There are many studies on disease risk prediction based on artificial intelligence technology. Aditya Khosla et al. proposed an integrated machine learning method for stroke prediction [4]. This method is an automatic feature selection algorithm based on support vector machine (SVM). Jiayu Zhou et al. proposed a multi-task learning formula for predicting disease progression using the cognitive subscale (ADAS-Cog) score in the Alzheimer’s disease (AD) assessment scale [5]. This formulation treats the prediction of each stage time point as a task and formulates the prediction of disease progression as a multi-task regression problem. Ankit Agrawal et al. proposed a lung cancer risk prediction model based on support vector machines, artificial neural networks, and random forest technology using lung cancer imaging data from the National Cancer Institute of the United States [6]. Based on this model, they developed an online lung cancer risk prediction system. Alceu Ferraz Costa et al. proposed a feature extraction method for classifying interstitial lung disease in computed tomography (CT) scans based on support vector machines. This method achieved an accuracy rate of 84.36% when performing classification tasks [7].

D. Shiloah Elizabeth et al. proposed an automated method for segmenting lung tissue from chest CT images based on artificial neural networks [8]. This method can be used for preprocessing before the diagnosis of lung diseases to improve the performance of system diagnosis. Md Jamiul Jahid et al. proposed a cancer disease prediction model that can be used in clinical practice using a support vector machine (SVM) as a classifier [9]. Shuo Xiang et al. proposed a multi-source data “two-layer” learning model based on random forests to solve the problem of the loss of block-level data leading to the decline in the prediction accuracy of Alzheimer’s disease (AD) [10]. Based on the unified formula, this model handles feature-level and source-level analysis, imputing missing elements.

Roland Assam et al., based on the conditional random field (CRF), used the extracted sample feature vector to capture the latent features of the freeze of gait (FOG) time series of Parkinson’s disease patients [11], analyzed the motion time series data of Parkinson’s disease patients, and analyzed the patient’s freezing of gate state (FOG) for effective prediction. Matthew Seeley et al. used a structured method of integrated learning to compare the accuracy of multiple model combinations for predicting Alzheimer’s disease (AD) [12]. Finally, they obtained the characteristic attributes that affect the prediction results. Nida Khateeb et al. used the K-nearest neighbor classifier [13], and by using 14 attributes, the accuracy of heart disease prediction reached nearly 80%. In order to reduce the false alarm rate of pulmonary nodule monitoring, Jiaxing Tan et al. proposed a two-stage deep learning framework based on a deep neural network and a convolutional neural network [14].

Allison M. Rossetto et al. proposed an integrated method based on a convolutional neural network to improve the accuracy of automatic primary diagnosis of lung cancer using deep learning on lung CT [15]. This ensemble method consists of two separate convolutional neural networks. This ensemble method achieved an average accuracy of 85.91%. Based on a convolutional neural network [16], Joongwon Kim et al. designed a system that can automatically diagnose the risk of lung cancer on chest CT. In order to realize the automatic classification of cystic fibrosis lung disease (CFLD) lesion degree in computed tomography (CT) [17], Xi Jiang et al. proposed a framework based on deep convolutional neural networks and transfer learning. In order to reduce the labeling work in the supervised lung CT image segmentation training [18], Yuan Huang et al. proposed a lung plaque feature extraction method based on a fully convolutional neural network and a generative adversarial network. In order to ensure the high accuracy and performance of fatty liver disease (FLD) prediction [19], Ming Chen et al. proposed a multi-layer random forest (MLRF) model using a medical examination dataset with standardized indicators. This model consists of an input data layer, a processing layer, and an output data layer. Among them, the processing layer comprises multiple random forests (RF).

Hatim Guermah et al. used a dataset containing physiological indicators (such as cell count, red blood cell count, and arterial blood pressure) and dietary attributes to predict the risk of chronic kidney disease based on context ontology and linear SVM [20]. They achieved an accuracy rate of 93.3%. Xuan Chen et al. proposed a weighted loss function to reduce the impact of class imbalance on the prediction results when using pancreatic magnetic resonance images (MRI) for pancreatic cancer risk prediction [21]. Furthermore, based on the ResNet18 model, a classification experiment was carried out, and the experimental results reached an accuracy rate of 91%. Lena Ara et al. used machine learning algorithms to predict peripheral arterial disease [22]. The findings also reduced variability in readouts in vascular laboratories. Amanda H. Gonsalves et al. used naive Bayesian (NB), support vector machine (SVM) [23], and decision tree (DT) to compare the risk prediction of coronary heart disease (CHD). The prediction effect of the model is relatively good.

Anik Saha et al. used a dataset containing physiological indicators (such as blood pressure and albumin) to predict chronic kidney disease (CKD) based on random forest, naive Bayesian [24], and multi-layer perceptron. They obtained an accuracy rate of 97.34%. Md. Golam Sarowar et al. used the “tuberous sclerosis” disease dataset obtained from the National Center for Biotechnology Information (NCBI) based on a hybrid of convolutional neural network (CNN) and particle swarm optimization (PSO) [25]. An optimized CNN algorithm to predict the disease of tuberous sclerosis (TSC) achieved an accuracy rate of 83.47%. Iftikhar Ahmed et al. used the “Myocardial Infarction (MI)” dataset in the UCI machine learning library based on support vector machines [26], multi-layer perceptrons, random forests, additive regression, and ant colony optimization (ACO). Convolutional neural networks (CNN) combine the CNN-ACO algorithm to predict myocardial infarction. The CNN-ACO algorithm achieved an accuracy rate of 95.78%.

When using multi-parameter magnetic resonance imaging (MRI) for prostate cancer (PCa) risk prediction [27], Paulo Lapa et al. found through comparative experiments that the accuracy of classification using semantic learning machine (SLM) is higher than that of CNN XmasNet. Muhammad Mubashir et al. proposed a new method based on a deep convolutional neural network to solve the problem of insufficient classification accuracy due to feature selection and signal analysis using wrist pulses to diagnose lung cancer [28]. This method consists of a 1D fifteen-layer deep convolutional network. This method can identify lung cancer based on the obtained wrist pulse signal and achieved a recognition accuracy of 97.67%. In order to better diagnose lung cancer [29], Yanhao Tan et al. proposed a method for automatically segmenting pulmonary nodules in CT images based on convolutional neural networks. N. Nemati et al. proposed a lightweight classification model based on a deep convolutional neural network and using EEG signals obtained from the CHEG-MIT scalp EEG database for seizure prediction [30], with an accuracy of 99%. Through comparative experiments [31], Rekka Mastouri et al. proved that in CT medical imaging analysis, the detection performance of the fine-tuned model based on the VGG16 model is higher than that of the model formed by transfer learning based on the pretrained VGG16 model.

Chulwoong Choi et al. propose an indexing method based on convolutional neural networks to automatically retrieve lung images from many DICOM medical images generated after PET-CT imaging [32]. This method achieved 70% accuracy. In order to improve the accuracy of detecting COVID-19 through chest X-ray images [33], Jonathan David Freire et al. proposed a new evaluation method based on the Resnet-34 architecture. This method uses data enhancement techniques for image preprocessing, including global histogram equalization and pink mapping. This method achieved an accuracy rate of 97.81%. Based on the Resnet-50 architecture [34], Xiang Yu et al. proposed a model for classifying breast abnormalities using mammographic imaging. In order to improve the accuracy of this model in predicting breast cancer, histogram equalization was used in the data preprocessing stage. Experimental results show that the model achieved an overall accuracy of 95.74%. Yu Lu et al. proposed a lung cancer risk prediction model based on the VGG-16 model and the expanded convolutional pulmonary nodule segmentation network [35]. This model achieved an accuracy rate of 97.1%.

S. I. Lopes et al. used radar-based or infrared thermal imaging technology to perform non-contact monitoring of body temperature and heart rate in nursing homes [36], without connecting physical electrodes, for early detection and prediction of COVID-19 in patients. R. Tsuzuki and others developed an online health chart system based on visualization [37]. This system allows nursing home nurses, medical staff, and older adult family members to view the health trends of the older adult. B. Braga et al. designed a low-cost infrared thermal imaging system for nursing homes for body temperature screening [38]. D. -R. Lu and others have developed an intelligent medical system for nursing homes, which uses the monitoring of service robots to record the current status of the older adult [39]. When the older adult falls, he presses the emergency button on his body to activate the alarm system of the service robot, and the nearby medical staff will receive an emergency call for help.

As shown in Table 1, in the current research on related topics, the accuracy rate of risk prediction for heart disease, lung cancer, chronic kidney disease, pancreatic cancer, tuberous sclerosis, myocardial infarction, epilepsy, and breast cancer has reached 80%. We produced statistics on some studies on the risk prediction of Parkinson’s disease, and the statistical results are shown in Table 2.

As shown in Table 2, the accuracy rates of models constructed using decision trees, convolutional neural networks, deep learning, and recurrent neural networks exceed 80%. Using the dataset collected by non-medical equipment, the prediction accuracy of the Diplin model proposed by using transfer learning and a lightweight neural network reached 98%.

In the abovementioned previous studies on disease risk prediction, there are many analyses of medical imaging images. However, in the application scenario of a nursing home, the nursing home does not have medical imaging equipment, and it is difficult to ensure the integrity of collecting gait, motion, and audio data. Therefore, we use datasets acquired by non-medical imaging devices in our proposed research work. Our disease risk prediction model based on WGAN, transfer learning, and EfficientNetV2 can run on ordinary tablet computers. In short, research based on datasets collected by professional medical testing equipment are not suitable for the business needs of nursing homes. The Diplin model does not rely on the datasets collected by professional medical testing equipment, so nursing homes are advised to use the Diplin model.