Fidan: a predictive service demand model for assisting nursing home health-care robots

Abstract

While population aging has sharply increased the demand for nursing staff, it has also increased the workload of nursing staff. Although some nursing homes use robots to perform part of the work, such robots are the type of robots that perform set tasks. The requirements in actual application scenarios often change, so robots that perform set tasks cannot effectively reduce the workload of nursing staff. In order to provide practical help to nursing staff in nursing homes, we innovatively combine the LightGBM algorithm with the machine learning interpretation framework SHAP (Shapley Additive exPlanations) and use comprehensive data analysis methods to propose a service demand prediction model Fidan (Forecast service demand model). This model analyzes and predicts the demand for elderly services in nursing homes based on relevant health management data (including physiological and sleep data), ward round data, and nursing service data collected by IoT devices. We optimise the model parameters based on Grid Search during the training process. The experimental results show that the Fidan model has an accuracy rate of 86.61% in predicting the demand for elderly services.

KEYWORDS:

• Nursing service
• big health
• internet of things
• machine learning
• smart service

1. Introduction

Population aging has become a challenge for every country in the world (Yang et al., 2022). Developed countries also face this problem (Johnson, 2022). In the context of population aging, the traditional elderly care service model is facing transformation (Zhang, 2023). In the elderly care service, the demand for caregivers is also increasing. The increased number of older adults and the decrease in newborns have caused a human resource crisis in the nursing industry (Marć et al., 2019). Our research is to solve this crisis better. Researchers are exploring solutions to the current problem of increased caregiver workload due to an aging population. In order to solve the problem of human resources and the increase in the workload of nursing staff, it has become common to use robots to assist nursing staff in their work. However, traditional robotic solutions’ limitations are that the robots currently used in nursing homes perform pre-set tasks. Robots performing pre-set tasks need to reduce the workload of nursing staff effectively.

The workflow of this type of robot that executes the set task type, taking the disinfection robot as an example, the main steps are: first, the disinfection robot will move around the office area where it is working before the first work to complete the work of building the map; then, the staff using the disinfection robot will mark critical locations on the constructed map, such as meeting rooms, administrative offices, front desk, etc.; finally, the staff will set the start time and place of disinfection in the background system, and the disinfection robot will Start working at the appointed time and position, and return to the charging position after completing the disinfection task. Since this type of robot does not have the capabilities of subjective analysis, judgment, decision-making, and operational execution, it adds new workloads in application scenarios where requirements change frequently. Our research results can empower robots to be upgraded from robots performing pre-set tasks to robots actively performing services, providing more direct and practical assistance to nursing staff.

Robot technology has developed rapidly in recent years and has been widely used in medical, logistics, agriculture, service industries, and other fields (Lopez et al., 2017). The emergence of edge computing technology has also made the application of artificial intelligence in intelligent elderly care even more powerful. In addition to basic offloading capabilities, edge computing provides the smart elderly care industry with value-added capabilities, application capabilities, and customer self-service capabilities. For example, cash register robots, security patrol robots (Trovato et al., 2017), warehousing robots, caddy-following robots, welcome robots (Draghici et al., 2022), food delivery robots, disinfection robots, etc., are more common in the service industry (Du et al., 2022). Such robots belong to the type of robots that perform set tasks.

After our research, we found that the nursing home provides services for individuals with nearly 20 tasks, such as nurses measuring blood pressure and massage for individual users, cooperating with nurses to conduct ward rounds, etc. Therefore, to realise the intelligent service of nursing home health-care robots, the service needs of individuals in nursing homes must be effectively predicted. The service requirements of individuals in nursing homes are mainly influenced by the behavioural attributes of individuals, so effective prediction of the behavioural attributes of individuals in nursing homes is the core task of intelligent service for nursing home health-care robots. The demand for intelligent nursing home health-care robots is to provide services to individuals after relevant analysis and decision-making.

The intelligent service scenario in Figure 1 has doctors and nurses overseeing and inspecting the services. The daily physiological monitoring, sleep monitoring, air monitoring, service matter, and room check monitoring values of all individuals in the nursing home are uploaded in real time to the health profile, health monitoring, and service management databases. From these databases, we extract relevant data for feature analysis, build models, train the constructed models on relevant features using the Gradient Boosting Decision Tree (GBDT) based machine learning model, and finally, the influence of each independent variable on the dependent variable is analyzed using the SHAP framework. The care server robot proactively provides services to individuals in nursing homes based on the predicted outcomes of the Fidan model.

Figure 1. Illustration of an intelligent service scenario.

This research aims to predict the demand for nursing home services based on the LightGBM (Light Gradient Boosting Machine) algorithm and the machine learning interpretation framework SHAP (SHapley Additive exPlanations) and data collected by IoT devices. This paper's contribution is innovatively applying the algorithm to health-care robots in nursing homes. The application value of this research is that in the application scenario of nursing homes, the Fidan model can predict the service demand of the elderly and improve the efficiency of health-care robots. The industrial significance of this research is that in the application scenario of nursing homes, nursing homes can add analysis and decision-making capabilities to nursing robots by calling the Fidan model. Nursing robots using the Fidan model can actively perform services, reducing the workload of nursing staff more directly and effectively. The main contributions of this paper are as follows:

• We propose a feature extraction model for predicting service demand based on the comprehensive analysis of prediction results from data in the health records database, health monitoring database, service management database, and a machine learning interpretable framework.

• We propose a model for predicting service demand based on the gradient boosting framework. Based on an exhaustive search, combined with relevant feature variables, we optimised the model parameters and obtained the optimal model for predicting service demand through training.

• We performed a complexity and overhead analysis while proposing the model. Based on the analysis results, we propose a human–computer interaction method and a communication method between the model algorithm layer and the robot control layer.

The following is how the rest of the paper is organised. Section 2 is a review of related work. Section 3 describes the design of the intelligent service model. Section 4 describes the implementation of the model. Section 5 describes the analysis of the experimental results. Section 6 presents the conclusions.

2. Related work

In recent years, there has been much research on service demand forecasting. In order to solve the problem that companies participating in the bicycle rental system are facing balanced supply and inconsistent demand, Y. Qin et al. uses the Washington DC shared bicycle system data set (Qin et al., 2018), based on principal component analysis and generalised recurrent neural network model, proposes a method for predicting public bicycle demand. This method achieved minimum RMSLE and MAPE values of 0.391 and 0.108, respectively. M. Phankokkruad et al. proposed a method to predict service demand through the number of patients served by medical services based on convolutional neural networks to ensure the availability of resources for the required drugs (Phankokkruad & Wacharawichanant, 2019). This method obtained an MAE value of 1.152 and an RMSE value of 1.366 on the daily outpatient data set of the obstetrics and gynecology department of a hospital in Bangkok.

In order to reduce the logistics cost of each process of service companies, R. Ramos-Carrasco et al. proposed a method of predicting tourism demand based on an artificial neural network (Ramos-Carrasco et al., 2019). This method's absolute average percentage error (AMPE) is 11.21%, which reduces the logistics cost of the service company by nearly 33%. In order to make more effective use of resources and reduce losses caused by overproduction (Çetinkaya & Erdal, 2019), Z. Çetinkaya et al. proposed a method for predicting daily dietary needs in the cafeteria of Kırıkkale University based on the artificial neural network model using student data sets. The average MSE value of the studies conducted was 0.00864. In order to facilitate vehicle scheduling, improve mileage utilisation, and reduce passenger waiting time (Jin et al., 2020), Y. Jin et al. proposed a model to predict online car-hailing demand based on stacked ensemble learning methods and large-scale data sets. The MAE and RMSE of this model increase by 6.0% and 5.2%, respectively, within a 30-min interval, demonstrating the proposed model's effectiveness and feasibility. In order to improve the accessibility of urban traffic, reduce passenger waiting time, improve the safety of taxi stations, and reduce passenger transportation costs, D. Faial et al. proposed a prediction method using flow learning (Faial et al., 2020). The method was 78 percent accurate in predicting taxi demand.

In order to realise the prediction of the demand for emergency medical services, K. Bandara et al. proposed a global prediction and reasoning framework based on the extended short-term memory network (Bandara et al., 2020). When this framework is validated using an Australian alcohol, drug, and self-harm-related emergency medical services dataset, experimental results outperform many state-of-the-art techniques, achieving competitive results in predictive accuracy. L. Munkhdalai et al. proposed a service demand prediction method based on deep learning in response to the sharp increase in demand for postal delivery services during unique festivals (Munkhdalai et al., 2020). The MAPE value of this method in the experiment is less than 0.15. In order to improve the refinement of aviation material support and increase the predictability of aviation material support (Niu et al., 2020), P. Niu et al. proposed a material demand forecasting method based on LVQ neural network, Elman neural network, and SOM neural network. Examples prove that this method has a good effect. M. A. Zháo et al. proposed a prediction method based on a support vector regression machine (Zháo & Jayadi, 2021), using variables such as seasonality, public holidays, and order peak times to realise the prediction of daily restaurant visitors and menu demand. This method achieved a MAPE of 31.2% in the experiments.

There is also a lot of research on nursing robots. Ahamed et al. used infrared sensors and Arduino UNO microcontrollers to make a robot that can remind patients or the elderly to take medicine in time and provide pure drinking water when patients need it. The robot uses infrared and ultrasonic sensors to sense the presence of cups and the level of drinking water (Ahamed et al., 2020). In order to reduce the social transmission of coronavirus disease, Manikandan et al. designed a robot that can provide patients with medicines, food, and clothes (Manikandan et al., 2021). In order to improve the comfort of the elderly robot for the elderly, Zhang et al. analyzed the vibration characteristics of the human-machine coupling system (Zhang et al., 2021). They designed a robot prototype that would not resonate with the elderly during walking. In order to ease the work of nursing staff, Jonatan et al. proposed a robot prototype for nursing homes that can deliver medicines according to a configured schedule (Jonatan et al., 2022). The above research mainly studies improving the physical experience during the robot-carrying process.

The nursing robots research above belongs to the type of robots that perform set tasks. Because task-based robots need to set tasks before running, during the actual work of nurses, the content of their work will change with the needs of the elderly in nursing homes. It then arises that while the robot is performing one task, it may be required to complete another. However, a robot can't interrupt an ongoing assignment with a task request (Zieliński et al., 2017). This interrupt handling is a complex issue for robot controllers. Some tasks cannot be terminated immediately during the startup phase, and some need to complete a series of operations before they can be suspended and resumed (Dudek et al., 2019). Therefore, task-based robots are limited in reducing nurses’ workload due to the need for more anticipation of tasks. The Fidan model proposed in this paper, based on machine learning, allows task-type robots to be transformed into active service-providing robots by predicting the corresponding service needs based on various behavioural attributes of individuals in nursing homes.

3. Model design

In this section, we first analyze the daily behaviours of all individuals in nursing homes and propose a feature extraction model. Secondly, we analyze various data and train the constructed training set based on the ensemble learning algorithm. Then, based on some task scheduling algorithms (Du et al., 2022; Zhou et al., 2022), the health-care robot application management system sends execution instructions to each health-care robot terminal. According to the business logic of intelligent service, it can be concluded that the core of intelligent service model design is the effective prediction of individual service demand (Wang et al., 2022; Zhou et al., 2021).

3.1. Feature analysis

According to our research, the service to be performed by the health-care robot in the nursing home is the behavioural needs of all individuals in the nursing home (Fuji et al., 2011). The individual demand behaviours in the nursing home are mainly divided into passive execution (Lu et al., 2021), active calling, and imperceptible participation (Khosla et al., 2013).

Passive execution behaviour means behaviour without consciousness and belongs to passive acceptance. The behavioural attributes include the detection of physiological indicators such as triglycerides, changing urine pads, changing bedding, changing clothes, indoor cleaning, diet, routine ward rounds, night patrols, etc.

The core of the intelligent service of health-care robots is the active service after the effective prediction of individual needs, which are the daily service needs of individuals. According to our research, these daily services can be categorised into five categories: physiological monitoring services, sleep monitoring services, air monitoring services, ward round monitoring services, and daily life services. Therefore, intelligent services need to predict the demand for these five categories of services effectively. Moreover, the feature analysis of these five types of data is the primary task of the intelligent service of health-care robots.

The initial data used in this study are extracted from three databases: health archives, health monitoring, and service management in the health management system. This paper extracts individual information, including gender, age, medical history, nursing level, evaluation level, physiological monitoring, sleep monitoring, ward round monitoring, air monitoring, and service items, which are more than 20 features of 10 categories as independent variables. We count the service workload of the individual, which is then used as the dependent variable, and use the user ID as the unique identifier, ultimately obtaining 13,000 groups of samples. According to the above analysis, we randomly divide the dataset into a test set for prediction evaluation and a training set for model training according to the ratio of 2:8.

3.2. Working principle of lightGBM

LightGBM is a new boosting framework proposed by Microsoft (Ke et al., 2017). The basic working principle of LightGBM is similar to that of XGBoost (Chen & Guestrin, 2016), both of which use decision trees based on learning algorithms. LightGBM has further optimised the training speed of the model. For example, LightGBM uses a histogram-based decision tree algorithm. This algorithm's basic idea is first to discretize the continuous floating-point eigenvalues into multiple integers and construct a histogram whose width is an integer number. The statistical index accumulates in the histogram when the model traverses the data according to the discretized value. After the model crosses the data once, the histogram collects the required statistics. Finally, according to the discrete value of the histogram, travel to find the optimal segmentation point.

The implementation process of the XGBoost algorithm is as follows: multiple decision tree models are used as the basic unit to form an integrated learner; based on the second-order Taylor expansion, the regular term is added to the loss function to realise the control of complexity and overfitting (Chen & Guestrin, 2016).

Formulas (1) (2) express the calculation method of the XGBoost evaluation function. (1) $\begin{array}{rl}{\mathcal{L}}^{(t)}& =\sum _{i=1}^{n}l(y_i,{\hat{y}}_i^{(t-1)}+f_t(x_i))+\mathrm{\Omega }(f_t)\end{array}$(1) (2) $\begin{array}{rl}\mathrm{\Omega }(f_t)& =\gamma \mathrm{T}+\frac{1}{2}\lambda \parallel \omega {\parallel }^2\end{array}$(2) In formula (1), ${\mathcal{L}}^{(t)}$ represents the objective function after $t$ iterations, $l(y_i,{\hat{y}}_i^{(t-1)}+f_t(x_i))$ represents the loss function, and $y_i$ represents the value of $i$ samples, ${\hat{y}}_i^{(t-1)}$ represents the i-th output after training $(t-1)$ times, $f_t(x_i)$ represents the state of $x_i$ after training $t$ times; in formula (2), $\mathrm{\Omega }(f_t)$ represents the output after training $t$ times, $\gamma$ represents the output parameter, $T$ represents the total dependent variable, $\lambda$ represents the relevant parameters in the objective function, and ω represents the corresponding degree of influence (Chen & Guestrin, 2016).

The LightGBM algorithm is an upgraded version of the XGBoost algorithm. The LightGBM algorithm uses the GOSS algorithm to reduce the number of samples (Ke et al., 2017). Suppose there are $i$data sets $\{x_1,\dots ,x_i\}$, each $x_i$ is an m-dimensional vector in $X^m$ space, and the loss function gradient is marked as $\{g_1,\dots ,g_i\}$. The processing method of the GOSS algorithm is as follows: first, arrange the training set in descending order of the absolute value of the gradient, retain the large gradient instance $a\times 100\mathrm{\%}$ subset $A$, and the small gradient instance $(1-a)\times 100\mathrm{\%}$ is the subset $A^C$, and randomly sample $b\times |A^C|$subset $B$, split the training set according to the variable gain value${\tilde{V}}_j(d)$ calculated on $A\cup B$ (Ke et al., 2017). As shown in formula (3). (3) ${\tilde{V}}_j(d)=\frac{1}{n}\left(\frac{{\left({\sum }_{x_i\in A_l}{g}_i+\frac{1-a}{n}{\sum }_{x_i\in B_l}{g}_i\right)}^2}{n_l^j(d)}+\frac{{\left({\sum }_{x_i\in A_r}{g}_i+\frac{1-a}{n}{\sum }_{x_i\in B_r}{g}_i\right)}^2}{n_r^j(d)}\right)$(3) In formula (3), $A_l=\{x_i\in A:x_{ij}\le d\}$, $A_r=\{x_i\in A:x_{ij}>d\}$, $B_l=\{x_i\in B:x_{ij}\le d\}$, $B_r=\{x_i\in B:x_{ij}>d\}$.

However, the LightGBM model is composed of N trees. For a specific training sample, we cannot confirm the influence of each feature in the training sample on the prediction result, especially for those samples that misjudged the model itself (Ke et al., 2017). The relationship between the results effectively improves the model effect or analyzes abnormal samples. Therefore, the Fidan model proposed in this paper based on the LightGBM algorithm uses the explainable machine learning framework SHAP (Shapley Additive exPlanations) to analyze the relationship between features and prediction results.

3.3. Working principle of SHAP

The principle of the SHAP value is the Shapley value, and the Shapley value is an algorithm derived from cooperative game theory. The Shapley value, created by Shapley in 1953, is an algorithm that assigns payouts to participants based on their contribution to the total payout. Participants cooperate in collective activities and obtain corresponding benefits from this cooperation. Interpreting machine learning predictions regarding Shapley, the Shapley value for each feature is the average marginal contribution of the feature value over all possible combinations of feature values. Where “total expenditure” is the model prediction for a single instance of the dataset, “participant” is the feature value of the sample, and “revenue” is the actual prediction for that instance minus the average forecast for all cases.

In the process of constructing the Fidan model, the calculation of the impact of feature variables on the prediction results is to use the machine learning interpretation framework SHAP developed by Lundberg et al. (Lundberg & Lee, 2017), as shown in formula (4). (4) ${\varphi }_i=\sum _{S\subseteq F\mathrm{\setminus }\{i\}}\frac{|S|!(|F|-|S|-1)!}{|F|!}[f_{S\cup \{i\}}(x_{S\cup \{i\}})-f_S(x_S)]$(4) In formula (4) (Lundberg & Lee, 2017), ${\varphi }_i$ represents the degree of influence of the i-th variable, $s$represents a subset of variables, $F\mathrm{\setminus }\{i\}$ represents a set of variables, $F$ represents the total input of variables, $f_{S\cup \{i\}}(x_{S\cup \{i\}})$ means the model output when the training set only has variable values in $x_{S\cup \{i\}}$, and $f_S(x_S)$ means the model output when the training set only has variable values in $x_S$.

3.4. Predictive model construction

The construction process of the Fidan model of the health-care robot task prediction model proposed in this paper mainly includes three steps: feature variable extraction, parameter optimisation, and result evaluation, as shown in Figure 2.

Figure 2. Forecast service demand model.

Figure 2 shows the implementation idea of the Fidan model, which is described in detail as follows:

• In steps 1 and 2, after analyzing the data in the health record database, health monitoring database, and service management database, we propose a feature extraction model for predicting service demand and use this model to construct a training sample set and test sample set.

• In steps 3, 4 and 5, we train the feature extraction model based on the LightGBM algorithm. The contribution of prediction results is analyzed based on the SHAP framework. Based on the analysis results, we updated the training and test sample sets.

• In step 6, we use LightGBM as the base classifier to train our proposed model for predicting service demand. Based on the prediction results of the model, we use Grid Search to optimise the model parameters.

• In steps 7 and 8, we use the test sample set to test the optimal output model and optimise the model again according to the test results.

We loop through steps 6, 7, and 8 until the accuracy reaches a preset threshold. Ultimately, we obtain an optimal model for predicting service demand.

Table

ALGORITHM 1: Fidan MODEL TRAINING
	Input:
	$D$: training dataset
	$T$: test dataset
	Output:
	Fidan model
1:	for$k$ in $1:epochs\mathrm{\_}N$ do:
2:		Train the constructed $Fidan$ model and load the $SHAP$ explanatory model
3:		Adjust the feature variables according to the contribution values of the $SHAP$ analysis
4:		Update the training dataset $D$
5:		Initially, set the model for each parameter range
6:		Load the base model, load $GridSearch$
7:		Train the constructed $Fidan$ model
8:		Update the model parameters based on the scores of each model parameter value
9:		Perform accuracy using test dataset $T$
10:		if Accuracy $>=Acc$:
11:			End of training
12:		End if
13:	End for
14:	Output Fidan model

In Algorithm 1, the $batch$ is the number of samples, the $epochs\mathrm{\_}N$ is the number of model-building exercises, and the $Acc$ is the accuracy pre-set threshold. Here the $Acc$ is a preset accuracy threshold. As shown in Algorithm 1, we first build training and testing sets using our proposed feature extraction model. In training the Fidan model, we use Grid Search to optimise the model parameters. We iteratively train the model until the accuracy reaches a pre-set threshold.

Parameter optimisation uses a way of controlling model complexity to prevent model overfitting and improve model performance. In this paper, five-fold cross-validation is used to obtain the best parameters of the model. 10,400 training samples are divided into 2,080 sets according to 5 sets as a unit. In training, the model structure parameters, parameters affecting accuracy, and parameters affecting overfitting are processed using the grid search method (GridSearchCV) to find the optimal parameters. Finally, we obtain the optimal parameter prediction model based on the LightGBM algorithm after multiple iterative tests and parameter adjustment.

3.5. Evaluation metrics

The evaluation metrics we use are Mean Square Error (MSE), coefficient of determination (R_Squared), Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and accuracy. Assuming that the predicted value is: $\hat{y}=\{{\hat{y}}_1,{\hat{y}}_2,{\hat{y}}_3,\dots \dots ,{\hat{y}}_n\}$, the real value is: $y=\{y_1,y_2,y_3,\dots \dots ,y_n\}$, the mean square error (MSE) calculation formula is shown in formula (5). (5) $MSE=\frac{1}{n}\sum _{i=1}^n({\hat{y}}_i-y_i{)}^2$(5) The value range $[0,+\mathrm{\infty })$ in formula (5), the larger the value, the larger the error. If the value is equal to 0, the predicted value is completely consistent with the actual value. The formula for calculating the coefficient of determination (R Squared) is shown in formula (6). (6) $R^2=1-\frac{{\sum }_{i=1}^n{({\hat{y}}_i-y_i)}^2}{{\sum }_{i=1}^n{({\overline{y}}_i-y_i)}^2}$(6) The numerator in formula (6) is the sum of the square difference between the preset value and the output value. The denominator is the sum of the square difference between the preset value and the output mean value. The value range of $R^2$ is [0,1]. The larger the output value, the better the training. The calculation formula of mean absolute percentage error (MAPE) is shown in formula (7). (7) $MAPE=\frac{100\text{\%}}{n}\sum _{i=1}^n\left|\frac{{\hat{y}}_i-y_i}{y_i}\right|$(7) The MAPE value range in formula (7) is $[0,+\mathrm{\infty })$. The larger the value, the greater the error. The calculation formula of the mean absolute error (MAE) is shown in formula (8). (8) $MAE=\frac{1}{n}\sum _{i=1}^n|{\hat{y}}_i-y_i|$(8) The MAE value range in formula (8) is $[0,+\mathrm{\infty })$. The larger the value, the greater the error. The accuracy calculation formula is shown in formula (9). (9) $\text{Accuracy}=\frac{TP+TN}{TP+TN+FP+FN}\times 100\mathrm{\%}$(9) In formula (9), TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative.

3.6. Method complexity and cost analysis

The Fidan model we proposed uses the LightGBM algorithm as the base classifier. The LightGBM algorithm uses a histogram algorithm when splitting. The histogram algorithm can reduce the complexity of the base model itself, thus reducing memory usage. When the histogram algorithm handles cache access, all the sample features obtain gradients in the same way so that to achieve continuous access, only the gradients need to be sorted. This mechanism significantly improves the hit rate of the cache. Since the histogram algorithm does not need to store the array of row index to leaf index, it reduces the storage capacity requirement and avoids the cache miss problem.

The LightGBM algorithm uses the unilateral gradient sampling technology GOSS (Gradient-based One-Side Sampling). The starting point of the unilateral gradient sampling technology GOSS algorithm is to reduce the number of samples. In the gradient boosting decision tree GBDT, each sample data has a different gradient value, and the sample with a slight gradient value has a relatively small training error. From this, the model learns the sample data with a slight gradient value well. Therefore, the GOSS algorithm balances sample data volume and prediction accuracy by deleting most of the small gradient samples, retaining all samples with large gradients, randomly sampling samples with small gradients in proportion, and using the remaining samples to calculate information gain. The number of samples increases the speed of model training.

3.7. Data collection

In proposing the Fidan model, we collected data from health records, monitoring, and service management. The main fields collected in health records include gender, height, weight, age, self-care assessment, level of care, living habits, dietary preferences, medical history, family medical history, cancer history, immunisation history, and medications. The equipment that collects health monitoring data mainly includes thermometers, oximeters, blood pressure monitors, blood glucose, uric acid, and cholesterol metres. These IoT devices will upload the collected data through the Internet to the server that interacts with the Fidan model. Service management data records users’ needs when they click on each service content on the APP.

4. Model implementation

In this section, we introduce the implementation of the Fidan model. Based on the lifting tree LightGBM algorithm, this model builds a basic model for predicting individual service demand based on the training data set. This model uses the method of Shapley's additional explanations (SHAP) derived from game theory to analyze the contribution of 9 categories of more than 30 characteristic variables to the output results. The relevant parameters are updated according to the analysis results, and finally, the best model parameters are obtained.

4.1. Introduction to sample data

The data set used in this paper is a non-public private data set when training the model, and the public data set Food Demand Forecasting on Kaggle is used to verify the design ideas of the model further. Some fields of the non-public private data set used in training the model are shown in Table 1.

Table 1. Description of characteristic independent variables.

Physiological monitoring	Sleep monitoring	Air monitoring	Ward round monitoring
Systolic blood pressure	Total sleep time	Carbon dioxide	Sleep inquiry
Diastolic blood pressure	Physical movements	Formaldehyde	Meal enquiry
Heart rate	Snoring	TVOC	Two questions
Blood sugar	Time out of bed	PM2.5	Mood inquiries
Uric acid	Breathing rate	PM10	Clinical chief complaint
Total cholesterol	Sleep heart rate	Temperature
Blood oxygen		Humidity
Body temperature
High density lipoprotein
Low-density lipoprotein
Triglycerides

We plan to desensitise the non-public private data set and make it public after obtaining relevant authorisation.

4.2. Individual behaviour attribute collection

According to our research, the task prediction of health care robots is mainly to effectively predict the service needs of all individuals in nursing homes. Individual service needs are mainly affected by individual behavioural attributes. Therefore, collecting individual behaviour attributes and analyzing historical records are essential tasks for medical robot task prediction. This paper's behaviour attribute acquisition modules include a central control module, signal acquisition module, and terminal use module developed by Google Android Studio, including login, health detection, health monitoring, service call, nursing service, and nursing service functions.

4.3. Implementation of health-care robot

After the individual service demand prediction is completed, sending execution instructions to each health-care robot terminal is necessary. The health-care robot application management system in this paper uses the MySQL database for data storage. It uses the Bootstrap framework and VUE framework for front-end management interface development. It also uses HTML5 Canvas for data visualisation development.

The medical robot application management system predicts the processing result according to the individual service demand and sends the movement direction to the navigation system. The individual behaviour attribute collection module's central control module controls the signal collection unit to collect signals through relevant hardware interfaces. After the central control module performs relevant processing and analysis, it finally sends the result to the business workload prediction module through the network communication module. The workload prediction model updates relevant parameters in real-time.

4.4. Human–computer interaction and communication

We integrated speech synthesis, recognition, and voice wake-up functions into the robot control software system.

Considering the interactive experience between health-care robots and the elderly, we have added a human–computer interaction module to the robot control software system. When the nursing robot arrives in front of the elderly, it will actively ask them elderly if they need services. The nursing robot will only start serving when the elderly say they need assistance. When the older adult replies that no service is necessary, the nursing robot will leave the older man and go to other older people. After the nursing robot completes the interaction with the elderly, it will send the received feedback results back to the server to provide data support for further optimisation of the model.

To gain an in-depth understanding of the service experience of the elderly with health-care robots connected to the Fidan model and those not related to the Fidan model, we have added a scoring module to the robot control software system. When the nursing robot completes the service for the elderly, the elderly will terminate the service evaluation through voice or touch screen. Evaluation indicators include:

• The timeliness of this service.

• The effectiveness of this service.

• The quality of this service.

As shown in Figure 3, to ensure that the nursing robot control layer can communicate with the model algorithm layer, we use two modes of mutual integration, cloud and local, when deploying model applications. We deploy the optimal model obtained through training locally and in the cloud. After the cloud model is updated, the locally deployed model will be updated synchronously. The robot control layer prioritises calling the models deployed in the cloud. If the data communication fails, the robot control layer will switch and call the locally deployed models in real-time.

Figure 3. The robot control layer communicates with the model algorithm layer.

We add a heartbeat communication protocol between the robot control and model algorithm layers. The protocol sets a period, and the robot control layer sends heartbeat data to the model algorithm layer. After receiving the heartbeat data, the model algorithm layer replies to the data according to the protocol. If the robot control layer receives the reply data over time, it will immediately start the model offline alarm.

5. Result analysis

5.1. Prediction results

Standard prediction algorithms are linear regression algorithm, logistic regression algorithm, support vector machine algorithm, model fusion algorithm, random forest algorithm, adaptive boosting tree algorithm, and gradient boosting decision tree algorithm. The Fidan model proposed in this paper is constructed based on a linear regression algorithm, logistic regression algorithm, support vector machine algorithm, model fusion algorithm, random forest algorithm, adaptive boosting tree algorithm, and gradient boosting decision tree algorithm. The evaluation results are compared and verified, as shown in Figure 4(a).

Figure 4. Comparison of prediction results of various methods. (a). Comparison of the prediction accuracy of GBDT with other algorithms; (b). Comparison of prediction accuracy of LightGBM with other algorithms.

As can be seen from Figure 4(a), the accuracy of the base models constructed based on the model fusion algorithm, the random forest model (Random Forest), and the gradient boosting decision tree algorithm all exceeded 80%. They all have high predictive ability, among which the basic model based on GBDT has the highest accuracy rate, and the accuracy rate of forecasting personalised service demand reaches 86.36%.

In order to further compare the evaluation results of the base model, the Fidan model proposed in this paper was constructed based on the XGBoost algorithm, Linear Regression, Random Forest, and LightGBM algorithm, respectively, during the construction process, and the evaluation was carried out. The results are compared and verified, as shown in Figure 4(b).

It can be seen from Figure 4(b) that the accuracies of the base models constructed based on the XGBoost algorithm, the Random Forest model, and the LightGBM algorithm all exceed 85%, and they all have high predictive ability. Among them, the base model based on the LightGBM algorithm has the highest accuracy, and individual service demand prediction accuracy reaches 86.61%.

The healthcare sector can turn their running nursing robots that perform set tasks into active service nursing robots with analysis and decision-making capabilities by calling the API interface encapsulated by our research results.

5.2. Analysis of the influence degree of characteristics

The individual service demand is mainly influenced by the service content and matters provided by the nursing home. The characteristic dependent variables involved in the study of this paper are the following service components: changing diaper pads, meal inquiry, calling for toilet, calling for water, and calling for massage. The correlation between the independent and dependent variables is shown in Figure 5.

Figure 5. Correlation between an independent variable and dependent variable.

Figure 5 reflects that individual service needs are strongly influenced by the evaluation level, nursing level, age, gender, medical history, and some physiological monitoring values in the health record.

5.3. Analysis of the contribution of predicted outcomes

In order to explore the factors that affect the prediction results of individual service demand, we use the SHAP algorithm to analyze the contribution of the prediction results in detail from two aspects. These two aspects are global characteristic independent variable SHAP values and analysis of individual sample characteristics.

The influence of the global independent variables on the prediction results is shown in Figure 6, where each row represents an independent variable, and the horizontal coordinate is the SHAP value. In the Figure, a point represents a sample, a bluer colour represents a smaller value of the characteristic independent variable SHAP, and a redder colour represents a more significant value of the characteristic independent variable SHAP.

Figure 6. Global feature independent variable SHAP value.

In a comprehensive analysis, it can be concluded that the predictive results of the Fidan model proposed in this paper are influenced by gender, age, medical history, level of care, level of evaluation, HDL, LDL, total cholesterol, triglycerides, and uric acid.

5.4. Model tuning and validation

In terms of model tuning, we tune the training samples. In terms of sample tuning, according to the conclusions of the previous section, the characteristic independent variables that directly affect the prediction results of the Fidan model proposed in this paper are gender, age, medical history, nursing level, evaluation level, high-density lipoprotein, low-density lipoprotein, Total cholesterol, triglycerides, uric acid. In this paper, we first fine-tune the sample, extract ten characteristic independent variables and characteristic dependent variables that directly impact the prediction results of the Fidan model from the sample, and then form a new sample.

To compare the influence of the samples before and after tuning on the prediction results of the Fidan model, two evaluation metrics, mean squared error (MSE) and the coefficient of determination (R Squared), were used. Evaluation of the prediction results.

Because the more significant the MSE value, the larger the error, and the larger the R variance value is within the range of values, the better the model-fitting effect will be. Therefore, we examine the following evaluation results. Figure 7(a) shows the results of evaluating the mean squared error of the samples before and after adjustment. Figure 7(b) shows the results of the evaluation of the R squared evaluation results of the samples before and after adjustment.

Figure 7. The MSE and R squared evaluation results before and after sample tuning. (a). Mean square error evaluation results before and after sample tuning; (b). R squared evaluation results before and after sample tuning.

We can see that the prediction accuracy of the Fidan model trained with the tuned samples for individual service demands is better than that of the Fidan model trained with the pre-tuned samples.

5.5. Physiological data collection

In this paper, during the actual software operation process, the picture of using IoT devices to collect individual physiological data is shown in Figure 8.

Figure 8. Physiological data collection

As shown in Figure 8, while collecting physiological data by IoT devices, the individual blood glucose data collected by the blood glucose monitor are then transmitted to the Android tablet. The Android tablet transmits the data through the Internet to the cloud server where the Fidan model has been deployed.

5.6. Food demand forecasting

The core of the intelligent service of health care robots studied in this paper is the prediction of the daily service demand of individuals. The daily service contains daily living service. Daily living services include changing urine pads, bedding, clothes, house cleaning, and eating and drinking. Therefore, to further validate the accuracy of the Fidan model prediction, we validated the Fidan model using Food Demand Forecasting, a publicly available dataset on Kaggle (Vidhya, 2020). The results of the validation are shown in Figure 9:

Figure 9. Comparison of different models for food demand prediction results.

In this validation, we performed the base model construction based on the linear regression algorithm, logistic regression algorithm, support vector machine algorithm, and gradient boosting decision tree algorithm in the construction process for the Fidan model proposed in this paper. As shown in Figure 9, the accuracy of our proposed Fidan model for food demand forecasting in the dataset Food Demand Forecasting, the base model constructed based on the gradient boosting decision tree algorithm, reached 82.9%. From this validation result, the Fidan model can provide practical support for the intelligent service of nursing home care robots. Using the model optimisation method described in the previous section, we will further use Food Demand Forecasting to verify the model optimisation ideas. The experimental results before and after optimisation are shown in Table 2.

Table 2. Comparison of experimental results before and after optimisation.

	MAPE	MAE	MSE	R2	Acc
before optimisation	0.3587	1.736	0.00258	0.7562	0.735
after optimisation	0.07567	0.9562	0.00753	0.8287	0.829

From Table 2, all indicators of the optimised model are better than those before optimisation. We compare the experimental results with the research on service prediction cited in this paper, and the details are shown in Table 3.

Table 3. Comparison with current research methods.

Researcher	Research objectives	Basic algorithm	MAPE
(Qin et al., 2018)	bicycle demand	Neural Networks	0.108
(Munkhdalai et al., 2020)	delivery needs	deep learning	< 0.15
(Zháo & Jayadi, 2021)	menu requirements	SVRM	0.312
Ours	food demand	LightGBM	0.0756

As can be seen from Table 3, in the current research related to this paper, the MAPE value of Y. Qin et al.’s forecast for public bicycle demand is 0.108; the MAPE value of L. Munkhdalai’s forecast for service demand is less than 0.15; M. A. Zháo et al. The predicted MAPE value of the restaurant’s daily visitors and menu demand is 0.312; the comparison in Table 3 shows that the Fidan model proposed by us based on SHAP and LightGBM has the best prediction effect.

This section describes the experimental results of the Fidan model. The experimental environment used in this paper to train the Fidan model includes Ubuntu 18.04.5 LTS (Bionic Beaver), Python 3.7.6, LightGBM 3.3.2, XGBoost 1.6.1, Numpy 1.18.5, Pandas 1.3.5, and sklearn 1.1. In collecting Internet of Things data, we used the web application framework Django of the MTV framework model. This framework is an open-source model powered by Python programming and follows MVC design and BSD copyright. The front end uses RESTful API to access the Django server. RESTful API is a REST-style API. REST is an architectural style with nothing to do with programming languages and platforms. It uses HTTP as the transmission protocol. The system uses MySQL to store health records, health monitoring, and service management data. We evaluate the experimental results using MSE, MAE, R_Squared, MAPE, and Accuracy. The Fidan model achieved an accuracy of 82.9% when verified using the Food Demand Forecasting, a public data set on Kaggle. The experimental results prove that the Fidan model can effectively support the service demand prediction of nursing home care robots.

6. Conclusion

This paper proposes a Fidan model for forecasting demand for nursing home services. The model combines the LightGBM algorithm and the explainable machine-learning SHAP algorithm. After iteration and optimisation, the Fidan model has a prediction accuracy of 86.61% for each service demand. Using our model, the health care department has realised the active service function of providing nursing robots with analysis and decision-making, providing more direct and practical assistance to nursing staff.

This paper completes the forecast of service demand with the day as the window unit. Future research will use hours as the time window unit. In addition, this paper completes the forecast of single service demand, then collects sample data of multiple service demands to study the forecast of multiple service demands. We plan to connect the Fidan model to nursing robots in nursing homes one after another and collect relevant scene data, including user experience data on the nursing robot that calls the Fidan model. Considering the large dimensions of scene data, in future research, we plan to use dimensionality reduction technology and deep neural networks to optimise further the Fidan model proposed in this paper, improve the robustness and generalisation ability of the model, and expand the application scope of the model, and more closely integrate models with real-world applications.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by Shanghai Science and Technology Project under Grant: [grant no No.22510761000]; National Natural Science Foundation of China under Grant: [grant no No.61873309, No.92046024, No.92146002]; National Key Research and Development Program of China: [grant no No.2019YFA0709502].