Applying Smart Healthcare and ESG Concepts to Optimize Elderly Health Management

Abstract

As the aging population grows, ensuring effective and sustainable health management for elderly individuals has become a critical challenge. This study explores the integration of smart healthcare technologies and ESG (Environmental, Social, and Governance) principles to enhance elderly health management through data-driven strategies. Using the MIMIC-III database, this study evaluates five machine learning models (Adaboost, Bagging, Catboost, GaussianNB, and SVC) through ten-fold cross-validation to predict 3-day and 30-day mortality rates among elderly ICU patients. The Bagging model achieved the best performance with an AUROC of 0.80, demonstrating the potential of smart healthcare in mortality prediction. These technologies enhance predictive accuracy, enabling the timely identification of high-risk patients and effective intervention. Through the application of smart data integration methods, this study demonstrates how combining clinical indicators with socioeconomic factors can improve healthcare equity and efficiency. Furthermore, by aligning smart healthcare development with ESG concepts, we emphasize the importance of sustainability, social responsibility, and governance transparency in future healthcare systems. The findings offer valuable contributions toward building an interoperable and ethical health ecosystem, supporting early risk identification, improved care outcomes, and the promotion of healthy living for the elderly population.

1. Introduction

As the trend of global population aging becomes increasingly evident, the health issues of the elderly population are receiving more and more attention. Changes in the population structure have led to a rapid increase in the demand for intensive care units (ICUs), surpassing the capabilities of clinical doctors. The physiological changes associated with aging affect every organ system, making specialized training and education for intensive care physicians crucial for guiding clinical care [1,2]. The primary purpose of establishing ICUs is to provide a certain level of critical care capability to alleviate the shortage of beds in specialized ICUs, and to evaluate the prognosis of patients admitted to ICUs, thus offering better indicators for resource allocation in ICUs. ICUs play a significant role in the healthcare system, and due to the frailty of elderly patients’ health, they often require more intensive care. However, ICU resources are limited, and effectively allocating these resources to meet the needs of elderly patients has become an urgent issue [3]. Despite the development of various prognostic scoring systems based on large-scale patient data, predicting clinical outcomes for ICU patients remains a significant and challenging task [4]. Therefore, this study analyzes a large amount of medical data, sourced from MIMIC-III, which is currently the most renowned anonymized open-access database. It integrates comprehensive clinical data from patients admitted to Beth Israel Deaconess Medical Center (BIDMC) in Boston, Massachusetts, USA. This dataset provides detailed patient information, including laboratory test results, demographic characteristics, microbiological results, treatment processes, and fluid balances. The MIMIC-III dataset has been widely used in numerous studies and publications across different fields, including clinical prediction, disease diagnosis, and treatment. In addition to the traditional use of structured data, the existing literature also emphasizes the integration of text, visual, and verbal signals to enhance the effectiveness of clinical predictions for patients [5,6,7,8]. Our study utilizes the extensive clinical data stored in the MIMIC-III database to specifically investigate mortality prediction models for elderly patients in ICUs using machine learning models. This method includes constructing prediction models using machine learning techniques and analyzing large-scale datasets with Python 3.8. It evaluates the integration of structured data collected during ICU stays (including vital signs and test data) to predict mortality rates at different time points and provide better treatment plans for healthcare personnel. Additionally, redesigning smart healthcare management processes can improve the efficiency and convenience of medical services [9], helping clinical staff to more accurately assess patient risks, conduct patient risk assessments in a timely manner, improve treatment outcomes, and effectively utilize hospital resources.

This study references past research, such as Wang et al. [10], which developed and validated a machine learning model that can predict the occurrence of sepsis in ICU patients more accurately and earlier than currently used methods. Soares Pinheiro et al. [11] identified modifiable factors associated with higher ICU mortality rates, including age, patients transferred from hospital wards, sepsis diagnosis, and other clinical factors. These findings utilized high-frequency data analysis to enhance the prediction of ICU mortality and potentially help reduce mortality rates. Retrospective studies of high-frequency data in electronic medical records provide dynamic and interpretable machine learning predictions of ICU patient mortality. However, few studies have comprehensively integrated multiple factors to predict mortality. Predictive models using simple demographic information (such as age) and basic health indicators (such as vital signs) offer a resource-efficient and practical approach, as vital signs are non-invasive and are widely regarded by healthcare professionals as essential for assessing health status. While these variables can partially predict ICU patient mortality rates, most studies primarily adopt quantitative data. Existing statistical predictive models excel at manipulating quantitative variables, yet standardizing and incorporating unstructured textual data remain challenging [12]. This emphasizes the need for advanced methods to effectively integrate different types of data to ultimately improve mortality prediction and patient outcomes in ICU settings. Therefore, this study aims to evaluate whether the integration of clinical and socioeconomic data through machine learning can improve mortality prediction accuracy among elderly ICU patients. Specifically, short-term mortality within 3 days is frequently used in critical care research to capture acute deterioration and early intervention windows [13]. The 30-day period is a widely recognized benchmark for post-ICU outcome evaluation and hospital quality metrics, allowing for the assessment of the medium-term prognosis [14]. Accordingly, we investigate whether 3-day and 30-day mortality predictions using structured ICU data can enhance early risk assessment and support more informed, equitable healthcare resource allocation. Through this approach, this study not only contributes to predictive model development but also aligns with broader ESG (Environmental, Social, and Governance) concepts, highlighting the potential of data-driven systems to support sustainable and ethical healthcare delivery for aging populations. The research framework is shown in Figure 1 and Figure 2.

2. Literature Review

2.1. Prediction Mortality Model for Elderly Patients

Mortality prediction models for elderly patients are based on their relevant characteristics and data, such as past medical data, clinical records, physiological indicators, and imaging results. These models predict the risk of mortality for elderly patients over a certain period. Previous studies have explored various methods, model constructions, prediction indicators, and accuracies, analyzing the strengths and weaknesses of these prediction models in forecasting elderly patient mortality. Brajer N, Cozzi B et al. [15] conducted prospective and external evaluations of machine learning models predicting in-hospital mortality at the time of adult admission. Their ability to accurately predict in-hospital mortality at admission can improve clinical and operational decisions and outcomes. Prospective and multi-site retrospective evaluations of these machine learning models indicate that they can effectively distinguish in-hospital mortality among adult patients at admission. Data elements, methods, and patient selection enable the implementation of models at the system level. Machine learning models developed for predicting in-hospital mortality rarely apply broadly to all adult patients across an entire healthcare system and are easy to implement. Similarly, few projects are implemented and prospectively evaluated with external validation. Woodman RJ, Bryant K, et al. [16] improved the 12-month mortality risk prediction for elderly hospitalized patients using multiple prognostic index domain scores, clinical data, and machine learning. They primarily assessed whether the accuracy of predicting 12-month mortality improved using maximum likelihood estimation logistic regression (LR-MLE) with three types of MPIs (Multidimensional Prognostic Indices), along with age and gender. The results indicated that implementing risk scores based on MPI domains and clinical data using ML prediction models can support clinical decisions regarding the risk stratification of follow-up care for elderly hospitalized patients. In studies comparing the application of machine learning techniques in predicting elderly mortality [17], various machine learning techniques were compared using features from the “Healthy Aging Initiative” study to predict elderly mortality. They evaluated multiple machine learning techniques, including feature engineering, feature selection, data augmentation, and resampling, to improve mortality prediction accuracy. Time series analysis plays a crucial role in quantitative disease prediction, with common methods including trend extension algorithms, particularly the autoregressive integrated moving average (ARIMA) model. This model is a powerful tool for time series data prediction, especially suited for handling non-stationary time series due to random process characteristics [18].

Additionally, machine learning and deep learning have shown excellent results in predicting medical risks in electronic health systems. Using the BERT language model to convert surgery-related text into vectors, combined with deep neural networks to predict postoperative mortality, effectively identifies high-risk patients, reduces misdiagnosis, and promotes early intervention and the effective management of medical resources. Mortality prediction for elderly patients involves using data analysis and statistical methods to estimate the mortality risk that elderly populations may face after receiving medical treatment. As the trend of population aging increases, the increasing medical needs of the elderly population make predicting adverse outcomes for elderly patients particularly significant. The process of establishing prediction models includes (1) data collection and preprocessing, (2) feature selection and modeling, (3) model training and evaluation, (4) model optimization, and (5) model application and interpretation. Mortality prediction models for elderly patients help healthcare providers more accurately assess patient risks in clinical practice, offering customized and effective medical care [19]. Chi-Hsien Huang et al. [20] developed a machine learning-based prognostic index to predict all-cause mortality in community-dwelling elderly people. Through laboratory tests, they developed and validated the prognostic index using machine learning models to predict mortality in middle-aged and elderly individuals. They found that the MARBE-PI (Machine Learning-based Routine Blood Examination Prognostic Index) is the most suitable risk stratification measure in busy clinical settings, potentially identifying elderly individuals at higher risk of death and aiding clinical decision-making.

2.2. Research on Big Data Applications in Healthcare Management and ESG

The common definition of big data is the four Vs, which include Variety, Volume, Velocity, and Veracity. The application of big data in healthcare is primarily divided into three major areas: pharmaceutical research and development, patient diagnosis and treatment, and public health. Future developments in healthcare big data will focus on the application of machine learning and artificial intelligence [18]. Models established by big data aim to reduce the workload of healthcare professionals rather than replace or deprive them of their jobs, enabling them to make more accurate medical judgments more easily. Smart healthcare applications in clinical healthcare include outpatient and emergency services, inpatient services, community health (including long-term care), teaching, administrative management, and environmental management [21]. Big data has shown great potential in healthcare management and ESG (Environmental, Social, and Governance) fields.

In healthcare management, big data technology helps doctors develop personalized treatment plans, making more accurate diagnoses and treatments based on patients’ genes, lifestyles, and health conditions. Research indicates that analyzing big data can improve the diagnosis and treatment for elderly patients. By analyzing electronic health records (EHRs) and wearable device data, big data can accurately track the progression of chronic diseases, enhancing the effectiveness of personalized treatment plans. It also helps doctors identify potential health risks [22]. Big data is significant in public health policy formulation, supporting the prediction of disease risks and healthcare needs by analyzing health data of the elderly population, thereby optimizing resource allocation. Especially during epidemic outbreaks, big data can help quickly formulate prevention strategies to reduce health threats. Additionally, by analyzing large datasets, big data can predict the development trends of certain diseases, achieving early detection and treatment, effectively managing medical resources, reducing medical costs, and improving operational efficiency.

In the ESG field, systematic reviews have studied how big data is applied to elderly health research and proposed an ecological framework. Big data is used as a data analysis tool, a decision support tool, and a modeling support tool in the environment and health interaction domain. Research emphasizes that big data can improve the efficiency of social resource allocation and promote the formulation of public health policies [21]. Other studies have reviewed how big data helps healthcare institutions improve their ESG performance. The higher the concern of enterprises about ESG-related issues, the higher their ESG scores (e.g., Refinitiv and RepRisk ratings), which are associated with lower environmental and social risks. Institutional investors are also more inclined to invest in companies with good ESG performance [23]. Xinyue Zhang, Xiaolu Gao et al. [21] explored the application of big data in public health policies, particularly in the sustainability challenges of the elderly population. Through big data technology, a better understanding of the health needs of the elderly, the optimization of resource allocation, and the promotion of healthy aging strategies can be achieved. These studies highlight the potential of big data in improving the ESG performance of healthcare services for the elderly, particularly in enhancing environmental sustainability, reducing social inequalities, and optimizing governance structures. Big data is used for environmental monitoring and is capable of analyzing and monitoring environmental changes, helping enterprises and governments formulate more effective environmental protection policies. It can also perform social impact assessments, ensuring corporate activities conform to social responsibilities, providing insights, and improving social impacts. In governance, big data increases corporate governance transparency, helping regulatory agencies and investors better understand the operations of companies. These applications demonstrate the significant value of big data in improving healthcare quality, promoting sustainable development, and advancing social responsibility.

2.3. Health Management and Treatment Plans for Elderly Patients

Managed Care is a concept that was first introduced in the United States. It involves healthcare insurance institutions systematically managing the health of their insured clients to effectively control the occurrence or development of diseases, reduce the rate of claims and actual medical expenses, and ultimately decrease losses from insurance payouts. Health management emphasizes the proactive prevention and early detection of potential health issues. The implementation of health management encompasses six cyclical processes: monitoring, evaluation, analysis, prediction, planning, and implementation [24]. Related research has expanded to optimize the allocation and management of medical resources to improve the efficiency and quality of healthcare services. In recent years, due to the scarcity of medical resources and high healthcare costs, there has been increasing global attention on the allocation of medical resources. Inefficiencies in resource allocation and healthcare operations remain common, such as shortages of medical professionals, the underutilization of personnel in value-added services, and inappropriate scheduling. Due to the high costs and limited supply of medical resources, decision-makers in the healthcare industry must enhance operational efficiency. One of their biggest challenges is how to effectively allocate limited resources to provide the highest quality of healthcare services [25]. Healthcare resource allocation mechanisms typically include expert analysis (such as technical assessments) and social participation to determine which services should be prioritized, removed, or restricted (e.g., establishing clinical guidelines). The goal of prioritizing healthcare resource allocation is to maximize population health while minimizing the opportunity costs associated with the delivery of healthcare services [26].

Regarding health management and treatment plans for elderly patients, research indicates that various factors need to be considered, including personal health status, social support, and economic conditions. These plans should emphasize personalized health management, covering medication treatment, lifestyle management, and rehabilitation programs. By integrating these factors, healthcare institutions can better provide comprehensive and targeted care for elderly patients, improving their quality of life and treatment outcomes. Prior studies have analyzed extensive personal health information collected from elderly patients, establishing quantitative relationships between genetic, environmental, lifestyle, and dietary factors and overall health status. This information helps predict the likelihood of certain specific diseases occurring or causing death within a certain period. Based on these predictions, health management and treatment plans are developed, including medication treatment, lifestyle management, and rehabilitation programs, offering targeted control and intervention. This approach helps governments, organizations, insurance companies, and individuals achieve maximum health benefits at minimal costs [20].

2.4. The Application of Machine Learning Methods in the Medical Field and Management

The integration of machine learning in healthcare has made significant progress, fundamentally transforming the way medical diagnosis, treatment, overall patient care, and management are conducted. By analyzing vast amounts of medical data, machine learning algorithms can automatically identify hidden patterns, assisting doctors in making faster and more accurate diagnoses, such as analyzing imaging diagnostics like X-rays or MRI images. In terms of treatment, machine learning provides personalized treatment recommendations by analyzing patients’ historical data, genetic characteristics, and health conditions, tailoring the most suitable treatment plans. Moreover, machine learning plays a crucial role in drug development, accelerating the discovery and development of new drugs, and predicting their effects and side effects. Machine learning is also used in managing chronic disease patients by monitoring and analyzing health data to provide timely warnings and suggestions. In the healthcare insurance field, it helps in analyzing patients’ health data to assess risks and formulate reasonable insurance plans. Additionally, machine learning is utilized to predict disease progression, enabling healthcare professionals to identify potential health risks early and intervene proactively, thereby improving patient outcomes. It plays a key role in optimizing the allocation of medical resources by predicting hospital bed occupancy rates, medication needs, and staffing, thus reducing resource waste and increasing operational efficiency in line with management efficiency and ESG principles [27]. In recent years, machine learning has become an indispensable method in data analysis for building predictive models. Its ability to effectively extract data features and overcome the limitations of assumptions in traditional statistical models has led to its successful application in various prediction problems in healthcare. For example, the random forest method has been used to assess fetal maturity, adjusting model parameters across different time scales to extend previous concepts based on heart rate and univariate complexity indices [28]. Another study developed an ICU mortality prediction model using the SICULA (Super ICU Learner Algorithm) and compared its predictive performance across multiple MIMIC-II datasets [29]. Additionally, research has shown that increasing the sampling frequency of blood pressure can significantly improve algorithm prediction performance, especially in predicting ICU hypotension events using machine learning techniques [30]. Poornima D. et al. [31] explored the application of machine learning techniques in real-time healthcare management, covering aspects such as resource allocation, patient monitoring, and medical decision-making, emphasizing the impact of machine learning on optimizing medical resources and improving patient management efficiency. Machine learning enhances healthcare service efficiency through automated diagnosis and predictive models, including the early diagnosis of cardiovascular diseases, cancer, and chronic diseases, and improving medical record analysis methods [32].

Previous reviews have highlighted various machine learning algorithms applied in clinical diagnosis, personalized treatment, and health prediction. The effectiveness of deep learning in image analysis and disease risk prediction underscores the importance of data quality and algorithm interpretability [33]. Supervised and unsupervised learning algorithms in healthcare applications, including time series analysis and health data processing, show that machine learning can enhance data analysis accuracy, contributing to the development of personalized healthcare [34]. Overall, the application of machine learning not only improves the accuracy and efficiency of healthcare services but also promotes more personalized, preventive, and precise patient care, significantly enhancing healthcare management efficiency. The application of machine learning in smart healthcare and management has become a key driver of technological innovation. In smart healthcare, machine learning techniques have significantly improved the precision of diagnosis and treatment. By analyzing medical images, genomic data, and EHRs, machine learning models can detect diseases early, predict disease progression, and support the development of personalized treatment plans. For example, breakthroughs have been made in the application of deep learning in cancer screening and chronic disease management. In management, machine learning helps organizations optimize resource utilization through data-driven insights, aiding smart healthcare, sustainable development, and efficient management.

3. Methodology

3.1. Dataset from MIMIC-III

This study used clinical data from the BIDMC in Boston, Massachusetts, sourced from MIMIC-III. MIMIC-III is one of the most well-known anonymized open-access databases, integrating comprehensive clinical data from BIDMC’s admitted patients, covering various types of ICU data from 2001 to 2012. This includes patients’ vital signs, medications, lab data, and observation records [13]. The MIMIC-III database includes 46,520 different patients with a total of 58,976 ICU admissions. Most patients are adults, with males accounting for 55.9%, a median age of 65.8 years, a median hospital stay of 6.9 days, and a median ICU stay of 2.1 days. The mortality rate during ICU stays is 8.5%. On average, each ICU admission generates a vast amount of data, including 6643 patient observation records, 83 clinical document records, and 559 laboratory test results. A co-author of this paper, Te-Nien Chien, completed the NIH’s online course, passed the human subjects protection exam, and obtained usage permission (certificate number: 35628530), securing ethical approval to access the MIMIC-III database, ensuring all data used in this study are appropriately safeguarded.

3.2. Data Preprocessing and Variable Selection

This study adopted an empirical research methodology, first collecting relevant data from elderly patients in ICUs, including physiological monitoring data and clinical assessment data. Based on the relevant literature [35], we selected eight easily collectable patient variables from the MIMIC-III database’s Admission Tables, Chartevents Tables, and Labevents Tables. These variables include the Glasgow Coma Scale, heart rate, respiratory rate, oxygen saturation, systolic blood pressure, diastolic blood pressure, body temperature, and age. We then used statistical analysis methods to analyze and model the data, establishing a mortality prediction model for patients. The data preprocessing adopted a three-stage method to handle missing values. First, patients with more than 30% missing data were excluded. Second, data with more than 40% missing in predictive variables were removed. Lastly, records with more than 20% missing data were excluded based on the above criteria, and the remaining missing values were filled using the mean value. To underscore the early predictive capacity, input to the predictive model was derived from data collected within the first 24 h of patient admission to the ICU. We adopted basic imputation methods (mean and median imputation). Additionally, to address the challenges posed by imbalanced datasets—an issue frequently encountered in medical data—we employed the Synthetic Minority Over-sampling Technique (SMOTE). By generating synthetic samples for the minority class, the SMOTE helps correct class distribution imbalances and improve model performance. In our study, we experimented with various SMOTE ratios to better mitigate the impact of class imbalance and enhance the robustness of our predictions [36].

After the model was established, we validated it using a separate validation dataset to evaluate its predictive accuracy and performance. Finally, based on the study results, we discussed and analyzed the findings, proposing improvement strategies and management recommendations. This comprehensive research method aims to help medical institutions more effectively manage treatment and resource allocation for elderly patients through accurate mortality prediction models.

3.3. Machine Learning Prediction Models and Evaluation Metrics

The machine learning methods used in this study included Adaptive Boosting (Adaboost), Bagging, Catboost, Gradient Boosting (GB), and Support Vector Classifier (SVC). These five classic classification algorithms were employed to elucidate the impact of clinical annotations and pathology reports on clinical outcome predictions. Adaboost is an ensemble learning method that continuously trains a series of weak classifiers through iterative computation and adjusts the weight of each sample based on the performance of the previous classifier. This method focuses on samples misclassified by the previous classifier, thereby improving the overall model’s prediction accuracy. Bagging, short for Bootstrap Aggregating, is an ensemble learning algorithm proposed by Leo Breiman in 1994. It involves randomly sampling multiple subsets from the training data, training an independent classifier on each subset, and then combining these classifiers’ predictions through voting or averaging to obtain the final prediction. Bagging can be combined with other classification and regression algorithms to improve accuracy and stability, reduce result variance, and avoid overfitting. It has been used in various microarray studies [37]. Catboost is a variant of the Gradient Boosting algorithm optimized for categorical features, offering better predictive performance and faster training speed. Gradient Boosting is also an ensemble learning method that trains a series of weak classifiers iteratively and adjusts each classifier’s weight based on the previous round’s errors to improve prediction accuracy. This method uses the negative gradient information of the model’s loss function to train poorly performing models and integrates the trained results cumulatively into the existing model [38]. This study used the scikit-learn library to implement Gradient Boosting, setting the maximum number of iterations to 100, with other hyperparameters using the default values from the scikit-learn library. SVC is a supervised learning algorithm that finds an optimal hyperplane in a high-dimensional feature space to maximize the margin between sample classes, thereby classifying the samples. These algorithms were applied in this study to predict clinical outcomes of elderly patients, comparing their prediction performance and applicability. All data mining tasks were completed using the Python programming language, leveraging its rich machine learning libraries and tools for data analysis and modeling.

To comprehensively compare the impact of integrating structured and unstructured data on predicting ICU patient mortality, five different metrics were chosen as evaluation tools for modeling in this study. These metrics included the AUROC (Area Under the Receiver Operating Characteristic Curve), precision, recall, the F1-Score, and accuracy. The AUROC is a metric used to evaluate the overall performance of a classifier. The ROC curve illustrates the relationship between the true positive rate (TPR) and the false positive rate (FPR) at various classification thresholds, providing a comprehensive view of the classifier’s capability. The AUROC value ranges from 0 to 1, with higher values indicating better performance. Notably, the AUC calculation assigns equal weight to all instances, regardless of the nature of the positive label. Precision refers to the proportion of true positive samples out of all the positive predictions made. Recall, also known as sensitivity, measures how well a model identifies actual positive cases. It is the proportion of true positives correctly predicted out of all actual positives. Sensitivity is crucial in medical diagnosis, as it helps assess a model’s ability to detect diseases or abnormalities, making it essential for reliable and effective clinical applications. The F1-Score is the harmonic mean of precision and recall. This metric is most commonly used to measure precision and is typically used to determine the accuracy of algorithms. Accuracy measures the proportion of correctly predicted samples by the classifier, representing the ratio of correctly classified samples to the total number of samples. It is used to evaluate the performance of a model in disease screening or diagnosis.

This study evaluated the impact of combining quantitative and textual data on the prediction of mortality in elderly ICU patients using five widely recognized binary classification evaluation metrics. These metrics are key indicators of model performance and precision:

Precision (PPV): In predictive analytics, precision denotes the accuracy of positive predictions. It is calculated as the ratio of true positives (TPs) to the sum of true positives (TPs) and false positives (FPs). Sokolova, M. et al. [39] provides a comprehensive explanation of how precision is defined and its role in evaluating model performance.

$$Precision=PPV={\displaystyle \frac{TP}{TP+FP}}$$

(1)

Recall (TPR): Recall quantifies the model’s ability to correctly identify positive cases. It is calculated as the ratio of true positives (TPs) to the sum of true positives (TPs) and false negatives (FNs). Chicco, D. et al. [40] discuss the various performance metrics used in machine learning, including recall, and evaluate their effectiveness in binary classification tasks.

$$Recall=TPR={\displaystyle \frac{TP}{TP+FN}}$$

(2)

F1-score: This is a comprehensive metric combining precision and recall, calculated as the harmonic mean of the two, offering a balanced measure of overall performance. Sokolova, M. et al. [39] provide a systematic analysis of various performance metrics for classification tasks and discuss their usage in machine learning evaluations.

$$F1-score={\displaystyle \frac{2\ast Precision\ast Recall }{Precision+Recall }}$$

(3)

Accuracy: Accuracy assesses the overall correctness of predictions and is calculated as the ratio of the sum of true positives (TPs) and true negatives (TNs) to the total number of samples in the test dataset. Powers, D. M. W. [41] provides a comprehensive discussion of various evaluation metrics used in machine learning, including accuracy, and highlights its significance in classification model evaluation.

$$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}}$$

(4)

AUROC: The AUROC is a common diagnostic metric that provides a comprehensive measure of classifier performance by plotting the relationship between the true positive rate (TPR) and the false positive rate (FPR) across various classification thresholds.

By employing the above evaluation metrics, this study aims to provide a comprehensive and standardized assessment of predictive models, specifically in the context of mortality prediction for elderly ICU patients.

4. Result

4.1. Descriptive Statistical Analysis

This study applied descriptive statistical analysis and statistical significance testing to examine the sociodemographic data of ICU patients, assessing how factors such as gender, age, insurance type, marital status, and ethnicity influence mortality rates. By analyzing the mortality rates of different demographic groups, we gained insights into the extent to which these characteristics affect patient survival, providing important references for subsequent healthcare management and preventive measures.

Table 1. Sociodemographic data of elderly patients in intensive care units and the relationship between in-hospital mortality and survival.

	Overall	Dead at ICU	Alive at ICU	p-Value 1
General	23,517 (100%)	3327 (14%)	20,192 (86%)	0.21
Gender (Male)	12,721 (54%)	1751 (14%)	10,970 (86%)
Gender (Female)	10,798 (46%)	1576 (15%)	9222 (85%)
Age				<0.001 **
Age (65–74 years old)	9820 (42%)	1072 (11%)	8748 (89%)
Age (75–84 years old)	10,477 (45%)	1626 (16%)	8851 (84%)
Age (Over 85 years old)	3220 (14%)	629 (20%)	2591 (80%)
Insurance				<0.05 *
Medicaid	315 (1%)	41 (13%)	274 (87%)
Private	2233 (9%)	299 (13%)	1934 (87%)
Medicare	20,897 (89%)	2971 (14%)	17,926 (86%)
Government	46 (0%)	9 (20%)	37 (80%)
Self-Pay	26 (0%)	7 (27%)	19 (73%)
Marital Status				<0.05 *
Separated	1428 (6%)	176 (12%)	1252 (88%)
Single	3431 (15%)	456 (13%)	2975 (87%)
Married	12,350 (53%)	1654 (13%)	10,696 (87%)
Widowed	5236 (22%)	731 (14%)	4505 (86%)
Ethnicity				<0.05 *
Asian	521 (2%)	91 (17%)	430 (83%)
Black	4400 (19%)	727 (17%)	3673 (83%)
White	18,113 (77%)	2443 (13%)	15,670 (87%)
Multi Race Ethnicity	483 (2%)	66 (14%)	417 (86%)

¹ Chi-square test. p-value thresholds: * p < 0.05 indicates statistical significance; ** p < 0.001 represents highly significant results.

presents the descriptive statistical analysis of elderly patients included in this study. Figure 3 provides visual comparisons of demographic characteristics (gender and age), facilitating the interpretation of distribution patterns and revealing potential risk differentials relevant to ICU outcomes.

Understanding the sociodemographic data of ICU patients is crucial for identifying factors that influence in-hospital mortality and survival rates. By analyzing key variables such as gender, age, insurance type, marital status, and ethnicity, we can reveal differences and patterns in patient outcomes, thus identifying areas that require optimization for specific population groups. Based on descriptive statistical analysis and statistical significance testing, elderly patients in ICUs show significant demographic differences. Male patients account for 54%, slightly higher than the 46% of female patients, but the in-hospital mortality rates between genders show no statistically significant difference (p = 0.21), with mortality rates of 14% for males and 15% for females. This suggests that gender has a similar impact on mortality risk, although further research is needed. In terms of age, mortality rates increase significantly with age (p < 0.001), especially among patients over 85 years old, who have the highest mortality rate at 20%. Correspondingly, survival rates decrease with increasing age, with the 65–74 age group having a survival rate of 89%, while the survival rate for those over 85 years old drops to 80%. These data indicate that age is a critical factor affecting the prognosis of ICU patients, especially for elderly patients, who require closer monitoring and care. The insurance type also demonstrates a statistically significant association with mortality rates (p < 0.05). Patients with self-pay and government insurance show relatively higher mortality rates, particularly those with government insurance, which may be related to unequal medical resource allocation or other structural issues that warrant further investigation. Additionally, marital status presents significant differences in mortality rates (p < 0.05), with widowed patients showing higher mortality compared to married individuals. This may reflect the negative impact of lacking social support on patient health outcomes, indicating that these patients may need additional social and psychological support. It is noteworthy that a higher proportion of patients over 85 years old use government insurance, reflecting that this age group may be more reliant on government-provided medical coverage. This table presents ICU mortality and survival data disaggregated by four ethnic groups: Asian, Black, White, and Multi-Race. Notably, the White group demonstrates the highest survival rate (87%) and the lowest mortality rate (13%), marking a significant contrast compared to the other groups. Both the Asian and Black groups exhibit identical ICU mortality rates (17%), suggesting a relatively elevated risk profile in terms of critical care outcomes. The Multi-Race group shows a slightly lower mortality rate (14%) than the Asian and Black groups, though still above that of the White group. These findings highlight potential disparities among ethnic groups in ICU outcomes, which may be influenced by a combination of factors such as underlying health conditions, the timing of ICU admission, access to critical care, and broader socioeconomic determinants. The results warrant further investigation into the structural causes of health inequities and may serve as a foundation for future studies on racial and ethnic disparities in critical care settings.

These results highlight the differences in mortality rates across different characteristics, suggesting the need for further research to develop more effective care strategies to address the specific needs of elderly patients. Overall, as age increases, the risk of mortality significantly rises (p < 0.001), and survival rates gradually decline, underscoring the need for more vigilant monitoring and personalized treatment measures for elderly patients in ICUs.

4.2. Machine Learning Modeling for Mortality Analysis in Elderly ICU Patients

This study applied five machine learning algorithms—Adaboost, Bagging, Catboost, GB, and SVC—to build prediction models. These models use machine learning methods to model and predict the mortality rates of elderly patients in ICUs. By comparing the accuracy and effectiveness of these models, we provide tools to assess the mortality risk of elderly patients. This will help medical institutions implement more targeted treatment and care measures for high-risk patients, thereby improving the quality and efficiency of healthcare services.

This study used ten-fold cross-validation to construct various machine learning models based on data collected within 24 h of admission for elderly patients in ICUs, predicting the risk of mortality within three days and thirty days. The models were constructed using five classic classification algorithms, Adaboost, Bagging, Catboost, GB, and SVC, with a detailed analysis and comparison of mortality predictions at different time periods to predict elderly patients’ mortality rates. According to this study’s results, the performance metrics of five different machine learning methods and how these predict mortality within three days and thirty days after admission were analyzed. For the three-day mortality group, the Bagging model had the highest AUROC value (0.7981), followed by SVC (0.7856). The AUROC measures the performance of the model across different classification thresholds, with values closer to 1 indicating better performance. In terms of accuracy, the Catboost model performed best (0.9061), followed by Adaboost (0.8323). Accuracy indicates the proportion of correctly classified samples out of the total number of samples, with higher values indicating higher overall predictive accuracy. The study results show that the Catboost model performs well in accurately classifying patients who die within three days. For the thirty-day mortality prediction, the Bagging model again had the highest AUROC value (0.7424), followed by Adaboost (0.7354). Compared to the three-day mortality predictions, the models’ performance generally declined, but Bagging still performed well. In terms of accuracy, the Catboost model performed best (0.8147), followed by GB (0.7596). Overall, the Bagging model performed relatively well in terms of AUROC and recall, especially in predicting mortality within three days after admission, making it suitable for identifying patients at high risk of death. The Catboost model performed best in terms of accuracy, particularly in predicting three-day mortality. The study results indicate this model’s advantage in accuracy, especially in short-term mortality risk prediction, effectively reducing false positives. Choosing the most suitable model for specific situations requires considering both the AUROC and accuracy comprehensively. Please refer to Appendix A for the parameters of the five machine learning models. Specific data and analysis results can be found in

Table 2. Mortality prediction of elderly ICU patients at different time periods using various machine learning methods.

	Adaboost	Bagging	Catboost	GB	SVC
3 Days
AUROC	0.7773	0.7981	0.7022	0.7748	0.7856
Precision	0.1683	0.1567	0.2706	0.1667	0.1451
Recall	0.7170	0.7916	0.3175	0.7132	0.7827
F1	0.2727	0.2615	0.2920	0.2702	0.2448
Accuracy	0.8323	0.8040	0.9061	0.8310	0.7882
30 Days
AUROC	0.7354	0.7424	0.7185	0.7315	0.7314
Precision	0.3372	0.3232	0.4014	0.3368	0.3312
Recall	0.7029	0.7490	0.5834	0.6920	0.7005
F1	0.4555	0.4511	0.4751	0.4528	0.4496
Accuracy	0.7584	0.7377	0.8147	0.7596	0.7534

and Figure 4. In Figure 4, the red dashed line is the baseline, it represents the random guessing line. If a model’s ROC curve is close to this line, it means the model performs no better than random guessing.

5. Discission and Conclusions

5.1. Research Findings

As the population ages, more elderly patients are entering ICUs, posing challenges for healthcare resource management. Elderly patients have complex and unstable conditions with higher mortality rates [42,43]. The timely prediction of mortality risk and personalized treatment are crucial for improving survival rates and optimizing resource utilization [44]. Smart healthcare technology emphasizes the use of data-driven techniques to enhance the efficiency, precision, and equity of medical services [45]. This study used big data and machine learning to develop a mortality prediction model for elderly ICU patients and analyzed the impact of demographic factors on mortality, aiding healthcare institutions in managing resources effectively and improving patient care quality [46].

This study used the MIMIC-III database and applied five machine learning algorithms (Adaboost, Bagging, Catboost, GB, and SVC) to predict the 3-day and 30-day mortality rates of elderly ICU patients. The Bagging model achieved an accuracy of 0.8040, demonstrating its potential in mortality prediction. This study also found that age, insurance type, and marital status are key factors affecting mortality rates in elderly ICU patients. The older the patient, especially those over 85, the higher the mortality rate [1]; patients with private insurance had lower mortality rates, possibly due to economic conditions and healthcare accessibility [47]; and divorced, single, and widowed individuals exhibited a higher combined mortality rate compared to married individuals, suggesting that marital status may be an important factor influencing survival in ICU settings.

The prediction model developed in this study not only helps accurately predict the mortality risk of elderly ICU patients but also aids healthcare institutions in allocating resources more effectively and reducing waste [15]. By accurately predicting mortality, healthcare professionals can identify high-risk elderly patients early and adjust treatment plans accordingly, improving the precision and efficiency of clinical care. This model enhances decision-making, especially in resource-limited situations, helps optimize resource distribution, and improves the quality of care for elderly patients, further advancing health management and clinical practice.

5.2. Research Innovation and Contribution

With the rapid advancement of information technology and machine learning, smart healthcare is significantly improving medical services for elderly patients. This study focused on the application of smart healthcare technologies in ICUs, particularly the value of machine learning models in predicting clinical risks for elderly patients [48,49]. By analyzing electronic health records and clinical data, the research highlights the significant impact of social factors such as insurance type and marital status on mortality rates [34,44], offering innovative tools for the early identification of high-risk patients and resource allocation. In terms of resource management, smart healthcare technologies optimize ICU resource distribution, ensuring appropriate care for elderly patients, which is crucial in addressing the growing medical demands of an aging population. Additionally, machine learning enhances healthcare efficiency and cost-effectiveness, providing essential support for precise policymaking and operational management by healthcare institutions [46,47]. These technologies accurately predict disease progression and create personalized treatment plans, significantly reducing mortality risks and improving clinical outcomes. Furthermore, expert support systems analyze multidimensional data to provide precise recommendations for medical decision-making, helping healthcare professionals effectively address the complex needs of elderly patients [1,33]. By advancing care quality and resource rationalization, smart healthcare technologies are paving the way for a more effective and sustainable healthcare ecosystem tailored to the needs of an aging society.

Smart healthcare technologies also align with ESG principles by improving operational efficiency, promoting equitable access to medical resources, and enhancing governance transparency. By optimizing healthcare delivery, reducing inefficiencies, and facilitating data-driven decision-making, these technologies contribute to a more sustainable healthcare model that supports both environmental and social objectives. Machine learning assists in modernizing governance frameworks, increasing institutional efficiency, and strengthening resource management strategies [15,42,50]. Recent findings from PatientView’s 2023/2024 survey further reinforce the significance of ESG considerations from the patient perspective. The study, which collected responses from 701 patient advocacy groups across 74 countries, underscores the increasing importance of ESG issues in healthcare decision-making. Specifically, 83% prioritize companies actively addressing social issues, emphasizing the expectation of long-term engagement and transparent communication [51]. These technologies support flexible strategies, continuously improving services for elderly patients, ultimately promoting sustainable and equitable healthcare development, and achieving comprehensive improvements in health services and social value.

By incorporating ESG-related discussions into healthcare analytics, this study contributes to ongoing discourse on responsible healthcare innovation and provides a foundation for future research to further explore the relationship between ESG principles and predictive healthcare models.

5.3. Management Implications

Since the implementation of the National Health Insurance program in Taiwan in 1995, the healthcare system has covered over 99% of the population, significantly increasing healthcare service utilization. Against this backdrop, the ICU has become one of the most resource-intensive areas of healthcare. Based on an analysis of the U.S. BIDMC database, this study demonstrates the significant potential of machine learning in predicting mortality rates for elderly patients, particularly in relation to factors such as marital status and insurance type. Over the past few decades, the proportion of ICUs in the U.S. and Taiwan has grown significantly, with the number of ICU beds in both countries far exceeding that in other Western countries [52]. Studies show that ICU admission and mortality rates in Taiwan are similar to those in North America and Europe for non-selected patients, with the ICU bed ratio per hundred thousand people comparable to that in the U.S. [53]. Management in this study refers to the optimization of healthcare resource allocation and policy formulation, particularly in relation to elderly ICU patients. Based on an analysis of the U.S. BIDMC database, this study highlights the potential of machine learning in supporting mortality prediction and guiding healthcare institutions in improving treatment efficiency. Given the similarities between ICU conditions in Taiwan and the U.S., these findings provide valuable references for Taiwanese medical institutions to refine local strategies for clinical care. Healthcare providers should leverage these results to enhance the precision of medical interventions, while policymakers should focus on promoting the equitable distribution of healthcare resources to improve overall service quality.

This study provides significant insights into mortality prediction and resource management for elderly ICU patients. However, some limitations need to be addressed. Future research should expand the data scope, including more cross-cultural samples to increase the model’s generalizability. Additionally, the deep application of unstructured data, such as clinical notes and medical imaging, needs further exploration to enhance the model’s accuracy and comprehensiveness. To validate the model’s stability, it should be tested in different regions and healthcare systems, coupled with advanced deep learning techniques to further improve its clinical applicability. Moreover, future studies should consider the ethical issues and data privacy of smart healthcare technologies, ensuring their legality and safety. Ultimately, by collaborating with healthcare institutions, research findings can be translated into practical applications to optimize resource allocation strategies and improve healthcare accessibility for disadvantaged groups.

In conclusion, this study emphasizes the potential of smart healthcare technologies in predicting mortality rates among elderly patients and optimizing medical resource allocation. With the intensification of population aging, these technologies can effectively support clinical decision-making, enhance healthcare efficiency, and promote fairness. By integrating ESG principles, smart healthcare not only improves the quality of medical services for elderly patients but also promotes the sustainable use of resources, addressing the challenges posed by global aging. Therefore, future research should focus on technological innovation and clinical needs, combining big data analytics with medical expertise to advance the sustainable development of healthcare systems and achieve more accurate and equitable medical services.