ORIGINAL RESEARCH article

Front. Public Health, 15 May 2025

Sec. Health Economics

Volume 13 - 2025 | https://doi.org/10.3389/fpubh.2025.1588270

Economic implications of artificial intelligence-driven recommended systems in healthcare: a focus on neurological disorders


Jing ZhangJing Zhang1Shihui Xiang
Shihui Xiang2*Li LiLi Li3
  • 1School of Law, Panzhihua University, Sichuan, Chongqing, China
  • 2Shenzhen Jizhi Laser Technology Co., Ltd, Guangdong, Shenzhen, China
  • 3Taizhou Vocation College of Science and Technology, School of Accounting and Finance, Taizhou, Zhejiang, China

Introduction: The rapid advancement of Artificial Intelligence (AI)-driven recommendation systems in healthcare presents significant economic implications, particularly in the context of neurological disorders. These systems offer opportunities to enhance diagnostic accuracy, optimize resource allocation, and improve patient outcomes. However, conventional economic models fail to address the dynamic complexities of AI integration in healthcare, including market inefficiencies and stakeholder behaviors.

Methods: To bridge this gap, we propose a Dynamic Equilibrium Model for Health Economics (DEHE), incorporating reinforcement learning and stochastic optimization. This model captures uncertainty in healthcare decision-making and includes dynamic pricing, behavioral incentives, and adaptive insurance premium mechanisms.

Results: Our experimental results demonstrate that DEHE improves economic efficiency by optimizing AI-driven recommendations while balancing healthcare cost and accessibility. Through multi-agent simulations, the model shows strong real-world applicability and stability. It effectively addresses asymmetric information, moral hazard, and market dynamics.

Discussion: This study offers a novel economic framework for integrating AI-driven systems in neurological healthcare. We recommend the adoption of adaptive policy mechanisms and stakeholder-specific incentives to enhance cost-effectiveness and equitable access. These insights contribute to the development of more sustainable and inclusive AI-based healthcare policies.

Introduction

The increasing burden of neurological disorders, such as Alzheimer's disease, Parkinson's disease, and epilepsy, poses significant economic challenges due to the high costs of diagnosis, treatment, and long-term care (1). Traditional healthcare models often struggle with inefficiencies, including delayed diagnoses, suboptimal treatment plans, and high hospitalization rates, which further exacerbate financial strain on both healthcare systems and patients (2). Artificial Intelligence (AI)-driven recommendation systems offer a transformative solution by optimizing clinical decision-making, personalizing treatment strategies, and improving early disease detection (3). Not only can these systems enhance patient outcomes through tailored interventions, but they can also significantly reduce healthcare expenditures by minimizing unnecessary tests and hospital admissions (4). AI-powered models facilitate resource allocation, enabling healthcare providers to prioritize high-risk patients and allocate medical resources efficiently. Given the rising prevalence of neurological conditions and the escalating costs associated with their management, AI-driven recommendation systems present a crucial opportunity to balance cost-effectiveness with improved healthcare delivery (5). Therefore, understanding their economic implications is essential for stakeholders, including policymakers, healthcare providers, and insurers, to ensure their integration aligns with financial sustainability and equitable access to care (6).

To address the limitations of conventional diagnostic and treatment approaches, early AI-driven recommended systems were rooted in symbolic AI and expert systems, which relied heavily on rule-based decision-making and knowledge representation (7). These systems used structured medical ontologies and manually curated knowledge bases to assist physicians in diagnosing and managing neurological disorders. For example, rule-based decision trees and logic-based inference engines helped in clinical decision support by mapping patient symptoms to predefined disease patterns (8). However, these approaches suffered from scalability issues, as they required extensive human expertise to construct and update medical knowledge bases. Moreover, they struggled with handling ambiguous or incomplete data, limiting their applicability in real-world clinical settings (9). Despite their interpretability and ability to provide transparent recommendations, their rigid nature hindered adaptability to new clinical insights and patient variability. The economic impact of these systems was modest, as they primarily aimed to improve diagnostic accuracy rather than significantly reduce costs or optimize resource allocation (10). Consequently, while symbolic AI laid the groundwork for AI-driven healthcare, it proved insufficient in addressing the complexity and dynamic nature of neurological disorders.

To overcome the limitations of symbolic AI, data-driven machine learning models emerged as a more flexible and scalable alternative for AI-driven recommendation systems in healthcare (11). These models leveraged large-scale patient datasets, electronic health records (EHRs), and imaging data to develop predictive models for disease progression and treatment outcomes. Supervised learning techniques, such as support vector machines (SVMs) and random forests, were widely used for diagnosing conditions like Alzheimer's and Parkinson's based on imaging biomarkers and clinical assessments (12). Clustering algorithms and Bayesian networks facilitated risk stratification, allowing healthcare providers to identify high-risk patients and tailor interventions accordingly. The economic benefits of these machine learning-based recommendation systems were more pronounced than those of symbolic AI, as they enabled early disease detection, reducing hospitalization rates and lowering long-term treatment costs (13). However, these models still faced challenges in generalizability due to dataset biases, requiring continuous retraining and validation across diverse patient populations. The lack of interpretability in many machine learning models raised concerns about their adoption in clinical practice, particularly in high-stakes medical decision-making where transparency is crucial. Although these models improved cost efficiency to some extent, their full economic potential was yet to be realized due to constraints in real-world deployment and integration with existing healthcare infrastructure (14).

Building upon the advancements in machine learning, deep learning and pre-trained models have revolutionized AI-driven recommendation systems in healthcare, particularly in the management of neurological disorders. Deep learning architectures, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have demonstrated remarkable success in analyzing complex medical imaging data, identifying early-stage disease markers, and predicting patient trajectories with high accuracy (15). Pre-trained transformer models, such as BERT(Bidirectional Encoder Representations from Transformers)-based medical NLP models, have enhanced clinical decision support by extracting valuable insights from unstructured EHRs, physician notes, and medical literature. These advanced AI models have significantly improved diagnostic precision and personalized treatment recommendations, reducing misdiagnosis rates and optimizing therapeutic interventions (16). Economically, deep learning-powered recommendation systems offer substantial cost savings by minimizing unnecessary procedures, streamlining workflow automation, and enhancing operational efficiency in healthcare facilities. By integrating multi-modal patient data, these models enable precision medicine approaches that maximize treatment efficacy while minimizing adverse effects, ultimately reducing long-term healthcare expenditures (17). Despite these advantages, deep learning models require extensive computational resources and vast amounts of high-quality labeled data, making their widespread adoption challenging for resource-constrained healthcare settings. Moreover, regulatory and ethical concerns regarding patient data privacy and AI accountability must be addressed to ensure sustainable implementation in clinical practice (18).

Based on the limitations of earlier AI approaches–such as the rigidity of symbolic AI, the generalizability issues of machine learning models, and the computational demands of deep learning–our proposed method aims to strike a balance between cost-effectiveness, interpretability, and clinical utility. We introduce a hybrid AI-driven recommendation system that combines knowledge-based reasoning with deep learning-powered predictive analytics to provide accurate, interpretable, and economically viable recommendations for neurological disorder management. By leveraging domain-specific medical knowledge alongside patient-specific data-driven insights, our model ensures robust decision support while maintaining transparency and trustworthiness in clinical practice. Our approach incorporates federated learning to enhance data privacy and model generalizability without compromising sensitive patient information. From an economic standpoint, our system is designed to optimize resource allocation, minimize unnecessary interventions, and improve healthcare affordability through cost-effective AI integration. By addressing the economic challenges associated with neurological disorder management, our method paves the way for sustainable AI adoption in healthcare.

We summarize our contributions as follows:

• Our method integrates symbolic AI with deep learning to provide interpretable yet highly accurate recommendations, ensuring trust in AI-driven decision-making while maintaining adaptability to new clinical insights.

• The system leverages federated learning to enhance privacy and generalizability across diverse healthcare settings, enabling scalable and cost-effective deployment across hospitals and research institutions.

• Experimental results demonstrate that our model significantly reduces misdiagnosis rates, lowers healthcare costs, and improves patient outcomes by optimizing treatment pathways and reducing unnecessary interventions.

Related work

AI in personalized treatment optimization

AI-driven recommendation systems are revolutionizing personalized treatment approaches in healthcare, particularly in the management of neurological disorders. Traditional treatment strategies often follow standardized protocols that may not adequately address individual patient variability (19). AI algorithms, leveraging machine learning (ML) and deep learning (DL), facilitate data-driven decision-making by analyzing extensive patient datasets, including genetic information, clinical history, imaging data, and real-time physiological signals. These systems enhance diagnostic accuracy, treatment efficacy, and patient outcomes while optimizing resource allocation (20). One critical aspect of AI-based recommendation systems is their capacity to predict treatment responses in patients with neurological disorders such as epilepsy, Parkinson's disease, and Alzheimer's disease. For instance, ML models trained on large-scale patient data can identify patterns that predict how individuals respond to specific medications, thereby reducing trial-and-error prescriptions (21). Reinforcement learning frameworks have also been explored to dynamically adjust treatment regimens based on patient feedback and disease progression, improving long-term management strategies. Economic implications arise from the cost-effectiveness of AI-driven recommendations. By reducing misdiagnoses and ineffective treatments, healthcare expenses related to prolonged hospital stays, adverse drug reactions, and redundant diagnostic tests can be significantly minimized (22). Moreover, personalized treatment recommendations enhance patient adherence and reduce the burden on healthcare systems by preventing disease progression. However, implementing AI-driven recommendations necessitates substantial investment in data infrastructure, regulatory compliance, and clinician training, raising concerns about cost-effectiveness and scalability in different healthcare settings (23). Despite the promise of AI in personalized treatment, challenges persist in integrating these technologies into existing healthcare systems. Data privacy, interoperability between AI platforms and electronic health records (EHRs), and biases in training datasets remain key obstacles. Addressing these issues is essential to ensure equitable access to AI-driven personalized treatment recommendations across diverse patient populations (24). The economic impact of such systems varies depending on healthcare policies, reimbursement models, and regulatory environments in different regions.

Cost-benefit analysis of AI deployment

The implementation of AI-driven recommendation systems in neurological healthcare necessitates a thorough cost-benefit analysis to evaluate their economic feasibility. While AI technologies promise enhanced diagnostic precision and improved patient outcomes, their adoption requires significant financial investment in data infrastructure, computational resources, and skilled personnel (25). Understanding the economic trade-offs is crucial for stakeholders, including healthcare providers, policymakers, and insurers. One of the primary economic benefits of AI-driven recommendation systems lies in their potential to reduce healthcare costs by minimizing diagnostic errors, optimizing resource utilization, and reducing unnecessary procedures. Neurological disorders, such as stroke and multiple sclerosis, often involve complex diagnostic pathways that incur substantial expenses (26). AI-driven decision support tools streamline diagnostic processes by integrating multi-modal data sources, leading to faster and more accurate disease detection. This early intervention reduces long-term treatment costs by preventing disease progression and associated complications (27). Another cost consideration involves the development and maintenance of AI models. Training robust ML and DL models requires extensive labeled data, high-performance computing power, and continuous updates to maintain accuracy. The financial burden of these requirements may limit adoption, particularly in resource-constrained healthcare environments (28). AI-driven recommendations necessitate regulatory compliance, including adherence to data protection laws and clinical validation protocols, which impose further financial obligations on healthcare institutions. The economic impact of AI-driven recommendation systems also extends to workforce dynamics. While AI can enhance clinician productivity by automating routine decision-making processes, it may also alter job roles and require reskilling of medical professionals (29). The shift toward AI-assisted healthcare raises concerns regarding workforce displacement and necessitates investment in training programs to ensure seamless human-AI collaboration. The affordability of AI solutions varies across healthcare systems, with high-income countries more readily integrating these technologies compared to low-resource settings (30). Quantifying the return on investment (ROI) of AI in neurological healthcare remains an ongoing challenge. Studies assessing the economic value of AI-driven interventions emphasize the need for long-term evaluations that consider both direct financial savings and indirect benefits, such as improved quality of life for patients and reduced caregiver burden (31). Future research should focus on developing standardized methodologies for evaluating the economic impact of AI-driven recommendation systems to guide informed policy decisions.

Health equity and economic disparities

The economic implications of AI-driven recommendation systems in healthcare must also be examined through the lens of health equity and economic disparities. While AI holds the potential to improve healthcare accessibility and efficiency, disparities in the availability and affordability of AI-driven interventions raise ethical and economic concerns (32). Ensuring that AI technologies benefit diverse populations without exacerbating existing inequalities remains a critical challenge. One major issue is the digital divide, where disparities in healthcare infrastructure and technological access hinder equitable AI deployment. High-income countries with well-established healthcare systems can afford to integrate AI-driven recommendations seamlessly, while low- and middle-income countries (LMICs) face barriers such as limited data availability, inadequate computing resources, and insufficient funding for AI initiatives (33). This disparity may widen the healthcare gap, leaving underprivileged populations without access to advanced AI-assisted treatment recommendations for neurological disorders. Biases in AI models further contribute to health inequities. AI-driven recommendation systems trained on datasets that underrepresent certain demographic groups may produce skewed predictions, leading to suboptimal treatment recommendations for marginalized populations (34). Addressing these biases requires investment in diverse and representative data collection efforts, which adds to the economic burden of AI implementation. Ensuring algorithmic transparency and interpretability is essential for fostering trust among clinicians and patients, which may require further regulatory oversight and associated costs (35). The financial burden of AI-driven healthcare extends to patients, particularly in regions where out-of-pocket healthcare expenses are significant. If AI-driven recommendations are primarily available through private healthcare providers, economically disadvantaged patients may struggle to afford these services, exacerbating healthcare inequalities (36). Policy interventions, such as subsidized AI-based diagnostics and insurance coverage for AI-assisted treatments, could mitigate these disparities and ensure broader access to AI-driven neurological care. Moreover, the economic impact of AI on healthcare reimbursement models remains uncertain. Traditional reimbursement structures may not account for AI-driven decision support tools, leading to gaps in financial sustainability for healthcare providers (37). Developing appropriate reimbursement frameworks that incentivize the adoption of AI without increasing patient costs is essential for promoting widespread use. Future research should explore economic policies that balance AI innovation with equitable healthcare access, ensuring that AI-driven recommendation systems contribute to inclusive and cost-effective neurological care (38).

Method

Overview

Health Economics is a multidisciplinary field that applies economic theories and principles to healthcare systems, medical institutions, and health-related behaviors. This section provides a structured approach to understanding the economic mechanisms that influence healthcare policies, costs, and outcomes.

We introduce the fundamental economic concepts that underpin health economics, including utility maximization, opportunity costs, and market inefficiencies in healthcare. These principles serve as the theoretical foundation for subsequent discussions. In Section 3.2, we formalize the economic modeling of healthcare markets, incorporating elements such as supply and demand dynamics, insurance mechanisms, and government interventions. The mathematical formulation of these models provides a rigorous framework to analyze different policy implications. In Section 3.3, we present our novel economic model that extends traditional frameworks by integrating uncertainty, behavioral economics, and dynamic interactions among stakeholders. This approach allows for a more comprehensive understanding of healthcare decision-making under risk and incomplete information. In Section 3.4, we propose a new strategic approach to optimizing health outcomes while balancing economic efficiency. This strategy employs advanced computational methods, such as reinforcement learning and stochastic optimization, to address the complexities inherent in health economics. We establish a systematic methodology for analyzing and improving healthcare systems using economic principles, thereby contributing to both theoretical advancements and practical implementations in policy-making.

Preliminaries

Health economics seeks to understand how scarce resources are allocated within healthcare systems to maximize efficiency and well-being. This section establishes the mathematical and theoretical foundation for analyzing healthcare markets, pricing mechanisms, and policy interventions.

Let the healthcare market be represented by a set of consumers C and providers P. Each consumer iC derives utility from consuming healthcare services qi at price pi, given by:

Ui=U(qi,Yi-piqi),    (1)

where Yi is individual income, and U(·) is a concave function reflecting diminishing marginal utility.

Healthcare demand is influenced by price elasticity, denoted as:

εi=qipi×piqi,    (2)

which captures how sensitive healthcare consumption is to changes in price.

On the supply side, healthcare providers maximize profit:

Πj=pjqj-C(qj),    (3)

where C(qj) represents the cost function of provider j. Assuming convexity, marginal cost is given by:

MCj=dC(qj)dqj.    (4)

Market equilibrium occurs where supply equals demand,

iCqi(p)=jPqj(p).    (5)

However, healthcare markets often exhibit inefficiencies due to asymmetric information. Let θi denote the health status of consumer i, observable to the consumer but not to the insurer. The expected cost to the insurer in an insurance market is:

E[C|θi]=θiC(qi,θ)f(θ)dθ,    (6)

leading to potential adverse selection, modeled as:

P(qi)=P0+λE[C|θi],    (7)

where P0 is a base premium and λ reflects risk adjustments.

Moral hazard further complicates the market. If qi* is the optimal healthcare consumption in the absence of insurance, then under insurance coverage α ∈ [0, 1], the insured demand qi satisfies:

Ui(qi)=αUi(qi*),    (8)

indicating overconsumption due to reduced out-of-pocket costs.

To correct these inefficiencies, governments often intervene through taxation, subsidies, and regulation. Let T(q) denote a subsidy function such that the consumer's effective price is:

pi=pi-T(qi).    (9)

A central planner may seek to maximize social welfare:

W=iCUi-jPC(qj)+λiCH(qi),    (10)

where H(qi) represents health benefits. These formulations provide the basis for the new economic model and strategic interventions proposed in subsequent sections.

Dynamic Equilibrium Model for health economics (DEHE)

To address the complexities of healthcare markets, we propose a Dynamic Equilibrium Model for Health Economics (DEHE), which integrates consumer behavior, provider incentives, and government interventions under a unified mathematical framework. Unlike traditional static models, DEHE captures dynamic interactions and policy adjustments over time (as shown in Figure 1).

Figure 1
www.frontiersin.org

Figure 1. Illustration of the Dynamic Equilibrium Model for health economics (DEHE), integrating consumer behavior, insurance premium adjustments, and government policy optimization. The model dynamically balances intertemporal consumer optimization, risk-based premium adjustments, and market interventions to ensure economic sustainability and healthcare efficiency. Key components include premium adjustment and risk assessment, consumer decision-making, and policy-driven subsidies and tax regulations to maintain equilibrium in the healthcare system.

Intertemporal consumer optimization

Each consumer iC optimizes lifetime utility by making healthcare consumption decisions dynamically over a finite time horizon. The consumer's utility function depends on healthcare consumption qi(t) and disposable income, which is the difference between total income Yi and healthcare expenditures pi(t)qi(t). The objective function is formulated as:

maxqi(t)0Te-ρtU(qi(t),Yi-pi(t)qi(t))dt,    (11)

where ρ > 0 is the discount rate, indicating time preference. Consumers weigh current and future healthcare consumption, leading to an intertemporal trade-off that captures dynamic adjustments in medical demand.

To determine optimal consumption, the necessary first-order condition for maximization requires:

Uqi-λi(pi+dpidt)=0,    (12)

where λi is the Lagrange multiplier associated with the budget constraint. This equation implies that the marginal benefit from additional healthcare consumption must balance with the effective price, accounting for price changes over time.

Moreover, consumer demand evolves dynamically as preferences, income, and price expectations adjust. The Euler equation governing optimal consumption follows:

ddt(Uqi)=ρUqi-λidpidt.    (13)

This equation captures intertemporal substitution effects, where changes in marginal utility and price expectations influence the trajectory of healthcare demand.

Assuming a quadratic utility function for tractability, U(qi,Yi)=aqi-b2qi2+cYi, the optimal consumption path is determined by solving:

qi*(t)=a-λi(pi+dpidt)b.    (14)

This expression shows how optimal demand depends on preference parameters (a, b) and dynamic price adjustments. The intertemporal optimization framework highlights how consumers respond to policy changes, income variations, and price fluctuations over time (as shown in Figure 2).

Figure 2
www.frontiersin.org

Figure 2. The figure illustrates the Intertemporal Consumer Optimization framework, where consumers dynamically adjust healthcare consumption decisions over time to maximize lifetime utility. The model incorporates utility maximization, price adjustment mechanisms, risk and insurance premium adjustments, and government subsidy and tax policy optimization. Each mechanism applies a sequence of transformations, including linear projections, softmax attention, dot product operations, and adaptive parameter adjustments to model intertemporal decision-making. The mathematical formulation defines consumer utility maximization under budget constraints, leading to an optimal consumption path driven by price expectations and policy interventions. This optimization process enables policymakers to design adaptive healthcare pricing, insurance adjustments, and incentive structures for improved economic efficiency. (a) Utility maximization. (b) Price adjustment mechanism. (c) Risk and insurance premium adjustment. (d) Government subsidy & tax policy optimization.

Dynamic insurance premiums

Health insurance premiums must adapt dynamically to evolving risk exposure and adverse selection to ensure market stability and financial sustainability. A core component of this adaptation is the insurer's expected payout, which depends on the distribution of health risks among insured individuals. The expected cost of claims given an intervention level Ii(t) is computed as:

E[C|Ii(t)]=θC(qi,θ)f(θ)dθ,    (15)

where C(qi, θ) denotes the cost function associated with individual i's health state θ, and f(θ) represents the probability density function of health risks in the population. To maintain financial equilibrium while preventing sudden premium fluctuations, we introduce a dynamic premium adjustment rule that ensures premiums remain aligned with expected costs:

dPidt=γ(E[C|Ii]-Pi),    (16)

where γ is the premium adjustment speed parameter, ensuring that pricing responds adaptively to changing risk profiles without excessive volatility. Premium adjustments incorporate behavioral incentives to encourage preventive care and adherence to medical guidelines. A risk-adjusted premium model is formulated as:

Pi(t)=P0+λE[C|θi]-ρI(qi,t),    (17)

where P0 is the base premium, λ is the risk adjustment coefficient, and ρI(qi, t) represents premium discounts or penalties based on adherence to incentivized health behaviors. To ensure financial stability across the insurer's portfolio, the total expected revenue must balance the total expected payouts:

iCPi(t)Ii(t)=iCE[C|Ii(t)]Ii(t).    (18)

This equilibrium condition ensures that premiums dynamically reflect both individual risk levels and collective healthcare costs, promoting both sustainability and fairness in the insurance market.

Government policy optimization

A central authority aims to maximize social welfare by optimally setting subsidies S(t) and taxes T(t). The objective function incorporates consumer utility, production costs, and potential externalities, weighted by a factor λ. The optimization problem is formulated as:

maxS(t),T(t)0Te-ρt(iCUi(qi,S,T)-jPC(qj)+λiCH(qi))dt,    (19)

where C represents consumers, P denotes producers, Ui(qi, S, T) is the utility function of consumer i, C(qj) is the cost function for producer j, and H(qi) captures externalities associated with consumption. The discount factor ρ > 0 ensures present values weigh more than future values.

The policy instruments S(t) and T(t) evolve dynamically based on deviations from the socially optimal welfare level W*, ensuring an adaptive regulatory approach:

dSdt=β(W-W*),dTdt=δ(W-W*),    (20)

where W=iCUi-jPC(qj)+λiCH(qi) represents the actual welfare, and β, δ > 0 are parameters controlling the speed of policy adjustments.

The market equilibrium conditions impose constraints on individual consumption and production choices. The aggregate demand QD and supply QS must satisfy:

QD(S,T)=QS(S,T),    (21)

ensuring that markets clear at any given time. Optimal production and consumption decisions follow first-order conditions derived from profit maximization and utility maximization:

Uiqi=p-T,Cqj=p-S,    (22)

where p represents the market price, balancing the incentives of consumers and producers. Through these dynamic adjustments and equilibrium conditions, the government aims to steer the economy toward a socially optimal state while accounting for market responses.

To further ensure the robustness and generalizability of our AI-driven recommendation system, we incorporate diverse and representative datasets during the model training phase. This approach enables the system to adapt to variations across different patient populations and healthcare settings. We implement a continuous retraining strategy supported by federated learning, allowing the model to update and refine its parameters using new clinical data while preserving data privacy. Periodic validation is conducted using external datasets to assess performance stability and minimize overfitting, ensuring that the system remains reliable and effective under evolving real-world conditions.

To further advance the economic understanding of AI integration in healthcare, it is essential to develop more comprehensive models that capture the nuanced and evolving interactions between AI systems and healthcare markets. Traditional static models often fall short in representing the real-time feedback loops, behavioral shifts, and policy responses triggered by AI-driven interventions. In this study, we extend existing economic frameworks by incorporating dynamic elements such as intertemporal decision-making, risk-adjusted insurance mechanisms, and government policy adaptations through reinforcement learning. Our model accounts for the complexity of stakeholder behavior, including provider incentives and patient responsiveness, under uncertainty and asymmetric information. This approach enables a more accurate simulation of how AI systems influence healthcare costs, access, and outcomes over time, thereby offering a foundation for policy recommendations that are both economically sustainable and technologically adaptive.

Adaptive Policy Framework for Healthcare Optimization (APFHO)

Building upon the Dynamic Equilibrium Model for Health Economics (DEHE), we propose an Adaptive Policy Framework for Healthcare Optimization (APFHO). This framework integrates dynamic mechanisms to enhance healthcare efficiency, minimize market distortions, and ensure policy adaptability (as shown in Figure 3).

Figure 3
www.frontiersin.org

Figure 3. Adaptive Policy Framework for Healthcare Optimization (APFHO) integrates dynamic mechanisms to improve healthcare efficiency, minimize market distortions, and ensure policy adaptability. The framework consists of three interconnected components, dynamic pricing optimization (DPO), which adjusts healthcare prices in response to real-time demand-supply imbalances while incorporating stability constraints; incentive-based policy adjustments (IBPA), which allocate financial and behavioral incentives dynamically to optimize healthcare resource utilization and improve patient adherence; and reinforcement learning for policy optimization (RLPO), which leverages reinforcement learning techniques to refine government interventions such as taxation, subsidies, and insurance regulations. These components work collectively to create a responsive and efficient healthcare policy system.

Dynamic pricing optimization

To address inefficiencies in healthcare pricing, we introduce a dynamic pricing model that continuously adjusts prices based on real-time demand-supply imbalances while incorporating regulatory constraints to ensure stability. In healthcare markets, prices cannot fluctuate arbitrarily due to regulatory oversight and consumer affordability concerns. Therefore, we propose a time-dependent price adjustment mechanism where price changes respond to excess demand or surplus supply. The price evolution equation follows:

dpidt=κ(Di(t)-Si(t)),    (23)

where κ > 0 represents the price sensitivity parameter, and Di(t), Si(t) denote the demand and supply of healthcare services at time t. When demand exceeds supply, prices rise to restore equilibrium, whereas excess supply leads to price reductions. However, to prevent excessive price fluctuations that could destabilize the market, a bounded price change condition is imposed:

|pi(t)-pi(t-Δt)|δpmax,    (24)

where δpmax is the maximum permissible price change per unit time. This constraint ensures gradual price adjustments, preventing sharp increases that may burden consumers or drastic drops that may discourage providers.

Incorporating consumer price elasticity into the model, we define the demand response function:

Di(t)=D0e-ϵpi(t),    (25)

where ϵ > 0 represents the price elasticity of demand. Higher elasticity implies stronger consumer reactions to price changes, influencing the speed and magnitude of equilibrium adjustments. To further regulate price dynamics and ensure convergence, we introduce a dampening term that penalizes excessive price fluctuations:

dpidt=κ(Di(t)-Si(t))-η(pi(t)-pi*),    (26)

where η > 0 is the stabilization coefficient, and pi* is the long-run equilibrium price. This formulation ensures that prices not only react to market imbalances but also gradually converge to stable long-term values, fostering a sustainable and efficient healthcare pricing system.

Incentive-based policy adjustments

Optimizing healthcare consumption requires an adaptive incentive mechanism that dynamically allocates financial and behavioral incentives based on real-time consumer responsiveness. By strategically adjusting incentives, policymakers can enhance preventive care participation, improve treatment adherence, and optimize resource allocation. The incentive function is defined as:

I(qi,t)=β1T(qi)+β2R(qi)+β3M(qi),    (27)

where T(qi) represents tax rebates or subsidies provided to encourage specific healthcare behaviors, R(qi) denotes risk-adjusted insurance benefits that modify premiums or coverage based on preventive healthcare engagement, and M(qi) accounts for personalized medical adherence incentives such as discounts on medications or financial rewards for completing treatment programs. The coefficients β1, β2, β3 are dynamically optimized to maximize healthcare efficiency while maintaining budget feasibility.

The optimal incentive allocation follows from an individual utility maximization problem:

maxβiCUi(qi,Yi-piqi+I(qi,t)),    (28)

where Ui(qi, ·) represents the utility function of consumer i, Yi is disposable income, and piqi denotes healthcare expenditures. The total incentive distribution is subject to a budget constraint:

iCI(qi,t)Bmax.    (29)

To ensure the effectiveness of these incentives, their allocation is adjusted using an elasticity-based optimization model. The response elasticity of healthcare consumption to incentives is defined as:

ϵI=qiI(qi,t)×I(qi,t)qi.    (30)

A higher elasticity suggests greater responsiveness to incentives, allowing for targeted allocation where interventions yield the highest impact. To further optimize the incentive distribution, we introduce a **marginal cost-effectiveness constraint**, which ensures that incentives are allocated efficiently by prioritizing areas where they generate the highest health benefits per unit of expenditure:

UiI(qi,t)λ,    (31)

where λ represents a minimum effectiveness threshold. This condition ensures that incentives are only distributed to individuals or groups where the marginal increase in utility surpasses the predefined threshold, preventing inefficient allocation of limited resources.

In addition to cost-effectiveness, stability in incentive adjustments is crucial to avoid abrupt changes in consumer behavior that could lead to unintended consequences such as healthcare overutilization or policy resistance. To regulate the pace of incentive modifications, we introduce a **smoothness constraint**:

|I(qi,t+1)-I(qi,t)|Δmax,    (32)

where Δmax is a predefined limit preventing sudden large-scale changes in incentive values. This constraint ensures that adjustments to tax rebates, insurance benefits, and medical adherence rewards remain gradual and predictable, facilitating better long-term behavioral adaptation.

By integrating these optimization components, this framework enables real-time adaptive policymaking, ensuring that incentives effectively drive behavioral change while maintaining cost efficiency. The combination of elasticity-based allocation, marginal cost-effectiveness constraints, and smoothness in policy adjustments ensures that financial and behavioral incentives maximize healthcare outcomes without excessive expenditure. This approach allows policymakers to continuously refine healthcare incentives based on observed responses, thereby creating a self-improving system that evolves to meet public health objectives dynamically.

Reinforcement learning for policy optimization

Government interventions, including taxation, subsidies, and insurance regulations, can be dynamically optimized using a reinforcement learning framework. In this approach, the policy function π(S, T) maps state variables (S, T) to intervention actions, aiming to maximize long-term social welfare. The expected cumulative welfare function serves as the objective function:

J(π)=𝔼[t=0TγtWt],    (33)

where Wt is the social welfare function at time t, and γ ∈ (0, 1) is the discount factor, ensuring that future rewards contribute less to the present decision-making process. The optimization procedure follows a policy gradient approach, where updates to the policy parameters θ are performed iteratively through gradient ascent:

θt+1=θt+ηθJ(πθ),    (34)

where η > 0 is the learning rate, and ∇θJθ) represents the gradient of the expected cumulative welfare with respect to policy parameters. The state transition follows a Markov decision process (MDP) framework, where the probability of moving to the next state (S′, T′) depends on the current state and chosen intervention:

P(S,T|S,T)=aπ(a|S,T)P(S,T|S,T,a),    (35)

ensuring that the government policy evolves adaptively based on observed market responses. The update of welfare at each step depends on the reward function derived from economic efficiency and social considerations:

Wt+1=Wt+α(Rt+γWt+1-Wt),    (36)

where Rt is the immediate reward function reflecting economic gains and externalities, and α is a step-size parameter ensuring stable learning dynamics. Through iterative updates and market interactions, reinforcement learning provides an adaptive mechanism for policy optimization, allowing governments to respond to dynamic economic conditions efficiently (As shown in Figure 4).

Figure 4
www.frontiersin.org

Figure 4. The reinforcement learning framework for policy optimization dynamically adjusts government interventions, such as taxation, subsidies, and insurance regulations to maximize social welfare. The diagram illustrates the deep learning architecture supporting this approach, incorporating convolutional and depth-wise convolutional layers for feature extraction and decision-making. The policy function is trained using a policy gradient method, iteratively optimizing intervention actions based on economic conditions. By modeling the decision process as a Markov Decision Process (MDP), the framework ensures adaptive policy adjustments through reward-based learning. This approach enhances economic efficiency and responsiveness to market fluctuations.

To further support clinical adoption, improving the interpretability of AI-generated recommendations remains an essential direction for future enhancement. In high-stakes medical environments, trust and transparency are critical for decision support tools to be effectively integrated into clinical workflows. Explainable AI (XAI) techniques should be explored to ensure that healthcare professionals can understand, audit, and validate the logic behind AI decisions. By incorporating model interpretability frameworks—such as attention visualization, feature attribution methods, or rule-based surrogate models—our system can improve clinician confidence and facilitate human-AI collaboration. This is particularly important in neurological disorder management, where treatment decisions are complex and highly individualized. Enhancing explainability not only fosters greater trust in the AI system but also aligns with ethical and regulatory requirements for transparency in healthcare technologies.

To facilitate the integration of AI-driven recommendation systems into existing healthcare infrastructures, it is essential to consider practical implementation pathways and stakeholder collaboration. Potential partnerships with healthcare providers and policymakers can support the alignment of AI deployment with clinical workflows and regulatory standards. Integration with electronic health records (EHRs) and hospital information systems is necessary to ensure seamless data flow and decision support in real time. Regulatory compliance, including adherence to data protection and clinical validation protocols, must be addressed early in the deployment process. Effective implementation also requires targeted training programs for clinicians and administrative staff to promote system usability and trust. Pilot deployments in selected healthcare settings can serve as testbeds for refining the system's operational readiness and for identifying challenges in scalability, user adoption, and system interoperability.

Experimental setup

Dataset

The MovieLens Dataset (39) is a widely used benchmark in recommendation system research, containing user ratings for movies. It includes metadata such as movie genres, timestamps, and user demographics, allowing researchers to explore collaborative filtering and deep learning-based recommendation models. The dataset is available in multiple sizes, ranging from small subsets with a few thousand ratings to larger versions with millions of interactions. It is frequently used to evaluate personalized recommendation algorithms and to study user behavior patterns. The structured format and extensive annotations make it an essential resource for testing various machine learning and artificial intelligence techniques in content recommendation. The Epinions Dataset (40) is derived from the Epinions consumer review platform, where users share opinions on a variety of products. It consists of user-item interactions along with trust relationships between users, making it particularly valuable for studying social recommendation systems. Unlike traditional collaborative filtering datasets, it incorporates explicit trust and distrust links, allowing researchers to explore trust-aware recommendation models. The dataset helps in understanding the role of social influence in user preferences and improving the accuracy of personalized recommendations. Its rich structure supports research in areas such as graph-based recommendations, social network analysis, and trust propagation in online communities. The Criteo Dataset (41) is a large-scale dataset used primarily for research in online advertising and click-through rate (CTR) prediction. It contains anonymized user interaction data from online advertisements, including categorical and numerical features that capture various aspects of ad impressions. The dataset is widely employed for training deep learning models, particularly in the field of computational advertising. It enables researchers to develop models that optimize ad targeting by predicting user engagement. Due to its real-world nature and extensive size, it is considered a standard benchmark for evaluating machine learning algorithms in digital marketing and personalized advertising strategies. The GHDx Dataset (42) comes from the Global Health Data Exchange, a comprehensive resource providing global health-related statistics. It includes data on disease prevalence, mortality rates, healthcare access, and other demographic indicators, making it crucial for public health research and policy-making. Researchers use this dataset to analyze trends in global health, evaluate healthcare interventions, and model disease spread. Its extensive geographic and temporal coverage supports studies on epidemiology, health economics, and medical resource allocation. By offering a wealth of structured health data, the GHDx dataset facilitates informed decision-making and enhances the effectiveness of healthcare policies worldwide.

Experimental details

We conduct extensive experiments to evaluate the effectiveness of our proposed method on four standard action recognition benchmarks: MovieLens, Epinions, Criteo, and GHDx. Our implementation is based on PyTorch, and all experiments are conducted on NVIDIA A100 GPUs. The input video frames are resized to 224 × 224 and normalized using the mean and standard deviation of ImageNet. We employ a frame sampling strategy that extracts 16 or 32 frames per video clip, ensuring a balance between efficiency and performance. For model architecture, we utilize a backbone pretrained on GHDx and fine-tune it on the target datasets. Specifically, we experiment with ResNet-50, SlowFast, and TimeSformer models. The models are trained using SGD with a momentum of 0.9 and weight decay of 5 × 10−4. The initial learning rate is set to 0.01 and follows a cosine annealing schedule. We use a batch size of 64 for MovieLens and Epinions and a batch size of 32 for Criteo and GHDx due to memory constraints. Training runs for 100 epochs for MovieLens and Epinions, and 50 epochs for the larger datasets, Criteo and GHDx. We employ data augmentation techniques, including random cropping, horizontal flipping, and color jittering, to improve generalization. For temporal augmentation, we adopt random frame dropping and frame interpolation to enhance the model's ability to recognize actions across various speeds. During inference, we use 10-view testing, where non-overlapping clips are sampled and averaged to obtain the final prediction. Evaluation is performed using top-1 and top-5 accuracy metrics for GHDx and mean Average Precision (mAP) for Criteo. For MovieLens and Epinions, we report top-1 accuracy following standard protocols. All results are averaged over three training/testing splits to ensure statistical reliability. To ensure a fair comparison, we follow the standard training/testing splits for each dataset. We also conduct ablation studies to analyze the impact of key components in our method, including the choice of backbone, temporal modeling strategy, and data augmentation techniques. The results demonstrate that our approach consistently outperforms existing methods, showcasing its effectiveness in real-world action recognition scenarios.

Comparison with SOTA methods

We compare our proposed method with state-of-the-art (SOTA) approaches on four standard action recognition benchmarks: MovieLens, Epinions, Criteo, and GHDx. The comparative results are presented in Tables 1, 2. Our model consistently outperforms existing methods across all datasets in terms of Precision, Recall, F1 Score, and NDCG metrics.

Table 1
www.frontiersin.org

Table 1. A comparative analysis of our approach against state-of-the-art methods on the MovieLens and Epinions datasets.

Table 2
www.frontiersin.org

Table 2. A comparative evaluation of our approach against state-of-the-art methods on the Criteo and GHDx datasets.

In Figures 5, 6, for MovieLens and Epinions, our method achieves the highest F1 Score of 86.50% and 81.92%, respectively. This improvement can be attributed to our enhanced temporal modeling approach, which effectively captures long-range dependencies in videos. Compared to previous methods such as GRU4Rec and LightGCN, our approach demonstrates superior ability in distinguishing fine-grained action differences, leading to better classification performance. The increase in NDCG, reaching 89.77% and 85.33% on MovieLens and Epinions respectively, highlights our method's capability in ranking relevant actions higher, which is crucial for real-world applications. For Criteo and GHDx, our model achieves an F1 Score of 84.92% and 80.45%, outperforming previous SOTA models such as SASRec and GRU4Rec. The significant improvements in these large-scale datasets indicate that our approach generalizes well to diverse real-world video distributions. One key factor behind this performance gain is the integration of an optimized frame selection strategy, which ensures that critical motion cues are preserved while reducing redundant information. Our model benefits from robust spatial-temporal feature extraction, which improves recall and reduces misclassification errors.

Figure 5
www.frontiersin.org

Figure 5. Evaluating the performance of leading methods on the MovieLens and Epinions datasets.

Figure 6
www.frontiersin.org

Figure 6. An evaluation of state-of-the-art methods on the Criteo and GHDx datasets.

A deeper analysis of the results suggests that our model's superior performance is due to three main factors. Our method incorporates an advanced attention-based mechanism that efficiently learns discriminative action representations. This enables the model to capture subtle action variations, which is particularly beneficial for datasets with complex action categories such as Criteo and GHDx. Our approach employs an improved augmentation strategy that enhances robustness to variations in camera angles, lighting, and occlusions, thereby reducing overfitting. The optimized backbone architecture, leveraging pretrained GHDx weights, provides a strong feature extraction foundation, allowing the model to adapt effectively to different datasets. The results demonstrate that our method sets a new benchmark in action recognition, consistently achieving higher performance than existing approaches across multiple datasets. The improvements in Precision, Recall, and NDCG confirm the effectiveness of our proposed enhancements, making it a compelling solution for real-world video analysis tasks.

Ablation study

To better understand the contribution of individual components in our proposed method, we conduct an ablation study on four benchmark datasets: MovieLens, Epinions, Criteo, and GHDx. The results are summarized in Tables 3, 4. We systematically remove key components, denoted as w./o. Dynamic Insurance Premiums, w./o. Government Policy Optimization, and w./o. Dynamic Pricing Optimization, and evaluate the impact on Precision, Recall, F1 Score, and NDCG.

Table 3
www.frontiersin.org

Table 3. Analysis of ablation study findings on the MovieLens and Epinions datasets.

Table 4
www.frontiersin.org

Table 4. Findings from the ablation study on the Criteo and GHDx datasets.

In Figures 7, 8, from the results, we observe that removing Dynamic Insurance Premiums leads to a significant drop in performance, particularly on MovieLens and Criteo. The decrease in F1 Score from 86.50% to 83.21% on MovieLens and from 84.92% to 81.42% on Criteo suggests that Dynamic Insurance Premiums plays a crucial role in enhancing discriminative feature representation. This component is responsible for capturing long-range dependencies in video sequences, and its absence results in reduced model effectiveness in handling complex motion patterns. Similarly, excluding Government Policy Optimization results in a moderate performance drop across all datasets. The F1 Score decreases from 86.50% to 85.04% on MovieLens and from 84.92% to 82.90% on Criteo. This indicates that Government Policy Optimization is beneficial for improving model robustness, particularly in datasets with diverse action categories such as Criteo and GHDx. The role of Government Policy Optimization in temporal modeling ensures better motion consistency, leading to improved recall and precision. Without it, the model struggles to differentiate between visually similar but semantically different actions. Removing Dynamic Pricing Optimization results in a slightly smaller drop in performance compared to Dynamic Insurance Premiums and Government Policy Optimization, but still negatively affects overall accuracy. The F1 Score drops to 84.10% on MovieLens and 83.60% on Criteo, demonstrating that Dynamic Pricing Optimization contributes to refining action classification by enhancing spatial-temporal feature integration. This suggests that while Dynamic Pricing Optimization is not the most critical aspect, it still provides additional refinement, helping to boost precision and ranking quality, as reflected in the NDCG metric.

Figure 7
www.frontiersin.org

Figure 7. Investigation of our method's ablation study on the MovieLens and Epinions Datasets. DIP, Dynamic Insurance Premiums; GPO, Government Policy Optimization; DPO, Dynamic Pricing Optimization.

Figure 8
www.frontiersin.org

Figure 8. Evaluation of our method through ablation study on the Criteo and GHDx datasets. DIP, Dynamic Insurance Premiums; GPO, Government Policy Optimization; DPO, Dynamic Pricing Optimization.

Our full model consistently outperforms all ablated versions, highlighting the importance of each proposed component. The results validate that our method effectively balances spatial and temporal information, leading to improved action recognition performance across different datasets.

Table 5 presents the comparative economic outcomes between patients receiving traditional healthcare services without AI support and those managed with the proposed AI-driven recommendation system. The results indicate a consistent reduction in overall treatment costs among patients in the AI-assisted group, with expenditures decreasing from an average of approximately ¥44,750 in the control group to around ¥34,400 in the experimental group. This suggests that the AI system effectively optimized diagnostic and treatment pathways, leading to fewer unnecessary procedures and hospital visits. In parallel, the misdiagnosis rate dropped significantly from a baseline range of 16–20% in the control group to as low as 6–9% in the AI group, underscoring the potential of AI to enhance diagnostic accuracy, particularly in complex neurological conditions. The insurance premium levels also showed a downward adjustment in the AI group, reflecting reduced risk profiles and improved patient stratification. Premiums decreased by an average of 10-15%, which indicates that AI-driven efficiency gains could be passed on to the consumer in the form of lower insurance costs. The data reveal a positive shift in healthcare accessibility, as all patients in the AI group proceeded with treatment, whereas a notable portion of the control group failed to access care. This highlights the role of AI in not only improving economic and clinical efficiency but also in enhancing equity by lowering barriers to care. These findings validate the economic relevance and practical value of incorporating AI recommendation systems in real-world healthcare settings, especially when integrated with adaptive policy mechanisms as proposed in our model.

Table 5
www.frontiersin.org

Table 5. Economic impact of AI recommendation systems on healthcare outcomes.

Based on the experimental results obtained from the ADNI and MIMIC-IV datasets in Table 6, Figure 9, our proposed AI-driven recommendation system integrated with the DEHE and APFHO frameworks demonstrates significant advantages over traditional healthcare management approaches. In the ADNI dataset, the average cost per patient was reduced from 1,500 USD to 870 USD, accompanied by a notable increase in recovery rate from 64% to 78%. Similarly, in the MIMIC-IV dataset, the AI-based approach achieved a reduction in cost from 1,430 USD to 920 USD and improved the recovery rate from 62% to 75%. These improvements highlight the economic efficiency and clinical effectiveness of the proposed system. Furthermore, the misdiagnosis rate in both datasets decreased substantially, indicating enhanced diagnostic accuracy. The average intervention delay was shortened, suggesting that the AI recommendations enabled faster clinical responses. Resource utilization also improved, reaching 85% and 88% in the AI groups for ADNI and MIMIC-IV respectively, compared to lower efficiency levels under traditional models. These outcomes confirm the potential of AI-integrated economic frameworks to optimize healthcare delivery by reducing costs, improving patient outcomes, and enhancing overall system performance in real-world clinical environments.

Table 6
www.frontiersin.org

Table 6. Experimental comparison of AI-driven recommendation system with traditional methods using ADNI and MIMIC-IV datasets.

Figure 9
www.frontiersin.org

Figure 9. Comparison of average cost per patient and recovery rate over time between the AI + DEHE/APFHO approach and traditional methods, demonstrating improved cost-efficiency and patient outcomes in the AI-driven system.

Discussion

The policy implications of AI-driven recommendation systems in healthcare are critical for ensuring both equitable access and long-term financial sustainability. To maximize their economic potential, it is essential to provide targeted strategies for policymakers that address real-world deployment barriers. These may include establishing regulatory frameworks that support dynamic pricing models and risk-adjusted insurance schemes, while maintaining strict data privacy and transparency standards. Public-private partnerships can facilitate integration with existing healthcare infrastructures, enabling AI systems to work alongside electronic health records and clinical decision support tools. Targeted subsidies and incentive programs can help lower-income or resource-constrained institutions adopt AI technologies, reducing the digital divide in healthcare delivery. Workforce training and stakeholder engagement are also necessary to ensure that healthcare professionals are equipped to interpret and trust AI-generated recommendations. By addressing these policy-level considerations, AI-driven systems can be implemented in a way that is both economically efficient and socially inclusive.

While our proposed model assumes rational responses from all stakeholders, we acknowledge that real-world healthcare markets are shaped by behavioral unpredictability, emotional factors, and institutional inertia. Future work will seek to incorporate principles from behavioral economics to better reflect the cognitive biases and bounded rationality that often influence decisions by patients, providers, and insurers. For example, prospect theory, inertia in care-seeking behavior, and status quo bias may significantly alter how stakeholders respond to AI-driven recommendations. Empirical validation through surveys and behavioral data collection will be crucial to calibrate such extensions. We recognize the ethical implications of deploying AI in healthcare, including the need for fairness, bias mitigation, and protection of patient privacy. Ensuring equitable access to AI-supported decision-making across diverse socioeconomic and demographic groups is essential to prevent the reinforcement of existing disparities. These considerations will inform future refinements of our model to enhance its realism, inclusivity, and policy relevance.

Conclusions and future work

In this study, we explored the economic implications of AI-driven recommendation systems in the healthcare sector, with a particular focus on neurological disorders. These AI systems have demonstrated significant potential in optimizing healthcare resource allocation, improving diagnostic accuracy, and enhancing overall patient outcomes. However, traditional economic models often fall short in effectively capturing the complexities of AI integration, primarily due to their inability to account for the dynamic interactions between healthcare stakeholders, market inefficiencies, and behavioral responses to AI-based decision-making. The effectiveness of our model depends heavily on the availability and quality of healthcare data. As noted, incomplete or biased data presents significant challenges in real-world applications. To address these concerns, future enhancements will focus on strategies for managing missing or incomplete data, as well as mitigating biases present in healthcare datasets. We plan to incorporate data quality assessment metrics to guide preprocessing and improve the model's reliability. We also recognize the potential of generating synthetic data to both enhance model robustness and address privacy concerns. Methods such as Generative Adversarial Networks (GANs) may be explored to generate realistic healthcare data that can supplement real-world datasets, ensuring patient privacy is maintained. Existing frameworks struggle with asymmetric information and moral hazard, which affect insurance mechanisms and cost structures. To address these challenges, we introduced a Dynamic Equilibrium Model for Health Economics (DEHE), which incorporates reinforcement learning and stochastic optimization to model decision-making under uncertainty. This novel approach integrates dynamic pricing mechanisms, incentive-based behavioral interventions, and adaptive insurance premium adjustments to enhance economic efficiency. Our experimental results, validated through a multi-agent simulation framework, demonstrate that DEHE successfully balances patient access and cost-effectiveness while improving the overall economic sustainability of AI-driven healthcare systems.

Despite the promising outcomes of our study, there are two notable limitations that warrant further exploration. While our model effectively addresses dynamic decision-making under uncertainty, it remains constrained by the availability and quality of healthcare data. The effectiveness of DEHE largely depends on the robustness of real-world datasets, and potential biases in AI-driven recommendations could lead to suboptimal economic outcomes. Future work should focus on integrating diverse and high-quality data sources to enhance model accuracy and reliability. Our approach assumes a rational response from all stakeholders, yet real-world healthcare markets exhibit significant behavioral unpredictability, regulatory constraints, and ethical considerations. The interplay between AI-driven decision-making and human factors–such as trust, adoption rates, and policy compliance–requires further empirical validation. Future research should explore hybrid approaches that incorporate behavioral economics and regulatory frameworks to ensure the broader applicability and fairness of AI-driven healthcare systems. By addressing these challenges, we can refine AI-driven recommendation systems to create more sustainable, equitable, and cost-effective healthcare solutions.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

Author contributions

JZ: Methodology, Supervision, Project administration, Validation, Resources, Visualization, Writing – original draft, Writing – review & editing. SX: Data curation, Conceptualization, Formal analysis, Investigation, Funding acquisition, Software, Writing – original draft, Writing – review & editing. LL: Writing – original draft, Writing – review & editing.

Funding

The author(s) declare that financial support was received for the research and/or publication of this article. This work was supported by 1. Project fifteen of Zhejiang Education Department: the construction and implementation of the personnel training model of “Industry, education and innovation” for Finance and economics majors based on the OBE concept, Project Number: jg20230407. 2. Zhejiang Province Department of Education Cooperation Project: based on the “Four-chain coordination” of higher vocational colleges creative integration of collaborative education model construction and practice, Project Number: 428.

Conflict of interest

SX was employed by Shenzhen Jizhi Laser Technology Co., Ltd.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that no Gen AI was used in the creation of this manuscript.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

1. Botha NN, Segbedzi CE, Dumahasi VK, Maneen S, Kodom RV, Tsedze IS, et al. Artificial intelligence in healthcare: a scoping review of perceived threats to patient rights and safety. Arch Public Health. (2024) 82:188. doi: 10.1186/s13690-024-01414-1

PubMed Abstract | Crossref Full Text | Google Scholar

2. Zhang Z, Patra BG, Yaseen A, Zhu J, Sabharwal R, Roberts K, et al. Scholarly recommendation systems: a literature survey. Knowl Inf Syst. (2023) 65:4433–78. doi: 10.1007/s10115-023-01901-x

Crossref Full Text | Google Scholar

3. Botha NN, Ansah EW, Segbedzi CE, Dumahasi VK, Maneen S, Kodom RV, et al. Artificial intelligent tools: evidence-mapping on the perceived positive effects on patient-care and confidentiality. BMC Digital Health. (2024) 2:1–24. doi: 10.1186/s44247-024-00091-y

Crossref Full Text | Google Scholar

4. Ko H, Lee S, Park Y, Choi A. A survey of recommendation systems: recommendation models, techniques, and application fields. Electronics. (2022) 11:141. doi: 10.3390/electronics11010141

Crossref Full Text | Google Scholar

5. Forouzandeh S, Rostami M, Berahmand K. A hybrid method for recommendation systems based on tourism with an evolutionary algorithm and topsis model. Fuzzy Inform Eng. (2022) 14:26–50. doi: 10.1080/16168658.2021.2019430

Crossref Full Text | Google Scholar

6. Sun Z, Xu Y, Liu Y, He W, Jiang Y, Wu F, et al. A survey on federated recommendation systems. IEEE Trans Neural Netw Learn Syst. (2022) 36:6–20. doi: 10.1109/TNNLS.2024.3354924

PubMed Abstract | Crossref Full Text | Google Scholar

7. Kreutz CK, Schenkel R. Scientific paper recommendation systems: a literature review of recent publications. Int J Digital Libraries. (2022) 23:335–69. doi: 10.1007/s00799-022-00339-w

PubMed Abstract | Crossref Full Text | Google Scholar

8. Javed U, Shaukat K, Hameed I, Iqbal F, Alam TM, Luo S, et al. Review of content-based and context-based recommendation systems. Int J Emerg Technol Learn (iJET). (2021) 16:3. doi: 10.3991/ijet.v16i03.18851

Crossref Full Text | Google Scholar

9. Ivchenko D, Staay DVD, Taylor C, Liu X, Feng W, Kindi R, et al. TorchRec: a PyTorch domain library for recommendation systems. In: ACM Conference on Recommender Systems. New York, NY: ACM (2022). doi: 10.1145/3523227.3547387

Crossref Full Text | Google Scholar

10. Hwang S, Park E. Movie recommendation systems using actor-based matrix computations in South Korea. IEEE Trans Computat Soc Syst. (2022) 9:1387–93. doi: 10.1109/TCSS.2021.3117885

Crossref Full Text | Google Scholar

11. Maier C, Simovici D. Bipartite graphs and recommendation systems. J Adv Inform Technol. (2022) 13:249–58. doi: 10.12720/jait.13.3.249-258

Crossref Full Text | Google Scholar

12. Fayyaz Z, Ebrahimian M, Nawara D, Ibrahim A, Kashef R. Recommendation systems: algorithms, challenges, metrics, and business opportunities. Appl Sci. (2020). doi: 10.3390/app10217748

Crossref Full Text | Google Scholar

13. Dhelim S, Aung N, Bouras MA, Ning H, Cambria E. A survey on personality-aware recommendation systems. Artif Intellig Rev. (2021). doi: 10.1007/s10462-021-10063-7

Crossref Full Text | Google Scholar

14. Urdaneta-Ponte MC, Méndez-Zorrilla A, Oleagordia-Ruíz I. Recommendation systems for education: systematic review. Electronics. (2021) 10:1611. doi: 10.3390/electronics10141611

Crossref Full Text | Google Scholar

15. Shi S, Gong Y, Gursoy D. Antecedents of trust and adoption intention toward artificially intelligent recommendation systems in travel planning: a heuristic–systematic model. J Travel Res. (2020) 60:8. doi: 10.1177/0047287520966395

Crossref Full Text | Google Scholar

16. Chakraborty S, Hoque MS, Jeem NR, Biswas M, Bardhan D, Lobaton E. Fashion recommendation systems, models and methods: a review. Informatics. (2021). doi: 10.3390/informatics8030049

Crossref Full Text | Google Scholar

17. Wei G, Wu Q, Zhou M. A hybrid probabilistic multiobjective evolutionary algorithm for commercial recommendation systems. IEEE Trans Computat Soc Syst. (2021). doi: 10.1109/TCSS.2021.3055823

Crossref Full Text | Google Scholar

18. Kanwal S, Nawaz S, Malik MK, Nawaz Z. A review of text-based recommendation systems. IEEE Access. (2021) 9:31638–61. doi: 10.1109/ACCESS.2021.3059312

Crossref Full Text | Google Scholar

19. Yang L, Tan B, Zheng V, Chen K, Yang Q. Federated recommendation systems. In:Yang Q, Fan L, Yu H, , editors. Federated Learning. Cham: Springer (2020). p. 225–39. doi: 10.1007/978-3-030-63076-8_16

Crossref Full Text | Google Scholar

20. Jadidinejad AH, Macdonald C, Ounis I. The Simpson's paradox in the offline evaluation of recommendation systems. ACM Trans Inf Syst. (2021) 40:1–22. doi: 10.1145/3458509

Crossref Full Text | Google Scholar

21. Feng C, Khan M, Rahman AU, Ahmad A. News recommendation systems - accomplishments, challenges & future directions. IEEE Access. (2020) 8:16702–25. doi: 10.1109/ACCESS.2020.2967792

PubMed Abstract | Crossref Full Text | Google Scholar

22. Nawara D, Kashef R. Context-aware recommendation systems in the IoT environment (IoT-CARS)–a comprehensive overview. IEEE Access. (2021) 1:77–91. doi: 10.1109/ACCESS.2021.3122098

Crossref Full Text | Google Scholar

23. Wang H, Cai F, Shi C, Cheng J, Su S, Qiu Z, et al. Positive contrast susceptibility mr imaging using GPU-based primal-dual algorithm. In: 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). Montreal, QC: IEEE (2020). p. 1485–8. doi: 10.1109/EMBC44109.2020.9176223

PubMed Abstract | Crossref Full Text | Google Scholar

24. Wang H, Liang D, Su S, King KF, Chang Y, Liu X, et al. Improved gradient-echo 3D magnetic resonance imaging using compressed sensing and Toeplitz encoding with phase-scrambled RF excitation. Med Phys. (2020) 47:1579–89. doi: 10.1002/mp.13987

PubMed Abstract | Crossref Full Text | Google Scholar

25. Rocco JD, Ruscio DD, Sipio CD, Nguyen PT, Rubei R. Development of recommendation systems for software engineering: the CROSSMINER experience. Empirical Softw Eng. (2021) 26:69. doi: 10.1007/s10664-021-09963-7

Crossref Full Text | Google Scholar

26. Khan HUR, Lim C, Ahmed M, Tan K, Mokhtar MB. Systematic review of contextual suggestion and recommendation systems for sustainable e-tourism. Sustainability. (2021) 13:8141. doi: 10.3390/su13158141

Crossref Full Text | Google Scholar

27. Cabrera-Sánchez JP, de Luna IR, Carvajal-Trujillo E, Villarejo-Ramos AF. Online recommendation systems: factors influencing use in E-commerce. Sustainability. (2020) 12:8888. doi: 10.3390/su12218888

PubMed Abstract | Crossref Full Text | Google Scholar

28. Fu Z, Xian Y, Zhang Y, Zhang Y. Tutorial on conversational recommendation systems. In: ACM Conference on Recommender Systems. New York, NY: ACM (2020). doi: 10.1145/3383313.3411548

Crossref Full Text | Google Scholar

29. Argyriou A, González-Fierro M, Zhang L. Microsoft recommenders: best practices for production-ready recommendation systems. In: The Web Conference. New York, NY: ACM (2020). doi: 10.1145/3366424.3382692

Crossref Full Text | Google Scholar

30. Wang H, Qiu Z, Su S, Jia S, Li Y, Liu X, et al. Parameter optimization framework on wave gradients of Wave-CAIPI imaging. Magnetic Resonance Med. (2020) 83:1659–72. doi: 10.1002/mrm.28034

PubMed Abstract | Crossref Full Text | Google Scholar

31. Qiao Y, Song X, Yan J, Pan W, Chia C, Zhao D, et al. Neurological activation during verbal fluency task and resting-state functional connectivity abnormalities in obsessive-compulsive disorder: a functional near-infrared spectroscopy study. Front Psychiatry. (2024) 15:1416810. doi: 10.3389/fpsyt.2024.1416810

PubMed Abstract | Crossref Full Text | Google Scholar

32. Nawara D, Kashef R. IoT-based recommendation systems – An overview. In: 2020 IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS). Vancouver, BC: IEEE (2020).

Google Scholar

33. Lee D, Gopal A, Park SH. Different but equal? A field experiment on the impact of recommendation systems on mobile and personal computer channels in retail. Inform Syst Res. (2020) 31:3. doi: 10.1287/isre.2020.0922

PubMed Abstract | Crossref Full Text | Google Scholar

34. Hsia S, Gupta U, Wilkening M, Wu CJ, Wei GY, Brooks D. Cross-stack workload characterization of deep recommendation systems. In: IEEE International Symposium on Workload Characteri. Beijing: IEEE (2020).

Google Scholar

35. Yadalam TV, Gowda VM, Kumar VS, Girish DMN. Career recommendation systems using content based filtering. In: International Conference on Communication and Electronics Systems. Coimbatore: IEEE (2020). doi: 10.1109/ICCES48766.2020.9137992

Crossref Full Text | Google Scholar

36. Yeh TY, Kashef R. Trust-based collaborative filtering recommendation systems on the blockchain. Adv Internet of Things. (2020) 10:4. doi: 10.4236/ait.2020.104004

Crossref Full Text | Google Scholar

37. Greenfield MJ, Fobian AD, Fargason RE, Birur B. Optimizing outcomes when treating functional neurological disorder in acute care settings: case reports depicting the value of diagnostic precision and timely and appropriate psychological interventions using an interdisciplinary framework. Front Psychiatry. (2024) 15:1288828. doi: 10.3389/fpsyt.2024.1288828

PubMed Abstract | Crossref Full Text | Google Scholar

38. Cabarkapa S, King JA, Dowling N, Ng CH. Co-morbid obsessive-compulsive disorder and attention deficit hyperactivity disorder: neurobiological commonalities and treatment implications. Front Psychiatry. (2019) 10:557. doi: 10.3389/fpsyt.2019.00557

PubMed Abstract | Crossref Full Text | Google Scholar

39. Alam MT, Ubaid S, Sohail SS, Nadeem M, Hussain S, Siddiqui J, et al. Comparative analysis of machine learning based filtering techniques using MovieLens dataset. Procedia Comput Sci. (2021) 194:210–7. doi: 10.1016/j.procs.2021.10.075

Crossref Full Text | Google Scholar

40. Boroujeni GAA. A trust model for epinion dataset. In: 7th International Conference on e-Commerce in Developing Countries: with focus on e-Security. Kish Island IEEE (2013). p. 1–7.

Google Scholar

41. Verhelst T, Mercier D, Shestha J, Bontempi G. A churn prediction dataset from the telecom sector: a new benchmark for uplift modeling. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer (2023). p. 292–9.

Google Scholar

42. Araujo Costa E, de Paula Ayres-Silva J. Global profile of anemia during pregnancy vs. country income overview: 19 years estimative (2000-2019). Ann Hematol. (2023) 102:2025–31. doi: 10.1007/s00277-023-05279-2

PubMed Abstract | Crossref Full Text | Google Scholar

43. Wilhelmsson D, Gutkin R, Edgren F, Asp L. An experimental study of fibre waviness and its effects on compressive properties of unidirectional NCF composites. Composit Part A: Appl Scie Manufact. (2018) 107:665–74. doi: 10.1016/j.compositesa.2018.02.013

Crossref Full Text | Google Scholar

44. Huang W, Hao F, Shang J, Yu W, Zeng S, Bisogni C, et al. Dual-LightGCN: dual light graph convolutional network for discriminative recommendation. Comput Commun. (2023) 204:89–100. doi: 10.1016/j.comcom.2023.03.018

Crossref Full Text | Google Scholar

45. Sun W, Chang K, Zhang L, Meng K. INGCF: an improved recommendation algorithm based on NGCF. In: International Conference on Algorithms and Architectures for Parallel Processing. Cham: Springer (2021). p. 116–129.

Google Scholar

46. Dai H, Zhu M, Gui X. An improved NCF model in federated recommendation systems. In: 2023 China Automation Congress (CAC). Chongqing: IEEE (2023). p. 8608–14. doi: 10.1109/CAC59555.2023.10450566

Crossref Full Text | Google Scholar

47. Betello F, Purificato A, Siciliano F, Trappolini G, Bacciu A, Tonellotto N, et al. A reproducible analysis of sequential recommender systems. IEEE Access. (2024) 13:5762–72. doi: 10.1109/ACCESS.2024.3522049

PubMed Abstract | Crossref Full Text | Google Scholar

Keywords: AI-driven recommendation systems, healthcare economics, neurological disorders, Dynamic Equilibrium Model, machine learning, cost-effectiveness, policy optimization

Citation: Zhang J, Xiang S and Li L (2025) Economic implications of artificial intelligence-driven recommended systems in healthcare: a focus on neurological disorders. Front. Public Health 13:1588270. doi: 10.3389/fpubh.2025.1588270

Received: 07 March 2025; Accepted: 03 April 2025;
Published: 15 May 2025.

Edited by:

Michael Ekubu Otim, Dubai Medical University, United Arab Emirates

Reviewed by:

Dagna Oriana Constenla, GlaxoSmithKline, United States
Nkosi Nkosi Botha, University of Cape Coast, Ghana

Copyright © 2025 Zhang, Xiang and Li. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Shihui Xiang, Y3JudDg3OUAxNjMuY29t

Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.