithms serve the interaction tasks, which often require real-time performance. Before applying algorithms, feature extraction is usually conducted, involving the extraction and selection of appropriate features, to enhance the performance and effectiveness of the algorithms. The quality of feature extraction often determines the final performance of the algorithms.

We have ranked the top four algorithms used in the 35 articles and conducted statistics on the features employed in the literature using these algorithms. The results are presented in Table 3.

In the field of EOG signal analysis, feature extraction is a critical step, and common methods can be categorized into two types: features extracted from preprocessed clean signals and features in the form of first-order differentials. The original EOG signals typically exhibit distinct temporal patterns, directly reflecting waveform variations, which make them perform excellently in fast real-time processing. However, some researchers believe that certain features are not easily distinguishable in the original signal but become more prominent in the first-order differential forms of the original signal. Therefore, they opt to apply first-order differentiation to enhance feature discrimination.

Regarding feature extraction methods, researchers have proposed various innovative algorithms. For instance, Lv et al. introduced a spatial domain feature extraction method based on the Independent Component Analysis (ICA) algorithm^[8]. They used ICA separation matrix to construct a set of ICA spatial filters, mapping the experimental signals to different eye movement direction spaces through linear projection to extract spatial domain features. Fang et al. focused on continuous eye-writing recognition tasks and obtained extended EOG features suitable for different sampling rates through preprocessing, downsampling, and feature concatenation, used for subsequent neural network classification^[31].

Commonly used algorithms for eye movement signal classification include thresholding, waveform matching, neural networks, and traditional machine learning classifiers. The choice of algorithm largely depends on the research objectives and data characteristics. In this regard, neural networks have shown excellent performance in the field of biosignal processing, especially deep neural networks. However, traditional machine learning retains a distinct advantage in terms of interpretability that neural networks cannot replace. To illustrate the impact of research objectives and data characteristics on algorithm selection, we have separately summarized neural networks and traditional machine learning classifiers.

The thresholding method is the simplest classification approach, where a series of corresponding thresholds are set for eye movement signal classification using different feature combinations. It is worth noting that the thresholding method requires prior calibration of thresholds using collected eye movement signals to improve the accuracy when applied in formal usage. In research^[21,35], thresholds were set based on empirical rules, which may influence the accuracy for different subjects or within the same subject.

Waveform matching is primarily applied in interactive tasks involving eye-writing recognition of numbers, letters, etc.. Among them, the dynamic threshold detection method proposed by Ding et al. segments the eye movement signal into stroke segments and then uses the Dynamic Time Warping (DTW) algorithm to calculate similarity for handwriting eye movement signal classification^[9]. Kim et al. proposed a method that uses the gradient value of the subject's horizontal EOG signal waveform to determine the smooth eye movement direction^[37]. And by comparing the waveform gradient with the waveform gradient of the sound source surrounding the subject, it is a method to judge whether the subject pays attention to one of the multiple sound sources.

The Support Vector Machine (SVM) is the most commonly used classifier in machine learning. In addition to SVM, López et al. used AdaBoost for classification, which combines weak learners to form a strong classifier through linear combination^[11].

Moreover, classification methods based on neural networks have been gaining popularity. Perez Reynoso et al. used an optimized genetic algorithm to obtain a multilayer neural network model for modeling and reduced the cost of pretraining users by dynamically calibrating thresholds^[4]. Barea et al. employed the Extreme Learning Machine (ELM) algorithm to train a Radial Basis Function (RBF) network to determine if a valid saccade occurs, an algorithm that is simple, has small training errors, and rapidly responds to users' eye movements^[28].

In addition to the above-mentioned four types of algorithms, Barbara et al. proposed a novel dual-channel input linear regression model that considers features extracted from horizontal and vertical EOG signal components, mapping eye potential displacements to gaze angle displacements^[12]. Barea et al. based their EOG system model (BiDimensional Dipole Model EOG, BiDiM-EOG) on physiological and morphological data of EOG to separate saccades and slow eye movements and determine fixations^[25].

Feature extraction and algorithm design play a pivotal role in EOG-based human-computer interaction tasks. The selection of appropriate features and algorithms is critical for building efficient and reliable eye movement human-computer interaction systems. Researchers need to customize feature selection and algorithm design based on the specific task requirements and interaction effectiveness to provide high-quality user experience and interaction efficiency. For example, for simple four-directional physical control tasks, the thresholding method may be more suitable as it can quickly distinguish eye movement signals from non-eye movement signals, generating simple control commands. For complex virtual keyboard designs or alphanumeric writing tasks, waveform matching may be more applicable, as it can combine waveforms of EOG signals generated by a series of eye movements to expand the number of executable commands.

From Table 3, it can be observed that the majority of research in the EOG-HCI field still relies on algorithms related to thresholds. Most of these studies were conducted in controlled laboratory environments, and a significant portion of them employed calibration techniques during data collection to ensure data quality. This helps to reduce individual differences and minimize signal variations over time, thereby lowering the robustness requirements for recognition technology. Furthermore, the time-domain information of EOG signals is easily interpretable by human researchers. This allows for the development of algorithms that combine threshold features with human prior knowledge to tailor recognition algorithms for specific eye movement classification tasks with different requirements. The high flexibility of threshold detection methods makes them widely applicable in various EOG-HCI research scenarios.

However, threshold detection methods are sensitive to parameter changes, leading to poor robustness. Furthermore, the features derived from thresholds are often designed based on human observations, resulting in low-dimensional, concrete features that cannot capture some high-dimensional, abstract features of different eye movements that are difficult for humans to understand. Hence, threshold detection methods often encounter issues related to suboptimal precision and robustness. In fact, nearly every study utilizing threshold detection methods necessitates the implementation of calibration techniques during data acquisition. In summary, while threshold detection methods offer simplicity and versatility, they tend to lack the desired level of robustness.

Additionally, four studies utilized neural network technology to construct eye movement classification models. Study^[4] used a multilayer neural network (MNN) to model the EOG signal as a mathematical function and optimize genetic algorithms for automatic signal calibration to obtain the threshold of each subject’s EOG signal. Barea et al.^[28] employed an Extreme Learning Machine (ELM) algorithm to train a radial basis function (RBF) network with an internal structure comprising 20 neural neurons to determine whether the user had performed a valid eye jump. Pérez-Reynoso et al.^[30] used a multilayer perceptron (MLP) to establish a neural network with four hidden nodes for eye movement classification. Fang et al.^[31] employed deep learning techniques (DL) to construct a Deep Neural Network-Hidden Markov Model (DNN-HMM). This network model comprises four hidden layers, with 132 units in the input layer, 200 units in the first hidden layer, 100 units in the second hidden layer, and the number of units in the output layer equal to the number of HMM states. The ultimate output of the network represents the probability distribution of eye movement classifications.

It can be observed that DL techniques have not received sufficient attention in the EOG-HCI field. On one hand, based on the studies surveyed in this review, researchers in this field have placed more emphasis on application and have not explored EOG recognition algorithms extensively. On the other hand, there is currently no openly available EOG dataset tailored for HCI, making it challenging to train high-performance DL networks based solely on the samples collected from various studies. Chang et al.^[49] used a database of ten Arabic digit EOG eye-writing tasks collected in their previous work^[35], consisting of 540 eye-written digit samples for training their model.

In summary, current research is more inclined to explore the feasibility of using EOG for HCI, and threshold-based methods, due to their high generality, have become a popular recognition technology for exploring EOG-HCI feasibility. However, once the feasibility of EOG-HCI is established, higher demands will be placed on recognition technology in terms of robustness, classification precision, the number of classes, etc. At that point, recognition technology developed based on deep learning, for example, will be developed to meet these requirements.

Classification of research results in EOG-HCI

For clarity, we define four concepts within the scope of this review:

1. Interaction Application: A program or system with interactive functionality and a complete interactive loop, designed to meet user needs.

2. Interaction Task: Tasks that need to be accomplished by users via interaction to fulfill their requirements.

3. Control Command: The command translated by the interaction interface based on the user's operational intent, used to complete the interaction task.

4. Interaction Modality: The mapping between trigger (specifically referring to eye movement in this review study) emitted by users’ interaction intention and control commands. In the field of EOG-HCI (Electrooculography-based Human-Computer Interaction), the interaction application represents the final presentation form of research. Users achieve their goals of using the interaction application by completing interaction tasks, where the type of interaction task determines the interaction modality. The interaction modality significantly influences the engineering solutions for EOG signal acquisition, processing, and algorithm design. Figure 5 classifies all interaction applications over the 21 years based on their interaction tasks, dividing them into eye movement control and eye movement information input categories, and further subdividing them according to the interaction modality. As some literature describes multiple types of interaction applications, this review only selects more typical or distinctive applications for statistical analysis. This summarization provides a more comprehensive understanding of the entire EOG-HCI domain.

Figure 5. Classification of EOG-HCI Publications (2002-2022).

Direct control of virtual buttons or functions:

In this form of interaction, users control virtual software, desktop, or mobile applications through a GUI interface without involving real objects. For example, Lv et al. developed an electronic book reading controller using a MATLAB/JAVA joint platform, where users can achieve six basic control commands, such as page flipping, pause/resume reading, and opening/closing the book, through six predefined eye movement patterns^[8]. Similarly, Ryu et al. designed desktop and mobile applications that utilize different eye movement patterns to act as an auxiliary controller for Windows applications and perform seven functions for multimedia control on mobile devices^[13]. These applications provide users with a convenient way to control virtual software and applications rapidly through eye movements.

GUI-manipulated real or VR objects:

In this form of interaction, users typically select control commands for a target object in a GUI which serves as a control panel, before sending the control commands to the object for execution. For instance, Li et al. designed a GUI interface that included wheelchair switch converter buttons^[5]. Users activate the EOG switch by blinking synchronously with the flashing, thereby sending control commands to navigate the powered wheelchair.

Direct control of real objects:

In this form of interaction, researchers usually directly send user-generated eye movement commands to a controller, which implements high-level control and generates specific operation commands for interactive objects, primarily wheelchairs and robotic arms. In Barea et al.'s research^[25], recorded eye gaze information is processed by an onboard computer to formulate the wheelchair's control strategy and generate linear and angular velocity commands. These commands are then transformed into angular velocities for each wheel based on the wheelchair's kinematic model and sent to the low-level control module, achieving two closed-loop velocity controls. Consequently, users can navigate the wheelchair's movement and steering through eye movement commands, enabling autonomous navigation.

On the other hand, Ubeda et al.'s research^[26] focused on designing a robotic arm application. This application converts users’ eye gaze information into control commands for the robotic arm's end effector, replacing human arms to operate various tools. This interaction modality allows users to perform complex robotic arm operations through simple eye movement commands, improving their self-care abilities and work efficiency.

Virtual keyboard form:

In this form of eye movement interaction, virtual keyboards with different character layouts are used, allowing users to select characters they need through eye movements to complete input. To simplify the number of eye movements required for character selection, researchers explore different stimulation modes. One mode involves selecting a button by blinking users’ eyes in synchrony with the flash of the button^[17]. Another mode is the row/column paradigm activation mode^[29], where characters are arranged in rows and columns. Users efficiently determine the character's position by selecting the row and column through eye blinking. This method streamlines the user's selection process, making eye movement commands more efficient.

Eye-writing form:

In information input-type eye movement interaction applications, the eye-writing form aims to use different eye movement patterns to write symbols, such as characters. Researchers combine different eye movement types into different eye movement patterns, with each pattern corresponding to a stroke. By calculating the similarity between the user's input eye movement pattern and predefined templates, the system outputs the character the user has written. For example, Ding et al. designed a stroke-based Chinese eye-writing system that maps five basic strokes to predefined eye movement patterns, enabling continuous writing of Chinese characters^[9]. Similarly, Lee et al. predefined 12 eye movement templates for outputting 26 lowercase letters and 3 control commands^[24]. These systems determine the most similar character as the output result by comparing the differences between the input eye movement trajectory and the character templates. This eye-writing form of interaction system provides a novel and efficient way of inputting information.

Question-and-answer system:

Kim et al. developed a novel EOG-based Augmentative and Alternative Communication (AAC) system aimed at helping users engage in binary yes/no communication^[37]. The interaction system plays two rotating sound stimuli to the user and asks them to focus on the specified sound source. During this process, horizontal EOG is recorded in response to the involuntary auditory ocular reflex, which answers the designed questions. This system provides a means for users with communication barriers to engage in simple communication through eye movements, bringing convenience and improvement to their lives.

Summary

We classified the interaction applications into six subcategories within two main categories based on the categories presented in the previous section and proposed two metrics to compare the related research of the two interaction application types:

● Degree of Interaction Freedom (DOF): Reflecting the availability of the application, i.e., the number of available predefined commands. This metric is based on the number of available commands.

● Interaction Efficiency (EEF): Reflecting the usability of the application in terms of how quickly users can complete commands. This metric is based on the maximum number of eye movements required to complete a command.

There is a correlation between DOF (Degree of Freedom) and EEF (Efficiency). In essence, as DOF increases for the same application, the efficiency of interaction tends to decrease. This correlation is referred to as the Interaction Benefit score, represented by C. The formula for calculating C is as follows:

Using log₂DOF instead of DOF in the calculation helps mitigate the sensitivity of C to changes in DOF, considering that increasing DOF is relatively easier. All the information is summarized in Table 5.

Research on control-type interaction applications primarily focuses on user-directed interaction tasks and designs corresponding interaction modalities based on the purpose of using the interaction application. This type of interaction modality covers a wide range of application scenarios, including appliance control in home IoT settings^[14,29], machine control^[23,25], navigation^[4,5,26], web browsing^[16], e-book reading^[21], eye-tracking games^[22], and various other aspects. The number of research papers in this category is nearly twice as much as that of information input-type interaction applications.

Among the existing control-type interaction applications, 17 out of 23 papers used Graphical User Interface (GUI) as an intermediate medium. GUIs offer high extendibility, allowing the functionality of control-type applications to be easily expanded by adding more interfaces, placing new functions in separate sub-interfaces, or adopting layouts similar to P300 speller to integrate multiple functions on the same interface, thus enhancing the freedom and extensibility of the interaction interface. Therefore, in theory, there is no upper limit to the number of functions in control-type interaction applications.

However, it is essential to consider user experience. As the number of functions increases, users' cognitive load may also increase, leading to higher error rates and reduced operational efficiency. This aspect is particularly critical for the target users of EOG-HCI, as a significant portion of them may have mobility impairments, such as Amyotrophic lateral sclerosis (ALS) patients. Therefore, research on control-type interaction applications needs to pay special attention to a reasonable analysis of interaction tasks to ensure that users can easily understand and operate the interaction interface.

Information input-type applications have more explicit interaction task purposes, namely, user communication with the outside world. However, since human communication purposes are challenging to predict, information input-type applications require higher freedom to meet users' varied needs. From Table 5, it can be observed that among the 35 EOG-HCI literature samples, the top five interaction applications designed in literature^{[6,12,16,35,38]} have the highest DOF. Among them, applications in literature^[16] is a comprehensive human-computer interaction application that includes both pages browsing through eye movements and information input through a virtual keyboard. Although categorized as a control-type application, it overall encompasses the functionalities of an information input-type application. The interactive applications mentioned in all these five studies have a virtual keyboard, which indicates a high requirement of freedom for information input type applications.

In research of information input-type applications, ensuring efficiency while maintaining degrees of freedom is essential. Therefore, investigations into information input-type interaction applications may encounter some challenges, but they have the potential to make significant contributions to the field.

It is worth noting that the evaluation of interaction design should be conducted within a unified context. This review covers a wide variety of application types, and each interaction design is tailored to different users and scenarios. Therefore, the evaluation metric C proposed in Table 5 can only serve as a reference point. Performance assessment needs to be carried out specifically for the selected user groups and within the context of each specific application type.