Mild Cognitive Impairment Identification System Based on Physiological Characteristics and Interactive Games ^†

Abstract

As the global aging population increases, the early detection and prevention of Alzheimer’s disease (AD) have become important in public health. To solve the problems of subjectivity and low timeliness of traditional assessment methods, this paper proposes a multimodal dementia prevention system that combines physiological sensing, a gamification interface, and a classification model. The system includes an interactive joystick to measure pulse and blood pressure. A Chinese music game app increases the participation of the elderly and reduces their sense of rejection through gamification interaction. After the physiological data were standardized by Z-score, they were input into three small sample classifiers (Gaussian Naïve Bayes, Fisher Linear Discriminant Analysis, and Logistic Regression) for the binary classification of AD. The system performance was evaluated using the Leave-One-Out cross-validation method. Experimental results show that Logistic Regression performed best in situations with extremely small samples and class imbalance, with an F1-score of 0.700, which was higher than the other two. Dynamic features and model fusion technologies need to be integrated to further enhance the clinical application potential of the system in the early prediction of dementia.

1. Introduction

As the global population ages at an accelerated rate, the prevalence of dementia continues to increase, posing a major challenge to public health worldwide. Current estimates indicate that more than 55 million people are affected by dementia, reaching 135 million by 2050 [1]. This growing burden places significant pressure on healthcare systems, long-term care services, and the broader social economy. Alzheimer’s disease (AD) is the most common form of dementia, characterized by a slow and irreversible progression. Mild cognitive impairment (MCI) often characterizes the preclinical stage. Patients with MCI may experience single or multiple cognitive impairments such as memory loss, language disorders, and decreased attention and executive function. However, most of them can still maintain daily functions [2]. Studies suggest that approximately half of the individuals with MCI will develop dementia within five years [2].

The early symptoms of MCI resemble typical signs of aging, including short-term memory loss, mood fluctuations, and decreased motivation. As a result, these symptoms are frequently overlooked or dismissed, leading to delays in medical consultation and missed opportunities for timely intervention. Enhancing awareness among both the general public and clinical practitioners is therefore critical to slowing disease progression [3]. Common diagnostic tools for AD include the Mini-Mental State Examination and the Montreal Cognitive Assessment [4]. However, these methods heavily rely on manual observation and subjective evaluation, limiting efficiency and hindering the provision of immediate, personalized feedback. To address these limitations, AI technologies have been applied to dementia detection and management, including wearable devices [5] and machine learning models [6]. In addition, physiological biomarkers such as heart rate [7], blood pressure, and skin conductance are being explored as potential auxiliary diagnostic tools.

Since no cure for AD currently exists, prevention and delay of disease progression must be prioritized in medical and public health research. AD is a multifactorial neurodegenerative disorder, and prevention strategies often emphasize lifestyle and behavioral interventions. Cognitive training is recognized as an effective non-pharmacological approach, with regular engagement in activities targeting memory, language, and executive function shown to slow cognitive decline and stimulate brain activity [8]. Social interaction also plays a protective role, as maintaining positive relationships and participating in group activities can promote emotional stability and cognitive stimulation [9].

Recently, gamified interventions have emerged as promising tools for cognitive prevention [10]. Cognitive assessment games, which integrate interactive tasks with neuroscience-based design, simulate real-life scenarios and enhance motivation for participation. These games have demonstrated strong correlations with clinical scales in measuring reaction speed, memory, and attention [3]. Similarly, music-based activities, such as playing instruments, singing in choirs, or rhythm training, have gained recognition as effective non-drug strategies. By engaging multiple sensory modalities, motor coordination, and emotional involvement, music interventions have been shown to reduce dementia risk and support cognitive health in older adults [11]. Effective prevention of AD requires multifaceted approaches that combine cognitive enhancement, social participation, and digital tools to establish sustainable, participatory intervention models that improve quality of life in aging populations.

Based on the previous study results, this paper proposes an early prevention system that integrates game-based interaction with physiological sensing to address MCI, a precursor to AD. The system collects pulse and blood pressure data through a joystick and music rhythm game application equipped with photoplethysmography (PPG) sensing. MCI risk is then assessed using three small-sample classification models: Gaussian Naïve Bayes, Fisher Linear Discriminant Analysis, and Logistic Regression.

2. Related Work

2.1. Biomarker Sensing Technology

Biomarker technology and Alzheimer’s disease have a close connection. Biomarkers refer to indicators that are measured and reflect physiological and pathological processes and treatment responses. In the research and diagnosis of dementia, the application of biomarker technology is gradually becoming an essential tool for early detection, classification, and prognosis assessment. In Ref. [12], blood pressure measurement was proposed as a method to assess dementia risk. The analysis suggests that in certain cases, higher systolic and diastolic blood pressure may be associated with a reduced risk of AD, indicating a potentially protective effect. Incorporating blood pressure records with family medical history could facilitate early risk prediction and serve as an effective tool when combined with wearable devices or artificial intelligence technologies.

In Ref. [13], a non-invasive biomarker approach for Alzheimer’s disease was proposed based on physiological indicators such as skin pH, moisture retention, elasticity, and capillary tortuosity. Patients with autosomal dominant AD typically exhibit more neutral skin pH and greater moisturization but reduced elasticity. Moreover, a higher degree of capillary distortion correlates with poorer cognitive function. Notably, among carriers of the ApoE ε4 gene, pronounced capillary distortion is associated with improved treatment response, suggesting that skin-based assessments may support early diagnosis and prediction of therapeutic efficacy. A three-stage microfluidic capture chip was developed to simultaneously detect multiple Alzheimer’s disease biomarkers (e.g., lactoferrin and Aβ-42) in saliva [14]. This system integrates immunomagnetic beads of varying sizes with specific antibodies, distributed automatically within a micro-column array featuring progressively narrower gaps. Such a design enables multi-label detection from a single sample injection. The chip demonstrates high sensitivity, requires minimal sample volume, allows rapid processing, and shows strong potential for clinical application.

2.2. MCI Forecast and Analysis

MCI is a precursor to Alzheimer’s disease, so close monitoring and early detection can effectively improve prevention and treatment outcomes. Wearable devices are used to measure heart rate variability (HRV) to effectively distinguish MCI patients from healthy individuals [7]. In the method, a multiple linear regression model is applied, with HRV indicators (e.g., standard deviation of NN intervals, root mean square of successive differences, high-frequency power) as dependent variables and cognitive status as the independent variable, while controlling for age and gender. This method is noninvasive, immediate, and highly accurate, making it a promising tool for early detection and continuous monitoring of MCI. In Ref. [15], dementia and MCI detection were explored through facial expression features extracted from natural conversation videos. By analyzing facial movements, emotion classification, emotion dimensions (Valence–Arousal), and face embeddings combined with machine learning models, the method demonstrated high recognition accuracy, highlighting its potential as an early screening tool.

In Ref. [16], a noninvasive home-based screening system was developed using sensors placed under the mattress to record body movements and breathing patterns during sleep. The time delay between these signals was identified as a key feature for determining the presence of MCI. When neural networks and statistical models were applied for classification, the system achieved an accuracy of up to 88%, underscoring its potential as a practical tool for early detection.

3. Game Interactive System

To reduce the elders’ aversion to diagnostic screening, this article used a gamification approach to package physiological signal testing and AD risk assessment. The system consists of an app in the Chinese language music game and an interactive joystick, as shown in Figure 1.

3.1. System Architecture

We designed a joystick that can perform interactive operations. As shown in Figure 2, this interactive device in-cludes an Arduino ESP32 development board, a DFRobot PPG sensor, an Arduino joystick module, and a TP4056 charging module, as shown in Figure 2.

The joystick is connected to the mobile phone APP via Bluetooth. During the game, the joystick is used to control the app screen and game operations. ESP32 transmits the data to the APP for integration. After the game is over, the blood pressure data are uploaded to the cloud database for calculation and dementia risk analysis, and the results are sent back to the app for display to the user. The joystick’s appearance was designed based on the prototype design of the somatosensory game joysticks on the market, allowing users to hold the handle comfortably. When holding the joystick, the thumb controls the upper joystick button, while the other four fingers grasp the entire joystick, as shown in Figure 1c. At the same time, the index finger is placed just above the PPG sensor.

3.2. Music Interactive Game App Design

The music rhythm game app is shown in Figure 3. On the homepage of the APP, users log into the game based on their identity as a player or administrator. The functions become different according to different bodies. The manager’s function is to effectively assist family members and caregivers in quickly understanding the status of the person being cared for.

Figure 4 shows the game flow of the APP, including the game process, and the results display after the game ends. Figure 4a shows the settings and process before the game starts. Users start the game after selecting the difficulty and music of the game. The whole game process takes about 1–2 min. Users use the joystick to explode the balloons on the screen to score points. During the process, the joystick is used to collect blood pressure data and upload it to the cloud database. When the game is over, the score, evaluation, and suggestions for MCI are shown in Figure 4b.

In addition, this system has the option of setting historical records for users to view. The data provided includes historical scores and MCI indicator analysis, as shown in Figure 5. Through historical records, users can better understand their physiological data, and can upload the records to provide a reference to doctors.

Figure 6 shows the management mode screen. Administrators manage the game history records of users to confirm whether today’s result data (including users’ game scores and MCI evaluations) have been uploaded.

In management mode, two functions are available: user editing and administrator editing, as illustrated in Figure 7a. The user editing function allows for the addition and deletion of players, as well as the modification of their basic information, as shown in Figure 7b. The administrator editing function enables the addition and removal of administrators and the adjustment of their permissions, as depicted in Figure 7c.

4. Dementia Prevention Prediction Model

The system extracts data on pulse, blood pressure, and gender from the original data and converts it into a seven-dimensional numerical feature vector. The blood pressure was analyzed and separated into systolic and diastolic pressures. To eliminate the impact of measurement differences between different features, all features were normalized using the Z-score. In the subsequent model training stage, seven-dimensional standardized feature vectors were used as input, and three classic small sample classifiers: Gaussian Naïve Bayes (GNB), Fisher Linear Discriminant Analysis (LDA), and Logistic Regression, were used to construct a binary classification model for dementia.

GNB is a supervised classification method whose core assumption is class-conditional independence, that is, given the class Y = k, the prediction variables are independent of each other. According to the Bayes theorem, when the values of the predictor variables X₁, …, X_p are known, the posterior probability of belonging to class k can be expressed as Equation (1) [17].

$$\widehat{P}\left(Y=k∣X_1,\dots ,X_P\right)={\displaystyle \frac{\pi \left(Y=k\right){\prod }_{j=1}^{P}P\left(X_j∣Y=k\right)}{{\sum }_{k=1}^K\pi \left(Y=k\right){\prod }_{j=1}^{P}P\left(X_j∣Y=k\right)}}$$

(1)

LDA also uses a seven-dimensional feature vector as input to improve the separability between categories by maximizing the Fisher discriminant function. This method maximizes the mean difference between two types of samples while minimizing the variation within each type using Equation (2) [18]. After projecting to one-dimensional space, the projection results are assumed to follow a Gaussian distribution to calculate the posterior probability and classify the results using 0.5 as a threshold. Given the limited number of samples, when the intra-class scatter matrix S is not fully ranked, the system uses the Moore–Penrose pseudo-inverse to solve the parameters.

$$\underset{\varphi }{\mathit{max}}\mathcal{J}\left(\varphi \right)\hspace{0.17em}=\underset{\varphi }{\mathit{max}}{\displaystyle \frac{{\varphi }^T{S}_{AB}\varphi }{{\varphi }^{T}S\varphi }}\hspace{0.17em}$$

(2)

The system adopts binary logistic regression using L2 regularization. The mathematical forms of the logarithmic odds, the prediction probability, and the logistic function are presented in Equations (3)–(5) [19]. The model parameters α and β are estimated through Maximum Likelihood Estimation, and L2 regularization is introduced to reduce the risk of overfitting. The regularization strength is set to C = 1. To address the problem of sample imbalance, the loss function is reweighted using balanced weights. The optimal solution of the model is calculated by the liblinear solver, which provides stable convergence characteristics in a small sample environment.

$$logit{\theta }_i\equiv \mathit{log}\left\{{\displaystyle \frac{{\theta }_i}{1-{\theta }_i}}\right\}=\alpha +\beta x_i$$

(3)

$${\theta }_i\equiv pr\left(Y_i=1\right)={\displaystyle \frac{e^{\alpha +\beta x_i}}{1+e^{\alpha +\beta x_i}}}$$

(4)

$$1-{\theta }_i\equiv pr\left(Y_i=0\right)={\displaystyle \frac{1}{1+e^{\alpha +\beta x_i}}}$$

(5)

This paper used a self-constructed dataset for experimental analysis. The dataset included physiological information on pulse, diastolic blood pressure, and systolic blood pressure from 10 patients with MCI and 11 healthy subjects, comprising 12 males and 9 females. All data were collected and analyzed with the participants’ knowledge and consent. To expand the sample size and improve measurement accuracy, each participant recorded two sets of physiological data (pulse and blood pressure) while engaging with the game.

Model performance was evaluated using accuracy, precision, recall, and F1-score metrics, calculated through Leave-One-Out cross-validation (LOOCV) [20]. LOOCV is a validation method in which one sample is reserved for testing, while the remaining samples are used for training. This method maximizes sample utilization efficiency and is particularly well-suited for small datasets. The method was used to quantify and compare the predictive performance of three models, thereby assessing their applicability under extremely limited sample conditions.

Experimental Results

Table 1. Model performance comparison.

	Accuracy	Precision	Recall	F1-Score
GNB	0.524	0.500	0.400	0.444
LDA	0.714	0.750	0.600	0.667
Logistic Regression	0.714	0.700	0.700	0.700

presents the evaluation outcomes of the three classifiers across 21 subjects (42 measurements). GNB showed an accuracy of 0.524 and an F1-score of 0.444, indicating that its assumption of class conditional independence does not hold for the seven-dimensional physiological features. In contrast, LDA and Logistic Regression presented an accuracy of 0.714, with precision and recall values of 0.700. Logistic Regression improved the F1-score to 0.700, outperforming LDA’s 0.667. These findings suggest that in scenarios involving extremely small and class-imbalanced datasets, Logistic Regression with weight compensation was the most balanced and stable predictive method, whereas GNB showed the worst performance due to its reliance on independence assumptions and limited sample support.

5. Conclusions

This paper developed an AD precursor prevention system that integrates a gamification interface, physiological sensing, and classification models. The system enhances user participation through interactive joysticks and music games and collects physiological data such as blood pressure and pulse in real-time. Experimental results show that the Logistic Regression model performs most stably under conditions of extremely small samples and class imbalance and is practical and scalable. The system needs to be optimized in the future, including expanding the number of samples, integrating more dynamic features, and introducing model fusion and uncertainty quantification to improve prediction accuracy and enhance its clinical application value in elderly care and early cognitive abnormality screening.