Monitoring Hospital Visitors Could Enhance the Prediction of the Plastic Waste Collection Demand and Its Management

Abstract

A novel framework is proposed to support the prediction of the plastic waste (PW) collection demand, route optimization, and overall management of PW from individual facilities. Based on electronic manifests from a local recycling company in Fukuoka, Japan, we developed a two-step artificial intelligence (AI)-based approach for predicting the demand for industrial PW (IPW) collection from a hospital. The daily hospital visitor numbers were introduced as a new independent variable in the IPW collection demand prediction. The stability (robustness) of each model was measured by its variance through experiments for two variable groups in four validation months. We found that introducing the visitor variables into IPW collection demand predictions was effective. A high monthly mean accuracy (85.06%) was achieved in predicting the daily IPW collection demand, which exceeded the accuracy of predictions using models without visitor records (84.44%). The stability of the Fine tree model with the highest prediction accuracy for March 2020 was 0.0466 $\mp$ 0.0174. Based on the findings of this study, we propose several strategies for waste management: enhancing prediction models, controlling visitor flows, and analyzing working patterns. This study successfully links AI techniques with a human mobility monitoring system (location data) for waste management using MATLAB.

1. Introduction

In recent years, plastic waste (PW) issues have raised increasing concern, particularly in marine pollution [1,2], biodiversity loss [3], and fossil fuel depletion [4]. Consequently, the transition from fossil fuel-based plastics to bioplastics has been promoted. In addition, a new business model based on a circular economy was proposed. However, the transition to a zero-waste business model requires additional time. Current challenges primarily focus on reducing the fossil fuel-based PW usage and generation, while increasing the recycling ratio. Accurately predicting the PW collection demand in advance is crucial for waste management, including personnel arrangements for collection and processing, route optimization, and the potential regulation of generation patterns. The PW collection demand is the amount of PW to be collected by recycling companies or waste collectors.

Previous studies have applied various statistical models, such as linear and multiple linear regression models [5], vector autoregression [6], and seasonal autoregressive integrated moving average models [7], for prediction problems. In recent years, artificial intelligence (AI) techniques have been increasingly used to analyze large datasets for prediction problems because they provide a higher accuracy in many fields such as mental healthcare [8], soil temperature prediction [9], and municipal solid waste generation [10]. However, few studies have focused on PW management [11,12,13]. In a previous study [11], an AI-based approach was developed for predicting industrial PW (IPW) generation in the wholesale and retail trade sectors, achieving a weekly mean accuracy of 93.6%. Another study [12] extended the prediction models to five additional sectors, namely supermarkets (monthly mean accuracy: 84.6%), hospitals (84.4%), logistics companies (78.2%), food manufacturing companies (82.5%), and building management companies (81.1%). As an AI-based application, study [13] integrated AI-based IPW collection demand predictions with route optimization. Despite these advancements, challenges remain in improving the prediction accuracy, evaluating the temporal robustness of each model, and developing solutions to regulate PW generation patterns for effective waste management.

Recent advancements in Information and Communication Technology have increasingly facilitated the collection of high-volume data (i.e., big data), allowing researchers and practitioners to monitor visitor patterns remotely with high temporal and spatial resolutions [14,15,16]. Several studies have focused on tracking changes in human mobility following lockdown policies [17,18,19,20,21]. However, location data have not yet been utilized in the waste management field, particularly for predicting the collection demand.

There is a significant research gap in the integration of IPW generation predictions with location data. In this context, this study aims to incorporate location data into AI-based predictions and waste management (demand response). Unlike previous studies that conducted one-step predictions [11,12,13], we developed a two-step AI-based approach for the IPW collection demand prediction, incorporating location data to enhance the prediction accuracy beyond previous research [12]. In addition to the model accuracy, the temporal robustness (stability) was validated for different seasons (four months). Moreover, we proposed new demand response solutions for PW management and updated the AI-based PW collection system of a previous study [13]. To achieve these objectives, an improvement study was conducted using data from a local recycling company in the Fukuoka Prefecture, Japan.

The remainder of this paper is organized as follows. Section 2 describes the study area and framework, including the data preparation on IPW records, the impact of holidays and weather, and hospital visitor data. This is followed by an explanation of the model fitting and validation. Section 3 discusses the results, and Section 4 presents a novel system and potential solutions for waste management and addresses the limitations and future directions. Section 5 concludes this paper by summarizing the major findings.

2. Materials and Methods

2.1. The Framework for the Proposed PW Management System

The framework of the proposed two-step prediction approach is illustrated in Figure 1. To conduct future predictions using AI, four independent variables were considered daily: past collection records, holiday impacts, climate conditions, and hospital visitor data. To confirm the effectiveness of the category conversion for weekday variables, two groups were prepared. The dataset is divided into training (learning) and validation (testing) subsets. Multiple models were trained using the Regression Learner application of the Statistics and Machine Learning Toolbox (v.24.1) in MATLAB R2024a. The training performances of all models were evaluated using the root mean square error (RMSE) values [22], mean squared error, coefficient of determination (R²), and mean absolute error. Because of the large number of variables that may affect the running time and prediction accuracy, previous studies have demonstrated the effectiveness of feature selection in machine learning [8,12,23,24]. The optimal group of variables for the prediction was determined using a stepwise selection method. Subsequently, future predictions were generated using all fitted models, and the prediction accuracy was validated over a one-month validation period in four different seasons using the mean absolute percentage error (MAPE) indicator [25]. To enhance the prediction accuracy, a two-step method was developed, which involved an initial prediction of hospital visitors, followed by a prediction of the IPW collection demand. In the second prediction step, the number of hospital visitors on the previous day was incorporated as an independent variable. Finally, the potential future applications of this approach to waste management are proposed.

2.2. Data Processing

The daily IPW collection demands of a hospital in Fukuoka Prefecture from April 2018 to September 2020 were obtained from an electronic manifest provided by a local recycling company. The raw manifest data contained individual records of the daily amount collected (kg), collection date (from 1 April 2018 to 30 September 2020), facility name, intermediate treatment facilities (destinations of collections), and name of the collector for each collection. From the list, the hospital with the most collection records was selected as the target, and no adjustments were made for the missing values owing to a lack of information on the collection schedule. The soft vinyl waste collected by this company was compressed into cube-shaped plastic fuel; thus, we excluded other non-recyclable PW (i.e., infectious waste) generated in the hospital from our analysis owing to data limitations. To ensure prediction accuracy, eight additional independent variables were included along with the response variable (daily amount collected), as presented in

Table 1. The description of the independent variables used for predicting the IPW collection demand.

Variable	Description	Source
IPW_fday	Amount of IPW collected on the former day	Electronic manifest
IPW_sd_fw	IPW collected on the same weekday of the former week
Num_day	Number of days since 1 April 2018
Interval	The interval between the present and former collection day
Clim_fd	Category type on climate impact of the former day: wind speed ≥ 6 m s−1, daily rainfall ≥ 5 mm, highest temperature ≥ 35 degree	Daily weather report [26]
Holiday_fd	Category type on the former day belongs to holiday or not	Japanese calendar for 2018–2020
Weekday	What day it was
Vis_fday	Visitor amount on former day	Location data [27]

. Two of these variables, which day of the week (Weekday) and whether the previous day was a holiday (Holiday_fd), were determined by referring to the Japanese calendar for 2018–2020. As high temperatures, strong winds, and heavy rainfall can limit consumer activities and IPW collection in hospitals, information on the previous day’s weather was added to the model based on daily weather reports [26]. Because hospital visitors (including staff and patients) influence IPW generation, the number of visitors from the previous day extracted from the location data [27] was incorporated into the model as an independent variable.

Location data (i.e., the number of designated facility users observed via mobile phone GPS) were obtained using the KDDI Location Analyzer (KLA) [27]. The KLA is an advanced Geographic Information System that utilizes GPS location data as a self-analysis tool to examine trading areas, visitor trends, and time-based movement patterns. Integrated with Google Maps, the KLA enables the estimation of visitor numbers and visit frequencies within a defined geographic area (geo-fence). Specifically, the KLA anonymized subscriber location data and extrapolated them using census data to estimate population distribution in Japan. In this study, a geo-fence was defined around hospital buildings and parking areas (Figure 2). Owing to the processing methodology of the KLA, visitors who remained within the geo-fence for more than 15 min were recorded as facility users. Daily user data from 2018 to 2020 were extracted for analysis.

The dataset used in this study was limited to KDDI smartphone owners aged 20 years and above, who provided consent for GPS data usage. Consequently, the number of visitors included in this study was lower than the actual population. As visitors (perhaps consumers of PW) may impact PW consumption, introducing this variable in the prediction will cause sampling bias and potentially underestimate the IPW collection demand in the prediction process.

Before fitting the model, outliers in training data could be identified by a 3-Sigma method (mean ± 3 × standard deviation), and linear interpolation could be used as a cleaning method to fill the outliers. A previous study [12] demonstrated that the prediction accuracy using IPW data without data smoothing and outlier embedding was higher than that with prior data processing. Therefore, we omitted this for better comparison. To verify the effectiveness of category conversion (for the weekday variable), another experiment was conducted using the converted weekday and other variables. To check the robustness of the models, we prepared test data for four months during different seasons (December 2019, March, June, and September 2020). Consequently, four datasets were prepared for training.

2.3. Data Learning and Prediction

To conduct the model training (learning) and future prediction (validation), MATLAB R2024a was used. Because MATLAB includes multiple machine and deep learning models, various models were compared to identify the best-performing prediction model. As described above, a two-step prediction approach was developed. In the first step, daily hospital visitor numbers were set as the response variables, whereas the independent variables included factors similar to those used for the IPW collection demand prediction (Figure 1 and Table 1). In the second step, the number of visitors from the previous day was incorporated as an independent variable for training (fitting) the IPW collection demand model, and the predicted visitor numbers from Step 1 were used as input variables for future IPW collection demand predictions. To assess the impact of incorporating visitor numbers, two prediction cases were defined. In Case 1, all variables presented in Table 1 were included. In Case 2, the same variables as used in a previous study [12] were prepared (Case 1 variables except for the visitor variable). Previous studies performed feature selection using a sensitivity method [11,12], and the prediction accuracy of IPW from hospitals by the best model without climate impact variables (Clim_fd) was higher than that with climate impact variables. Thus, it was excluded from the models for the visitor number and IPW collection demand prediction. The default hyperparameters for each model and tuning ranges for some models [28,29] set by MATLAB developers were used without additional adjustments (see default settings in SI Table S1). Cross-validation was performed by repeatedly fitting all models, and the optimal model was selected based on the lowest MAPE. To evaluate the prediction accuracy of each model, the MAPE was calculated using Equation (1).

$$\mathrm{MAPE}={\displaystyle \frac{100}{n}}{\displaystyle \sum _{i=1}^n\left|{\displaystyle \frac{f_i-y_i}{y_i}}\right|}$$

(1)

where

• $n$: is the number of data;
• $y_i$: is the observed value of data $i$;
• $f_i$: is the predicted value of data $i$

2.4. The Measurement of the Stability

The stability of the model explicitly refers to the robustness of its computing performance, which significantly affects its application range [30]. A model with high stability can better resist external interference, that is, the robustness of the algorithm to data disturbances. Both the error and variance determine the model’s generalization error, and a high variance is the culprit of unstable behavior. Therefore, a stability test was conducted on the model, and the specific formula is as follows:

$$S_{var}(i)=\mathit{Var}\left(\left|{\displaystyle \frac{y_i-f_i}{y_i}}\right|\right)$$

(2)

2.5. Sensitivity Analysis

To confirm error propagation by introducing visitor numbers, a sensitivity analysis of the prediction accuracy was conducted for the visitor number and IPW collection demand. To show representative results, the validation month with the lowest MAPE was selected for the analysis. Using the one-factor-at-a-time method [31,32,33], the MAPE was recorded for each model in validation when the daily values of visitor numbers were changed from −10% to 10%. Finally, we checked the sensitivity as the variation ratio of the MAPE before and after the change in the visitor number.

3. Results

3.1. Model Performance in Training Process

The RMSE and R² values for all models evaluated during the training process on the IPW collection demand are presented in

Table 2. The fitting performances for the RMSE, and the R² for all models of the IPW collection demand (examples for the September 2020 validation period).

Category	Model	RMSE	R2
Linear regression	Linear Regression	7.9329	0.3073
	Interactions Linear	7.8545	0.3210
	Robust Linear	7.9409	0.3059
	Stepwise Linear	7.9304	0.3078
Decision tree	Fine Tree	8.3854	0.2234
	Medium Tree	7.7126	0.3430
	Coarse Tree	7.4149	0.3928
Support vector machine (SVM)	Linear SVM	8.0155	0.2904
	Quadratic SVM	7.9639	0.2995
	Cubic SVM	8.0335	0.2872
	Fine Gaussian SVM	8.5623	0.1903
	Medium Gaussian SVM	7.8641	0.3170
	Coarse Gaussian SVM	7.8844	0.3134
Effectively trained linear regression	Efficient Linear Least Squares	8.8464	0.1357
	Efficient Linear SVM	9.5039	0.0024
Ensemble	Boosted Trees	7.3209	0.4081
	Bagged Trees	7.2590	0.4180
Gaussian process regression (GPR)	Rational Quadratic	7.7533	0.3361
	Squared Exponential	7.6905	0.3468
	Matern 5/2	7.6633	0.3514
	Exponential	7.7315	0.3398
Kernel regression	SVM Kernel	7.8289	0.3231
	Least Squares Kernel	8.2627	0.2460
Neural network	Narrow Neural Network	11.1677	−0.3774
	Medium Neural Network	8.4143	0.2181
	Wide Neural Network	8.2451	0.2492
	Bi-layered Neural Network	7.8983	0.3110
	Tri-layered Neural Network	7.7375	0.3388
Optimal models	Optimal Tree	7.2842	0.4140
	Optimal SVM	7.9653	0.2993
	Optimal Efficient Linear	8.8458	0.1358
	Optimal GPR	7.2924	0.4127
	Optimal Kernel	7.7967	0.3286
	Optimal Ensemble	7.1936	0.4285
	Optimal Neural Network	9.5113	0.0009

. The results indicated that the Bagged Trees model achieved the lowest RMSE in this phase

3.2. Prediction Accuracy of Each Model

As described above, two variable groups were prepared and used for model training and the validation of the IPW collection demand.

Table 3. The MAPE of predictions for four individual months with all models explored using machine learning using the category-converted weekday variable and others.

Category	Model	MAPE for Validation Periods
		December (2019)	March (2020)	June (2020)	September (2020)	Average
Linear regression	Linear Regression	0.2577	0.2538	0.1995	0.2433	0.2386
	Interactions Linear	0.2652	0.2343	0.1915	0.2689	0.2400
	Robust Linear	0.2575	0.2531	0.1991	0.2480	0.2394
	Stepwise Linear	0.2634	0.2414	0.2011	0.2556	0.2404
Decision tree	Fine Tree	0.1715	0.1355	0.1977	0.1852	0.1725
	Medium Tree	0.1454	0.1255	0.1761	0.1737	0.1552
	Coarse Tree	0.1458	0.1405	0.1641	0.1904	0.1602
Support vector machine (SVM)	Linear SVM	0.2466	0.2426	0.2159	0.2884	0.2484
	Quadratic SVM	0.2511	0.2109	0.2149	0.4713	0.2871
	Cubic SVM	0.2308	0.1997	0.1914	0.1905	0.2031
	Fine Gaussian SVM	0.3402	0.2121	0.2057	0.1631	0.2303
	Medium Gaussian SVM	0.2312	0.1749	0.1638	0.1494	0.1798
	Coarse Gaussian SVM	0.2439	0.2300	0.2131	0.3194	0.2516
Effectively trained linear regression	Efficient Linear Least Squares	0.2534	0.2566	0.2010	0.2004	0.2279
	Efficient Linear SVM	0.3047	0.3002	0.2206	0.1885	0.2535
Ensemble	Boosted Trees	0.1706	0.1693	0.1638	0.1670	0.1677
	Bagged Trees	0.1705	0.1468	0.1594	0.1681	0.1612
Gaussian process regression (GPR)	Rational Quadratic	0.2485	0.2173	0.1626	0.1562	0.1962
	Squared Exponential	0.2518	0.2110	0.1580	0.1532	0.1935
	Matern 5/2	0.2610	0.2028	0.1555	0.1592	0.1946
	Exponential	0.2604	0.1995	0.1539	0.1580	0.1929
Kernel regression	SVM Kernel	0.2768	0.2335	0.1707	0.2366	0.2294
	Least Squares Kernel	0.2779	0.2129	0.1619	0.1846	0.2093
Neural network	Narrow Neural Network	0.3420	0.1904	0.2972	0.1512	0.2452
	Medium Neural Network	0.2702	0.2159	0.1919	0.2061	0.2210
	Wide Neural Network	0.1900	0.2159	0.1661	0.1925	0.1911
	Bi-layered Neural Network	0.2435	0.2069	0.1966	0.2334	0.2201
	Tri-layered Neural Network	0.2363	0.2222	0.1640	0.1560	0.1946
Optimal models	Optimal tree	0.1472	0.1990	0.1796	0.1918	0.1794
	Optimal SVM	0.2691	0.3058	0.1773	0.1878	0.2350
	Optimal Efficient Linear	0.2533	0.2566	0.1953	0.2004	0.2264
	Optimal GPR	0.2162	0.1403	0.1597	0.1628	0.1698
	Optimal Kernel	0.2298	0.2188	0.1615	0.1518	0.1905
	Optimal Ensemble	0.1567	0.1630	0.1584	0.1699	0.1620
	Optimal Neural Network	0.2732	0.2538	0.1964	0.1903	0.2284
The model with the lowest MAPE		Medium tree	Medium tree	Exponential	Medium Gaussian SVM	Medium tree

lists the MAPE of the predictions for four individual months, with all models explored using machine learning using category-converted weekday variables and others. For December 2019, the Medium tree model was the best model (with the lowest MAPE of 0.1454 and a monthly mean prediction accuracy of 85.46%). For March 2020, the Medium tree model was the best model (with the lowest MAPE of 0.1255 and an accuracy of 87.45%). For June 2020, the Exponential model was the best model (with the lowest MAPE of 0.1539 and an accuracy of 84.61%). For September 2020, the Medium Gaussian SVM model was the best model (with the lowest MAPE of 0.1494 and an accuracy of 85.06%). Figure 3 shows the performance during the training phase. Although the Bagged Trees model achieved the lowest RMSE in this phase, its prediction accuracy was not the best in the validation period. This indicated that the model with the lowest RMSE during training did not necessarily guarantee the highest predictive accuracy in the validation phase. Focusing on the average value among these four months, the Medium tree model was the best model (with the lowest MAPE of 0.1552 and an accuracy of 84.48%). The prediction accuracy of each model for the visitor number prediction using the category-converted weekday variable group validated for four individual months is presented in SI Table S2.

Table 4. The MAPE of predictions for four individual months with all models explored using machine learning using the non-category-converted weekday variable and others.

Category	Model	MAPE for Validation Periods
		December (2019)	March (2020)	June (2020)	September (2020)	Average
Linear regression	Linear Regression	0.2508	0.2490	0.1880	0.2379	0.2314
	Interactions Linear	0.2633	0.2015	0.1711	0.2934	0.2324
	Robust Linear	0.2492	0.2470	0.1887	0.2439	0.2322
	Stepwise Linear	0.2561	0.1978	0.1700	0.2895	0.2284
Decision tree	Fine Tree	0.1765	0.1014	0.1844	0.1822	0.1611
	Medium Tree	0.1464	0.1363	0.1539	0.1579	0.1486
	Coarse Tree	0.1220	0.1427	0.1757	0.1868	0.1568
Support vector machine (SVM)	Linear SVM	0.2328	0.2301	0.2017	0.2772	0.2355
	Quadratic SVM	0.2352	0.1727	0.2255	0.3700	0.2508
	Cubic SVM	0.2235	0.1564	0.1578	0.2059	0.1859
	Fine Gaussian SVM	0.3151	0.1405	0.1709	0.1580	0.1961
	Medium Gaussian SVM	0.2251	0.1441	0.1731	0.2204	0.1907
	Coarse Gaussian SVM	0.2335	0.2185	0.1888	0.2823	0.2308
Effectively trained linear regression	Efficient Linear Least Squares	0.2575	0.2603	0.2013	0.1955	0.2286
	Efficient Linear SVM	0.3068	0.3035	0.2184	0.1916	0.2551
Ensemble	Boosted Trees	0.1602	0.1521	0.1725	0.1720	0.1642
	Bagged Trees	0.1501	0.1269	0.1621	0.1688	0.1520
Gaussian process regression (GPR)	Rational Quadratic	0.2459	0.1934	0.1629	0.2101	0.2031
	Squared Exponential	0.2450	0.1905	0.1584	0.2072	0.2003
	Matern 5/2	0.2466	0.1797	0.1555	0.1939	0.1939
	Exponential	0.2427	0.1907	0.1575	0.2071	0.1995
Kernel regression	SVM Kernel	0.2720	0.2157	0.1586	0.2660	0.2281
	Least Squares Kernel	0.2296	0.2103	0.1507	0.2663	0.2142
Neural network	Narrow Neural Network	8.8 × 1022	3.0 × 109	7.3 × 1011	2.8 × 1011	2.2 × 1022
	Medium Neural Network	0.2218	0.2325	0.1786	0.3101	0.2357
	Wide Neural Network	0.2477	0.1984	0.1960	0.1988	0.2102
	Bi-layered Neural Network	0.2142	0.1640	0.1512	0.1700	0.1749
	Tri-layered Neural Network	0.2193	0.1478	0.1577	0.1804	0.1763
Optimal models	Optimal Tree	0.2355	0.1377	0.1591	0.1616	0.1735
	Optimal SVM	0.3084	0.2429	0.2229	0.2504	0.2561
	Optimal Efficient Linear	0.2575	0.2603	0.1999	0.1985	0.2290
	Optimal GPR	0.2450	0.1407	0.1456	0.1992	0.1826
	Optimal Kernel	0.2205	0.1621	0.1507	0.2017	0.1837
	Optimal Ensemble	0.1542	0.1508	0.1644	0.1870	0.1641
	Optimal Neural Network	0.2521	0.2490	0.1895	0.1601	0.2127
The model with the lowest MAPE		Coarse tree	Fine tree	Optimal GPR	Medium tree	Medium tree

lists the MAPE of predictions for four individual months, with all models explored using machine learning using the non-category-converted weekday variable and others. For December 2019, the Coarse tree model was the best model (with the lowest MAPE of 0.1220 and a monthly mean prediction accuracy of 87.80%). For March 2020, the Fine tree model was the best model (with the lowest MAPE of 0.1014 and an accuracy of 89.86%). For June 2020, the Optimal GPR model was the best model (with the lowest MAPE of 0.1456 and an accuracy of 85.44%). For September 2020, the Medium tree model was the best model (with the lowest MAPE of 0.1579 and an accuracy of 84.21%). Focusing on the average value among these four months, the Medium tree model was the best model (with the lowest MAPE of 0.1486 and an accuracy of 85.14%). The MAPE showed that the Fine tree model achieved the highest prediction accuracy for December 2019 in this experiment. Table 3 shows that the prediction accuracy of the best models using the non-category-converted weekday variable group was higher than that using the category-converted weekday variable group, except for the validation period of September 2020. This implies that the category conversion was effective in the validation period of September 2020. The prediction accuracy of each model for the visitor number prediction using the non-category-converted weekday variable group validated for four individual months is presented in SI Table S3.

3.3. Stability of Each Model

The stability (robustness) of each model was measured using its variance.

Table 5. The stability test result of each model explored using machine learning using the category-converted weekday variable and others.

Category	Model	Variance
		December (2019)	March (2020)	June (2020)	September (2020)
Linear regression	Linear Regression	0.0187	0.0288	0.0391	0.0547
	Interactions Linear	0.0160	0.0256	0.0357	0.0737
	Robust Linear	0.0185	0.0292	0.0398	0.0574
	Stepwise Linear	0.0163	0.0216	0.0370	0.0678
Decision tree	Fine Tree	0.0482	0.0763	0.0551	0.0621
	Medium Tree	0.0259	0.0565	0.0237	0.0380
	Coarse Tree	0.0274	0.0566	0.0205	0.0371
Support vector machine (SVM)	Linear SVM	0.0220	0.0386	0.0615	0.0792
	Quadratic SVM	0.0208	0.0536	0.0518	0.1550
	Cubic SVM	0.0226	0.0702	0.0387	0.0470
	Fine Gaussian SVM	0.0232	0.0269	0.0182	0.0211
	Medium Gaussian SVM	0.0170	0.0524	0.0184	0.0359
	Coarse Gaussian SVM	0.0204	0.0409	0.0669	0.0898
Effectively trained linear regression	Efficient Linear Least Squares	0.0178	0.0193	0.0303	0.0334
	Efficient Linear SVM	0.0106	0.0051	0.0153	0.0161
Ensemble	Boosted Trees	0.0355	0.0296	0.0139	0.0252
	Bagged Trees	0.0368	0.0586	0.0229	0.0413
Gaussian process regression (GPR)	Rational Quadratic	0.0156	0.0312	0.0194	0.0339
	Squared Exponential	0.0142	0.0302	0.0191	0.0342
	Matern 5/2	0.0149	0.0292	0.0192	0.0370
	Exponential	0.0162	0.0296	0.0188	0.0360
Kernel regression	SVM Kernel	0.0133	0.0260	0.0346	0.0692
	Least Squares Kernel	0.0201	0.0339	0.0157	0.0415
Neural network	Narrow Neural Network	0.0508	0.0881	0.1329	0.0170
	Medium Neural Network	0.0229	0.0308	0.0227	0.0533
	Wide Neural Network	0.0175	0.0356	0.0221	0.0382
	Bi-layered Neural Network	0.0206	0.0511	0.0375	0.0597
	Tri-layered Neural Network	0.0180	0.0436	0.0192	0.0294
Optimal models	Optimal Tree	0.0272	0.0443	0.0236	0.0484
	Optimal SVM	0.0190	0.0057	0.0222	0.0472
	Optimal Efficient Linear	0.0179	0.0193	0.0313	0.0334
	Optimal GPR	0.0161	0.0330	0.0290	0.0350
	Optimal Kernel	0.0183	0.0345	0.0163	0.0339
	Optimal Ensemble	0.0458	0.0517	0.0193	0.0449
	Optimal Neural Network	0.0186	0.0288	0.0391	0.0189
The model with the lowest variance		Efficient Linear SVM	Efficient Linear SVM	Boosted Trees	Efficient Linear SVM

presents the stability test results for each model explored using the category-converted weekday variable group. For December 2019, March 2020, and September 2020, an Efficient Linear SVM was the most stable model (with the lowest variances of 0.0106, 0.0051, and 0.0161, respectively). For June 2020, the Boosted Trees model was the most stable model (with the lowest variance of 0.0139). As presented in Table 3, the stability of the Medium tree model with the highest prediction accuracy for December 2019 and March 2020 is 0.0401 $\mp$ 0.0164. The stability of the Exponential model with the highest prediction accuracy for June 2020 was 0.0261 $\mp$ 0.0099. The stability of the Medium Gaussian SVM model with the highest prediction accuracy for September 2020 was 0.0347 $\mp$ 0.0177.

Table 6. The stability test result of each model explored using machine learning using the non-category-converted weekday variable and others.

Category	Model	Variance
		December (2019)	March (2020)	June (2020)	September (2020)
Linear regression	Linear Regression	0.0196	0.0311	0.0429	0.0574
	Interactions Linear	0.0119	0.0152	0.0311	0.0732
	Robust Linear	0.0194	0.0320	0.0443	0.0607
	Stepwise Linear	0.0116	0.0142	0.0333	0.0829
Decision tree	Fine Tree	0.0523	0.0291	0.0640	0.0482
	Medium Tree	0.0258	0.0364	0.0274	0.0291
	Coarse Tree	0.0417	0.0550	0.0205	0.0303
Support vector machine (SVM)	Linear SVM	0.0217	0.0468	0.0624	0.0799
	Quadratic SVM	0.0145	0.0346	0.0515	0.1102
	Cubic SVM	0.0122	0.0216	0.0207	0.0392
	Fine Gaussian SVM	0.0259	0.0284	0.0393	0.0186
	Medium Gaussian SVM	0.0142	0.0259	0.0254	0.0547
	Coarse Gaussian SVM	0.0196	0.0408	0.0571	0.0825
Effectively trained linear regression	Efficient Linear Least Squares	0.0177	0.0196	0.0283	0.0333
	Efficient Linear SVM	0.0110	0.0053	0.0150	0.0151
Ensemble	Boosted Trees	0.0237	0.0212	0.0191	0.0283
	Bagged Trees	0.0409	0.0476	0.0366	0.0338
Gaussian process regression (GPR)	Rational Quadratic	0.0117	0.0170	0.0242	0.0517
	Squared Exponential	0.0110	0.0161	0.0234	0.0495
	Matern 5/2	0.0115	0.0173	0.0222	0.0434
	Exponential	0.0108	0.0160	0.0231	0.0497
Kernel regression	SVM Kernel	0.0097	0.0185	0.0256	0.0809
	Least Squares Kernel	0.0161	0.0169	0.0183	0.0752
Neural network	Narrow Neural Network	2.0 × 1047	1.6 × 1020	1.1 × 1025	1.7 × 1024
	Medium Neural Network	0.0147	0.0135	0.0333	0.0845
	Wide Neural Network	0.0292	0.0164	0.0294	0.0479
	Bi-layered Neural Network	0.0140	0.0192	0.0230	0.0360
	Tri-layered Neural Network	0.0163	0.0199	0.0256	0.0409
Optimal models	Optimal Tree	0.0267	0.0617	0.0206	0.0269
	Optimal SVM	0.0115	0.0385	0.0163	0.0669
	Optimal Efficient Linear	0.0177	0.0196	0.0284	0.0337
	Optimal GPR	0.0110	0.0287	0.0254	0.0579
	Optimal Kernel	0.0137	0.0237	0.0181	0.0339
	Optimal Ensemble	0.0397	0.0274	0.0197	0.0477
	Optimal Neural Network	0.0194	0.0311	0.0426	0.0298
The model with the lowest variance		SVM Kernel	Efficient Linear SVM	Efficient Linear SVM	Efficient Linear SVM

presents the stability test results for each model explored using the category-converted weekday variable group. For March 2020, June 2020, and September 2020, an Efficient Linear SVM was the most stable model (with the lowest variances of 0.0053, 0.0150, and 0.0151, respectively). For December 2019, the SVM Kernel model was the most stable model (with the lowest variance of 0.0097). As presented in Table 4, the stability of the Coarse tree model with the highest prediction accuracy for December 2019 is.0378 $\mp$ 0.0173. The stability of the Fine tree model with the highest prediction accuracy for March 2020 was 0.0466 $\mp$ 0.0174. The stability of the Optimal GPR model with the highest prediction accuracy for June 2020 was 0.0344 $\mp$ 0.0235. The stability of the Medium tree model with the highest prediction accuracy for September 2020 was 0.0311 $\mp$ 0.0053.

3.4. Error Propagation Checked by Sensitivity Analysis

As described in Section 2.5, the variation ratio of the MAPE before and after the change in visitor numbers was analyzed to confirm the error propagation by introducing visitor numbers. As detected using the MAPE (Table 4), the prediction for March 2020 using the Fine tree model was the most accurate (MAPE 0.1014, accuracy 89.86%) and was made the target of the sensitivity analysis. The sensitivity results for all models predicted for March 2020 when changing the visitor numbers from −10% to 10% are presented in SI Table S4. For the visitor number prediction, the magnitude of the sensitivity ranged from −100% to (2.1 × 10¹³)%. The variation in the MAPE for the Fine tree model was small, ranging from −(7.4× 10⁻¹⁴)% to −(3.7 × 10⁻¹⁴)%. For the IPW collection demand prediction, the magnitude of the sensitivity also varied by model, ranging from −100% to (2.0 × 10⁴)%. The variation in the MAPE from the Fine tree model was stable at 9.89% when the daily values of visitor numbers were changed from −10% to 10%. The variation in the MAPE from the Boosted Trees model was stable. The MAPEs from the Stepwise linear, Medium tree, and Coarse tree models were confirmed without variation, even when the daily values of visitor numbers changed.

4. Discussion

4.1. A Comparison with a Previous Study

A previous study [12] achieved a monthly mean prediction accuracy of 84.44% for September 2020. As presented in Table 3, a higher monthly mean prediction accuracy of 85.06% is achieved using the proposed approach for the same validation period. Figure 4A compares the predicted and observed values of the IPW collection for 1–30 September 2020 using the Rational Quadratic model (without the visitor variable) in a previous study [12]. Figure 4B presents the same comparison using the Medium Gaussian SVM model (with visitor variables). These results demonstrate the effectiveness of incorporating visitor variables into the IPW collection demand predictions. Previous studies have demonstrated the role of tourists in the generation of PW [34,35] and visitors to hospitals in the generation of medical waste [36,37]. Consistent with these studies, we demonstrated that the number of visitors to a facility influences plastic packaging consumption and PW generation.

4.2. Future Prediction

Figure 5 shows an example of predicting the daily IPW collection amount and hospital visitor numbers for a future week in October 2020 using the proposed approach. This method demonstrates the feasibility of making accurate short-term predictions for response variables when the daily independent variables are continuously updated.

4.3. The Proposal for an Intelligent Waste Management System

Building on the findings of a previous study [13], a framework for an improved system is proposed, as illustrated in Figure 6. The AI techniques employed in this study include the AI-based forecasting of the collection demand, demand response strategies for PW collection through the regulation of independent variables such as visitor numbers, and optimization of vehicle routing problems [38,39]. As demonstrated in this study, the independent variables play a crucial role in predicting the collection demand for the IPW. These findings suggest that managing these variables could effectively reduce future IPW.

4.4. Potential Solutions for Waste Management

Based on the findings of this study, we propose the following strategies for managing key variables to control future IPW generation and disposal.

1. Advanced predictions using accumulated big data. The accurate forecasting of future waste generation can support the development of effective waste management plans. This is particularly important during peak seasons, such as summer and winter, when adjusting personnel schedules to account for extreme weather conditions could improve the working conditions in recycling facilities.

2. Visitor control policies. Previous studies have demonstrated the role of hospital visitors in the generation of medical waste [36,37]. Regarding the effectiveness of the visitor control policy, one study showed evidence for its effect on the safety and security of healthcare facilities [40], two studies demonstrated its role in solid waste management [41,42], and another study assessed the carbon dioxide emission reduction aspect [43]. Consistent with these studies, we demonstrated that the number of visitors to a facility influences plastic packaging consumption and PW generation. Hospital administrators and staff can implement policies to regulate visitor-related activities and schedule them on non-peak days to avoid an IPW collection rush.

3. The adjustment of work patterns. Similarly to the visitor impact, work schedules also affect plastic consumption and waste generation in hospitals. For instance, hospitalization schedules could be modified to distribute patient admissions more evenly throughout the week rather than concentrating them on peak days such as Mondays and Fridays.

4.5. Limitations, Applicability of This Framework, and Future Plans

The prediction accuracy in this study was constrained by the model selection. The ability to predict the IPW collection demand can be enhanced by developing more advanced AI models. Although visitor data have proven valuable for IPW collection demand predictions, data acquisition remains costly and inaccessible to non-smartphone users. The dataset used in this study was limited to KDDI smartphone owners aged 20 years and above, who provided consent for GPS data usage. Consequently, the estimated population was lower than the actual population. However, this dataset accounted for a substantial proportion of the population [16,44]. Additionally, the prediction accuracy is limited by the reliability of future temperature forecasts.

The PW collection records (electronic manifest) include those from other types of facilities (e.g., factories, schools, restaurants), and location data are also available for these facilities. Thus, this approach is applicable to PW management in other types of facilities and can be used across Japan. For regions with varying levels of data accessibility, time-consuming and laborious paper-based manifestations for PW collection and visitor registration forms are helpful for data acquisition.

Moving forward, we aim to incorporate more variables (such as hospital occupancy rates and waste bin fill levels), consider ensemble methods, and adjust the hyperparameters of each model to improve the model accuracy, develop AI-based applications, and validate their effectiveness in different settings. For instance, we will explore real-time data integration, reservation information utilization, and mobile app-based visitor tracking as more inclusive and flexible alternatives to third-party datasets. Subsequently, a cost–benefit or environmental impact analysis [13] will be conducted on the proposed intelligent waste management system. Once the latest data on the IPW collection are available, we will update the results immediately.

5. Conclusions

This study explored an AI-based approach to predict future IPW collection demands by integrating location data from a hospital in Kitakyushu, Japan. The model achieved a high monthly mean accuracy of 89.86% (using the Fine tree model in the March 2020 validation period) in predicting the daily IPW collection demand. Additionally, incorporating visitor records into the model resulted in a higher accuracy (85.06%) compared to models in a previous study that excluded visitor data (84.44%). These results highlight the effectiveness of visitor data in improving the IPW collection demand prediction. This implies that regulating hospital visitor numbers could serve as a helpful strategy to manage future IPW generation. The model performance varied across the different approaches, and the best-performing model was selected for validation based on the monthly accuracy metrics. The stability (robustness) of each model was measured by its variance through experiments with two variable groups over four validation months. For instance, the stability of the Fine tree model with the highest prediction accuracy for March 2020 was 0.0466 $\mp$ 0.0174. Despite its usefulness, the practical impact of marginal accuracy improvements (optimizing sample size, advanced techniques) is limited.

A framework for an improved waste management system was proposed, incorporating the AI-based forecasting of the collection demand, demand response strategies through the control of independent variables such as visitor numbers and working patterns, and optimization of vehicle routing problems. Based on these findings, several strategies for waste management can be considered, including refining prediction models, regulating the visitor flow, and assessing working patterns. Moving forward, additional predictive variables will be incorporated to improve the model accuracy, and AI-based applications will be developed to optimize waste management processes.