Transforming Hospital HVAC Design with BIM and Digital Twins: Addressing Real-Time Use Changes

Abstract

Traditional HVAC designs often struggle to respond promptly and accurately to dynamic changes in complex environments like hospital usage. This paper introduces a novel framework that integrates Building Information Modeling (BIM), digital twin technology, and practical medical processes to transform HVAC design for hospital construction. The framework ensured a smarter (with a reduction of 90% in calculation time and an improvement of 38.20–53.24% in respondence speed) and cleaner environment after identifying and calculating the rational layout of functional areas and optimizing intersecting flow lines. A key innovation of this research was the application of Support Vector Machine (SVM) and deep learning algorithm (Long Short-Term Memory) networks for real-time pedestrian traffic prediction. The implementation was validated through multiple simulations and applications including horizontal and vertical traffic flow and negative pressure analyses for three distinct departments. The findings underline the potential of BIM and digital twins to optimize HVAC systems and hospital design, providing adaptive, data-driven solutions for both routine operations and emergency scenarios. This framework offers a scalable approach for modernizing healthcare infrastructure, ensuring resilience and efficiency in diverse operational contexts.

1. Introduction

The design, construction, and operation of indoor environments are undergoing significant technological transformations, particularly through digitization, which improves comfort, sustainability, and efficiency [1,2,3]. However, heating, ventilation, and air conditioning (HVAC) designs remain challenging as they must balance energy consumption, environmental impact, business or service outcomes, and staff well-being [4,5,6]. In hospital settings, ventilation systems play a critical role in balancing indoor air quality, thermal comfort, infection control, and energy efficiency while meeting health and safety standards and the expectations of patient recovery and staff productivity [7,8,9,10]. Despite advancements in HVAC control mechanisms, traditional systems often rely on reactive strategies that fail to adapt dynamically to real-time environmental changes, leading to inefficiencies and suboptimal performance.

One promising solution lies in the integration of Building Information Modeling (BIM) and digital twins, which allow the real-time monitoring, control, and predictive optimization of HVAC operations. Digital twins create a bidirectional connection between the physical HVAC system and its virtual counterpart, enabling seamless data exchange and real-time adjustments. BIM provides a structured digital representation of the HVAC system, facilitating simulation, automation, and decision-making [11]. When combined with Artificial Intelligence (AI), Computational Fluid Dynamics (CFD) simulations, and the Internet of Things (IoT) (e.g., sensors and mobile devices), digital twins can enable the intelligent adaptation of HVAC systems to real-world conditions [12,13]. However, while digital twins are widely recognized for their potential, their full integration with BIM, CFD, and AI-driven predictive models remains underexplored. There is still a lack of an integrated framework that leverages these technologies to enhance real-time adaptability and operational efficiency [11,14,15]. Existing control strategies for HVAC systems primarily operate on preset thresholds that determine indoor climate settings based on static rules. These approaches struggle with dynamic and unpredictable factors such as variations in occupancy, temperature fluctuations, and external environmental influences [16,17,18,19]. While researchers have proposed AI-driven optimizations and IoT-based monitoring [20,21], a fully integrated BIM–digital-twin framework capable of both real-time analysis and predictive control remains undeveloped. Hence, a proactive framework should be established to incorporate real-time data analytics, CFD simulations for airflow modeling, and machine learning for predictive optimization.

To tackle these issues, this paper proposes a BIM–digital-twin framework that leverages real-time data, AI-based predictive control, and CFD simulations to enhance HVAC system efficiency and adaptability. By integrating these technologies, the proposed framework aims to enable dynamic adjustments based on real-time conditions while optimizing indoor air quality and thermal comfort across hospital environments. Unlike previous studies that focused on BIM, digital twins, CFD, or AI in isolation, this work presents a holistic approach that bridges these technologies in an adaptable system.

2. Literature Review

BIM is beneficial to a project’s lifecycle in multiple ways. For example, integrating BIM and GIS at the city scale has improved management, enhanced efficiency, and reduced costs [22]. Recent technological advancements in BIM and AI-driven applications have reduced project risks and cut the time spent on data retrieval and validation, making project execution more seamless and reliable [23,24,25]. BIM–AI applications have also lowered human and financial expenditures compared to traditional methods [22,25]. Yet, interoperability is still an issue in smart BIM platforms for analytical computations [22,25]. Actionable strategies are urgently needed to translate theoretical advancements into practical solutions for diverse applications [26]. Building on these advantages, researchers have increasingly explored BIM applications for indoor environments, leveraging BIM’s capabilities to optimize spatial analysis, airflow modeling, and real-time monitoring.

BIM applications for indoor environments have gained growing research interest and technical focus. BIM technology supplies precise geometric and semantic building information, enabling CFD simulations to model airflow and temperature distributions [26,27]. CFD results can be combined with machine learning (ML) to predict thermal comfort indices such as the Predicted Mean Vote (PMV) and Predicted Percentage of Dissatisfied (PPD) [28,29,30,31]. BIM can also serve as a knowledge base to enable multi-layered map generation and path planning for unmanned ground vehicles (UGVs), incorporating mobility-based expansions, collision checking, and cost-setting [25,32]. Beyond thermal modeling and airflow analysis, BIM’s integration with IoT technologies has further enhanced its ability to provide real-time insights, fostering data-driven decision-making for energy management and indoor comfort. By integrating IoT sensors, BIM systems can access real-time building performance data for informed energy interventions [33,34]. For example, a common data platform was developed using low-cost IoT sensors and Revit, enhanced by Dynamo and a specific API for dynamic data exchange [33]. This platform visualizes real-time and historical data on indoor conditions (e.g., temperature and luminance) and energy consumption, but BIM technology still needs significant algorithmic refinement, e.g., [35].

Table 1. State-of-the-art examples of BIM for indoor environments (E represents ‘example’).

E	Details	Advantages	Disadvantages and Limitations
[27]	A BIM-enabled, data-driven framework to optimize indoor thermal comfort and reduce energy consumption. Zone connectivity and interzonal airflow.	Provide input for ML models to leverage graph theory’s adjacency matrices and capture spatial features.	Limitations in fully meeting thermal comfort standards, especially across seasonal variations.
[32]	A global path planning system integrating BIM and a physics engine for UGVs in indoor environments. Validation in a university building.	Specific parameters enhance efficiency and adaptability. Improvements in trajectory time, steering jerk, and collision avoidance efficiency. Interoperability. Comprehensive experiments.	Manual intervention, dependence on prior BIM, limited dynamic obstacle handling.
[33]	The integration of IoT sensors with BIM via Revit, Dynamo, and an application program interface is practical and low-cost.	Energy audits with IoT–BIM platforms for retrofit decisions.	Its reliance on proprietary software limits broader adoption. It requires improved handling of data uncertainties.
[35]	A deep learning-based semantic segmentation model improves classification accuracy.	Evaluations on synthetic and real data demonstrate high precision (96–99%) and improved model completeness.	Struggles with close objects and shape variations. Difficulty with curved walls, close object differentiation, and shape variation.

shows the diverse applications of BIM in managing built environments.

Machine learning algorithms have been implemented in BIM digital twins to improve efficiency.

Table 2. Comparison of machine learning algorithms.

Model	Accuracy	Computational Cost and Real-Time Applicability	Dataset Requirements
SVM	High for small-to-medium datasets; struggles with very large datasets	Moderate costs (depends on kernel choice). Suitable for real-time applications with small datasets.	Performs well with small datasets; requires feature engineering.
LSTM	High for sequential data with long-term dependencies	High costs (due to sequential computations). Limited real-time applicability due to processing costs.	Requires a large, labeled dataset to generalize well.
GRU	Comparable to LSTM, sometimes slightly lower	Lower than LSTM (fewer parameters). More suitable for real-time applications than LSTM.	Performs well with medium-sized datasets
Trans.	Very high (especially for large datasets)	Very high (parallel processing but computationally expensive). Limited real-time applicability due to high computational demands.	Requires very large datasets for optimal performance.

summarizes Support Vector Machine (SVM), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and transformer (Trans.) approaches based on accuracy, computational cost, real-time applicability, and dataset requirements. The SVM was selected in this study based on its robustness in handling high-dimensional feature spaces and its efficiency in scenarios with limited labeled training data [36].

While the SVM is typically suited for linearly separable problems, we incorporated kernel functions (e.g., radial basis function) to enhance non-linearity handling [37]. Given that real-time pedestrian movement prediction requires quick and interpretable decision-making, SVM was computationally efficient over deep learning models, particularly in real-time edge computing environments where computational resources were constrained. LSTM networks were selected due to their proven ability to model long-term dependencies in sequential data, a crucial factor for capturing pedestrian movement patterns over time. While GRUs are computationally more efficient, they have difficulties in preserving long-range dependencies [38]. Similarly, transformers offer advantages in parallelization but typically require larger datasets and higher computational resources for effective training, making them less suitable for real-time applications in our study [39]. Despite these advancements, there remains a need to further refine ML algorithms to improve BIM’s adaptability in representing complex indoor environments with higher accuracy and efficiency.

Building Information Modeling and digital twins for real-time control in healthcare and medical facilities enables dynamic monitoring and predictive management [40]. However, the digital twin (DT) prototype presented in the study [40] is primarily a conceptual framework with four theoretical scenarios rather than a fully implemented system. Hence, the technologies require a high level of technical expertise. Sourcing appropriate sensors that meet the DT’s specific requirements can also be difficult. The study [40] lacked real-world implementation due to restricted access to critical healthcare facility data, systems, and equipment. Additionally, bidirectional coordination (a key feature of DTs) may face resistance due to the sensitive nature of hospital operations and concerns over the external control of critical systems. In another example, BIM digital twins integrated real-time sensor data to create a dynamic virtual representation of hospital spaces for infection spread analysis and mobile asset management [41]. The airflow patterns were visualized to identify potential hotspots for infection spread and evaluate the effectiveness of various mitigation strategies. The system’s real-time capabilities allowed for continuous monitoring and prompt adjustments to environmental controls. BIM frameworks for indoor environments should utilize data-driven methods to support applications like real-time HVAC system optimization, predictive maintenance through IoT integration, and exploring design scenarios for energy-efficient building layouts [27]. Current deficiencies include (1) incorporating graph convolutional neural networks to refine spatial feature analysis and improve predictive accuracy [27], (2) automating critical steps using AI (e.g., natural language processing for knowledge extraction) [42], and (3) implementing real-time BIM updating with as-built data integration [42,43,44,45]. While these digital twin applications demonstrate promising advancements, their effectiveness relies heavily on integrating heterogeneous data sources, particularly for critical building systems like HVAC.

To support real-time environmental control and predictive maintenance, BIM and digital twin frameworks must process vast amounts of data from HVAC systems, which draw from multiple sensor inputs to maintain indoor air quality and optimize energy usage. The heterogeneous data sources for HVAC systems include temperature, humidity, pressure, air quality (detecting pollutants, CO₂, or volatile organic compounds), occupancy, airflow, light, and motion [40,41]. For larger and more complex environments like hospitals, data centers, or multi-zoned office buildings, HVAC systems employ advanced control strategies to maintain consistent indoor temperature, humidity, and air quality while optimizing energy consumption by avoiding overcooling, overheating, or unnecessary operation [11,41]. Sensors and controls, integrated with BIM, digital twins, and AI, can significantly enhance the automation, adaptability, and sustainability of HVAC systems [8,10,27,46]. For example, Demand-Based Control uses sensors to continuously monitor parameters like temperature, humidity, CO₂ levels, and occupancy [47]. The system dynamically adjusts its operation to meet real-time needs, reduce waste, and ensure comfort. Zoning and Variable Air Volume (VAV) divide a building into multiple zones with individual controls to provide tailored heating, cooling, and ventilation to different areas based on usage and requirements [46]. Furthermore, predictive control uses AI and ML to analyze historical data, weather forecasts, and occupancy patterns to optimize HVAC operations proactively [48]. Using learning-based thermal comfort and energy consumption models, HVAC systems can integrate energy recovery techniques, such as heat exchangers, to utilize waste heat from one process (e.g., exhaust air) to pre-condition incoming air [42,49].

Although these advanced HVAC strategies enhance energy efficiency and occupant comfort in general buildings, hospital environments require even more sophisticated control mechanisms to meet stringent health and safety regulations. Hospital HVAC systems are more sophisticated than those in homes or standard office buildings due to the unique and critical environmental requirements [8,9,10]. The control strategy of a hospital HVAC system depends on the complexity of the building and its operational needs. Key measures include maintaining pressure environments, using strategic diffuser placement, maintaining thermal comfort, proper airflow directionality (e.g., positive pressure in operating rooms and negative pressure in isolation rooms), and responding to emergency conditions (e.g., fire or system failure) to prioritize safety and operational continuity [43,44,45]. Given the complexity of hospital HVAC systems, CFD emerges as a crucial tool for analyzing airflow behavior, pollutant dispersion, and thermal conditions to ensure optimal indoor air quality and patient safety.

Computational Fluid Dynamics is a numerical tool for predicting fluid behavior, particularly airflow in building interiors. It optimizes ventilation systems to control pollutant dispersion and improve indoor air quality, for example, in hospitals [2,45]. CFD includes Navier–Stokes equations (for momentum transfer), continuity equations (for mass conservation), and energy equations (for heat transfer) [50]. Integrating CFD, ML, and BIM enhances energy efficiency, adaptive design, and sustainable architecture [51,52,53,54]. To further illustrate the role of CFD in optimizing ventilation systems,

Table 3. CFD Details.

Category	Details	Ref.
Definition of CFD	A numerical tool for predicting fluid behavior, particularly airflows in building interiors. Used to optimize ventilation systems, control pollutant dispersion, and improve indoor air quality.	[2,45]
Key equations in CFD	Navier–Stokes Equations: These describe momentum transfer in fluids, relating velocity, pressure, and kinematic viscosity. Continuity Equations: These ensure mass conservation. Energy Equations: These account for heat transfers due to conduction, convection, and radiation.	[50,59]
Challenges of CFD	High computational costs and resource-intensive simulation processes hinder real-time applications and extensive scenario testing.	[51]
Emerging techniques	These reduce iterations and execution time for CFD simulations. They combine CFD with ML for applications of dynamic HVAC control and aerodynamic shape optimization.	[30,52]
Applications of ML with CFD	Optimization: GA and PSO algorithms optimize design parameters like geometry and boundary conditions. Surrogate Models: These use trained ML models to predict airflow in HVAC systems, reducing airborne pathogens and improving air quality.	[55,60,61]
Simulation process	One must define the problem domain, create a mesh, and specify boundary/initial conditions; solve discretized equations iteratively; and visualize (velocity fields and pressure distribution) and analyze data.	[62]

presents key computational techniques for environmental analysis through numerical modeling. Incorporating sensor data improves CFD’s predictive capabilities, adaptability, and analytical robustness [55]. Integrating CFD, ML, and BIM enhances sustainability and energy efficiency, improving design and performance while minimizing reliance on physical prototypes [53,54,55,56,57,58].

Knowledge gaps persist in advanced control strategies for hospital HVAC systems, particularly in maintaining pressure differentials, ensuring redundancy and fault tolerance, enabling real-time monitoring and alerts, and meeting stringent standards such as ASHRAE 170 [63] and CDC infection control guidelines. Current challenges include the difficulty of accurately predicting indoor environmental changes due to fluctuating factors like patient loads and activity levels. While BIM offers static insights and CFD focuses on dynamic scenarios, their integration with ML for real-time system optimization remains rudimentary [57,58]. Issues such as computational limitations, inadequate real-time frameworks, and the lack of seamless data exchange protocols hinder their application in hospital environments [62]. Research addressing these challenges is vital to enhance environmental systems’ efficiency and responsiveness. This paper addresses the following research question: “How can BIM and digital twins be used to respond to real-time system usage changes?” The corresponding research objectives are to (1) develop real-time control strategies to optimize pressure differentials, redundancy, and fault tolerance in hospital HVAC systems; (2) design and validate a framework integrating BIM, CFD, and pattern identification for indoor environmental conditions; (3) build a framework for workflow and traffic flow simulations to optimize spatial layouts and airflow patterns within hospital environments; (4) design practical support for real-time feedback and simulation, ensuring accurate adjustments during operational shifts; and (5) implement the BIM-based hospital digital twin (DT) utilizing Support Vector Machines (SVMs) for computational efficiency and Long Short-Term Memory (LSTM) networks for predicting complex flow patterns, enabling streamlined design analysis, and enhancing system adaptability.

3. Methodology

Figure 1 shows the proposed solution of the BIM–digital-twin framework, including creating BIM models and installing IoT and sensors. System parameters like airflow and occupancy are defined, and the database for the digital twin is established. Next, the datasets and CFD simulation outputs are used for machine-learning purposes to predict key performance metrics. The framework uses SVM and LSTM for the analysis and pattern extraction of IoT and sensor data. Equation (1) is the Navier–Stokes equation of the CFD surrogate model for the motion of fluid substances, expressing Newton’s Second Law. Equation (2) is a continuity equation to ensure that mass is conserved in the fluid flow. Equation (3) is in an energy-based form, which uses internal, kinetic, and heat energy types for energy conservation within a fluid system, which is appropriate for HVAC airflow. Using the relationship between internal energy, temperature (T), and specific heat ($c_v$), Equation (4) couples the energy equation with the Navier–Stokes equation to describe the relationships among temperature variations, fluid density, viscosity, and velocity. Equation (5) simplifies Equation (4) for incompressible flows.

$$\rho {\displaystyle \frac{du}{dt}}+\rho \left(u·\nabla \right)u=-\nabla p+\mu {\nabla }^{2}u+f$$

(1)

$${\displaystyle \frac{d\rho }{dt}}+\nabla ·\left(\rho u\right)=0$$

(2)

$$\rho {\displaystyle \frac{de}{dt}}+\rho \left(u·\nabla \right)e=-\nabla ·q+\mathsf{\Phi }+f·u$$

(3)

$$\rho c_v{\displaystyle \frac{dT}{dt}}+\rho c_v\left(u·\nabla \right)T=k{\nabla }^{2}T+\mathsf{\Phi }$$

(4)

$$\rho c_p{\displaystyle \frac{dT}{dt}}+\rho c_p\left(u·\nabla \right)T=k{\nabla }^{2}T$$

(5)

Equation (6) is the optimization problem of the SVM algorithm in a primal form for a dataset ($x_i,y_i)$, where $x_i\in {\mathbb{R}}^n$ and $y_i\in \{-\mathrm{1,1}\}$. Equation (7) is the decision function, and Equation (8) is the kernel trick for the algorithm’s non-linear classification. SVM is a supervised ML algorithm used for classification and regression tasks. The key concept is to find a hyperplane that best separates data points into different classes. If the data is not linearly separable, SVM uses a kernel function like Equation (8) to map input data to a higher dimensional space: $\mathrm{K}\left(x_i,x_j\right)=\mathbf{F}\left(x_i\right)·\mathbf{F}\left(x_j\right)$

$$\underset{w,b,\xi }{\min }{\displaystyle \frac{1}{2}}{\left|\left|w\right|\right|}^2+C{\sum }_{i=1}^N{\xi }_i, \mathrm{subject} \mathrm{to}: y_i(w·x_i+b){\ge 1-\xi }_i, {\xi }_i\ge 0$$

(6)

$$g\left(x\right)=sign(w·x+b)$$

(7)

$$\begin{gathered} \mathrm{K}\left(x_i,x_j\right)=x_i·x_j, \mathrm{if} \mathrm{linear} \\ \mathrm{K}\left(x_i,x_j\right)={({\gamma x}_i·x_j+\tau )}^d, \mathrm{if} \mathrm{polynomial} \\ \mathrm{K}\left(x_i,x_j\right)={exp(-\gamma {\left|\right|x}_i-x_j||}^2), \mathrm{if} \mathrm{a} \mathrm{Radial} \mathrm{Basis} \mathrm{Function} \end{gathered}$$

(8)

The following equations allow LSTM networks to provide scalability and accuracy in predicting complex flow patterns. Equation (9) is the forget gate to determine what information to discard from the cell state. Equation (10) is the input gate to control new information to add to the cell state, where $\sigma$ is the Sigmoid activation function. Equation (11) is the Cell State Update function based on forget and input gates, where tanh is a hyperbolic tangent activation function. Equation (12) is the output gate. Equation (13) is the Hidden State function. W, U, and b are learnable variables for input, recurrent connections, and biases. Moreover,

Table A1. Nomenclature.

Term	Meaning
$e$	$\mathrm{specific} \mathrm{total} \mathrm{energy} (e=e_{internal}+{\displaystyle \frac{1}{2}}{\left|u\right|}^2$)
$\rho$	fluid density
$u$	velocity vector (u, v, w),
t	time
p	pressure
k	thermal conductivity
w	weight vector defining the hyperplane
b	Bias term
$q$	$\mathrm{heat} \mathrm{flux} (q=-k\nabla T$) by Fourier’s law of heat conduction
$\mu$	dynamic viscosity
$\mathsf{\Phi }$	viscous dissipation term
${\xi }_i$	slack variables for soft margin classification
C	regularization parameter controlling the trade-off between maximizing the margin and minimizing classification error
${\nabla }^{2}u$	the viscous term of Laplacian of velocity
$f$	$\mathrm{external} \mathrm{forces} (\mathrm{e}.\mathrm{g}., \mathrm{gravity}), (f·u)$ calculates the work carried out by external forces
$g\left(x\right)$	decision boundary function of SVM
$\rho {\displaystyle \frac{du}{dt}}$	unsteady inertia, representing the rate of change of momentum
$\rho \left(u·\nabla \right)u$	$\mathrm{advection} \mathrm{describes} \mathrm{the} \mathrm{transport} \mathrm{of} \mathrm{momentum} \mathrm{by} \mathrm{the} \mathrm{fluid} \mathrm{itself}; \mathrm{and} \mathrm{for} \mathrm{incompressible} \mathrm{flows}, \left(u·\nabla \right)=0$
$-\nabla p$	pressure gradient or driving force
$\mu {\nabla }^{2}u$	viscous forces and account for internal friction in the fluid
$\left|\left|w\right|\right|$	the norm of w, which is, by default, second-order (Euclidean) norm.

lists the details and explanations of the variables in the equations of the framework.

$$f_t=\sigma (W_f{x}_t+U_f{h}_{t-1}+b_f)$$

(9)

$$\begin{gathered} i_t=\sigma (W_i{x}_t+U_i{h}_{t-1}+b_i) \\ \widetilde{C_t}=\tanh (W_c{x}_t+U_c{h}_{t-1}+b_c) \end{gathered}$$

(10)

$$C_t=f_t\odot {(C}_{t-1}+i_t⨀\widetilde{C_t})$$

(11)

$$o_t=\sigma (W_o{x}_t+U_o{h}_{t-1}+b_o)$$

(12)

$$h_t=o_t\odot {\tanh (C}_t)$$

(13)

Equations (6)–(8) are for finding the optimal decision boundary between classes by maximizing the margin. They work well for smaller datasets or problems with clear class boundaries. The kernel trick enables it to handle non-linear relationships. LSTM captures long-term dependencies in sequential data by controlling how information flows through the network with its gating mechanisms.

Table 4. Parameter selection.

Model	Parameter	Explanation
SVM	Kernel Type	Experimented with different kernel functions (linear, polynomial, and radial basis function (RBF)) and selected the RBF kernel to handle non-linear relationships.
	Regularization parameters	A grid search technique was applied over a predefined range (e.g., 0.1 to 100) to balance the trade-off between model complexity and accuracy.
	Gamma	Using cross-validation to ensure an appropriate decision boundary.
LSTM	layers and hidden units	Two-layer LSTM with 128 hidden units per layer as it provided the best balance between complexity and performance
	Learning rate	Adaptive learning rate optimization using the Adam optimizer was applied, with initial values ranging from 0.001 to 0.0001.
	Dropout rate	A dropout rate of 0.2 was incorporated to prevent overfitting.
	Batch size and epochs	Batch sizes 32 and 64 were tested, and 32 was used for validation accuracy. The quantity of training epochs was selected using early stopping to avoid overfitting.

lists the parameter selections [64,65,66,67,68]. The training process includes (1) data preprocessing: filtering noisy sensor data, normalizing inputs, and segmenting time-series sequences; (2) training datasets derived from open-access datasets [64,65,69,70,71], incorporating sensor-based pedestrian movement records collected in hospital environments; and (3) performance metrics: we used Root Mean Square Error (RMSE) to evaluate model accuracy and robustness.

4. Digital Twin for HVAC System

Under the assumption that precise modeling is required for an indoor environment of non-linear building dynamics with complex airflow patterns, this research analyzed real-world and simulation datasets from application scenarios involving hospital lifecycle. The BIM model developed for this study was based on the Emergency Medical Center Project of the Shanghai Public Health Clinical (SPHC) Campus in Jinshan District, where four new buildings were newly built, including a medical complex, a multi-functional complex, an energy center, and a liquid oxygen station. The multi-functional complex rooms are for medical observation and training, including rooms such as basic teaching rooms, special teaching rooms (student dormitories, libraries, academic lecture halls, conference functions, training bases, etc.), and basic scientific research rooms. The hospital project was officially launched on 21 March 2022, with a focus on the overall situation of epidemic prevention and control and the construction of the city’s public health system. Eight hundred (800) new beds will be added when the project is completed in March 2026. The construction area is 152,777 square meters, including 106,542 square meters of above-ground construction area and 46,235 square meters of underground construction.

Table 5. Data collection and modeling.

Step	Objective	Process	Tools
1. Data Collection	Collecting data to build and validate HVAC digital twin models for hospital construction.	Site visits and BIM data acquisition: Gathering BIM models of the 800-bed full negative pressure emergency medical center; obtaining layouts, dimensions, material specifications, and airflow data. Operational data collection: Using IoT sensors to collect real-time HVAC system performance data, including airflow rates, temperature, and pressure; gathering pedestrian traffic patterns through surveillance data and RFID tracking in hospital functional areas. Medical process data: Mapping connections between hospital functional areas and identifying flow lines (e.g., patients, medical staff, and equipment).Historical data: Including data from past operations for variabilities in “peacetime” and “epidemic” modes [65].	IoT sensors, BIM software (Autodesk Revit v2024, Navisworks v2024), pedestrian tracking systems, and manual surveys.
2. Numerical Model Development	Integrating BIM, digital twins, and flow simulations.	HVAC simulation setup: Using Simcenter STAR-CCM+ for CFD simulations to analyze airflow dynamics in 3 departments; simulating negative pressure requirements. Pedestrian traffic prediction: Applying SVM for initial traffic predictions using limited datasets and comparing results with LSTM for large-scale data; implementing TensorFlow and PyTorch for deep learning models. Integration with BIM: Creating parametric models in Revit to link HVAC layouts with pedestrian flow simulations.	Revit, Navisworks. STAR-CCM. Traffic prediction using TensorFlow, PyTorch v2.0, and Scikit-learn. Data processing using Python v3.13.0 and MATLAB R2024b.
3. Main Modeling Process	BIM and Digital Twin	BIM-based modeling: Building a 3D parametric model of the hospital; adding functionality for switching between routine and epidemic modes. Negative pressure simulation: Conducting airflow simulations for three departments to meet negative pressure requirements during epidemics. Traffic flow simulation: Simulating pedestrian and equipment flow under different scenarios to identify bottlenecks and optimize layouts. Integration of digital twin: Developing a real-time feedback loop linking IoT sensors, CFD models, and the BIM platform for operational adjustments. Smart platform application: Creating a centralized dashboard to visualize HVAC performance, pedestrian flows, and scenario simulations.	Pathfinder for evacuation simulations.
4. Data Analysis and Interpretation	Data-driven decision-making	Analysis of simulation results: Assessing the performance of HVAC systems under different flow conditions (e.g., emergency vs. normal); using SVM and LSTM models to predict and compare pedestrian flow impacts on HVAC requirements. Validation: Comparing simulation results with real-world measurements from IoT sensors; adjusting models iteratively to improve accuracy. Interpretation: Identifying key factors for HVAC performance and layout efficiency; providing actionable recommendations for optimization.	MATLAB, Python (Pandas, NumPy, Matplotlib), Power BI for visualization.

lists the software details, basic information about the numerical model, and the main modeling process of this research. This methodology uses BIM and digital twin technologies while leveraging AI to address dynamic changes in hospital environments. The simulation models integrate both routine operations and specialized measures for epidemic prevention. Pressure, temperature, and indoor air quality (IAQ) sensors were installed to collect data. The sensor sampling frequency was 60 Hz.

4.1. Models

The numerical models (Step 2 in Table 5) include the AI-enabled pedestrian traffic and CFD analysis of negative-pressure areas to incorporate the shifts in demand for medical services driven by the changes in medical diagnosis and treatment under various situations. The operation plans are different for routine situations, conventional epidemic management, and pandemic prevention and control. Specifically, the operational departments of conventional epidemic management include the emergency department, radiology department, operating rooms, critical care department, inpatient department, and Peripherally Inserted Vascular Access Score office. The closed areas include the outpatient clinics, logistics support, public spaces within the hospital, and related open spaces.

Additionally, precise adjustments to the department operation plans can be made according to the actual epidemic situations, for example, when a pandemic happens. More accurate modeling and simulations were carried out to streamline the closed-loop scope and fine-tune zoning management while addressing the personalized medical diagnosis and treatment needs of special patients. During a pandemic, the following departments and units are in operation, resulting in a total of 12 distinctive operation plans. The north side of the hospital includes the Radiation Department (on the first floor or 1F); Hemodialysis, Blood Testing, and Transfusion (2F); Pathology (3F); inpatient wards (4F–13F); and basement (laundry and sewage treatment, nuclear medicine, and radiotherapy center). The south side includes ICUs (4F), operation rooms (5F), and the Central Sterile Services Department (5F). The following departments and units of the south side are closed when a pandemic happens, resulting in a total of 18 operation plans. On 1F are outpatient areas, the emergency department, registration and payment areas, outpatient/inpatient pharmacy, the intravenous preparation center, and blood collection reservation areas on the south side. 2F has an outpatient area and emergency department. 3F has an ultrasound, a functional examination department, a respiratory endoscopy, and an outpatient area. 4F has an outpatient area and an interventional center. The Endoscopy Center on the 3F of the north side and all the public corridors and spaces are closed.

Multiple rounds of evaluations were conducted to assess the tasks and operational needs of all hospital departments. After the operation plans for each department of the Emergency Medical Center Project were meticulously developed, three representative departments (i.e., wards, the Radiology Department, and operation rooms) were selected for in-depth simulations in this study. Each representative department received secondary medical process streamlining simulations, airflow simulations, and evacuation modeling to optimize the design and enhance the overall experience of patients and medical teams. This modeling process helped reduce the computation load from 18 + 12 = 30 models and simulations of operation plans to three ones, an improvement of 90% in time and efficiency.

4.1.1. Construction of Airflow Simulation Model for Negative-Pressure Ward

Figure 2 shows that a transferrable area has been set up for each floor wing to accommodate the function changes. All the patient rooms are convertible to negative pressure. Areas 1, 2, and 3 are pass-through changing and buffer rooms that meet basic usage requirements for routine operation plans of the ward suite. Area 1 is a buffer room, Area 2 is the first changing room, and Area 3 is another buffer room. These three areas are for routine operations. Area 4 is the second changing room, Area 5 is the third buffer room, and Area 6 is the dressing room. These three areas can be used for offices and storage in routine operations and are converted to designated usages when the hospital switches to epidemic prevention and control mode. The process of taking off personal protective equipment is thus: hand hygiene → taking off shoe covers → hand hygiene → taking off isolation clothing and outer gloves → hand hygiene → taking off goggles → hand hygiene → taking off protective clothing and inner gloves and boot covers → hand hygiene → taking off protective masks → hand hygiene → taking off hats. Hence, the isolation clothing and outer gloves are taken off in Area 2, and the protective gears are taken off in Area 4.

Figure 3a depicts the functional layout of negative-pressure patient rooms in a ward unit of six patient rooms. To ensure safety and containment in respiratory infectious disease wards (including negative-pressure wards and isolation wards), a directional airflow toward the bathroom must be maintained, with a relative pressure difference of no less than 5 Pa between adjacent and connected rooms. The hierarchy of negative pressure levels descends in the following order: bathroom, patient room, buffer room, and inner corridor (e.g., patient walkway and medical team walkway). The hospital’s mechanical air supply and exhaust system has been designed to progressively reduce air pressure from clean areas to semi-contaminated areas and finally to contaminated areas. Clean areas are maintained as positive-pressure zones while contaminated areas operate as negative-pressure zones. In clean areas, the air supply volume must exceed the exhaust volume, whereas in contaminated areas, the exhaust volume must surpass the air supply volume to ensure effective containment. Figure 3b shows the construction of an airflow simulation model for a negative-pressure ward and an ethylene glycol heat recovery system. Figure 3c shows circulation airflows.

The items in Figure 3d are explained here: 1—the direction of intake air, 2—the direction of intake air, 3—buffer room, 4—the restroom, 5—the main air conditioning supply outlet, 6—the secondary air conditioning supply outlet, 7—the patient room, 8—airflow direction, 9—exhaust vent, and 10—the direction of exhaust air. The device details are thus: 11—Venturi valve with sensors for precise and accurate airflow control (e.g., [66], 12—high-precision volume flow control (e.g., [67]), 13—nursing station monitoring unit, 14—connected control app showing the air pressure, 15—wall mounting temperature and humidity sensor, and 16—airflow sensor.

Venturi valves regulate airflow for room pressurization and fume hood containment, ensuring proper ventilation and protecting against hazardous airborne substances. Commonly used in environments like laboratories or hospitals with variable exhaust system pressures, these valves maintain consistent airflow by rapidly adjusting to pressure changes. An actuator connected to a lever arm positions a spring and cone assembly within the valve’s hourglass-shaped structure, passively balancing forces to control airflow without requiring active adjustments. Airflow is determined by a potentiometer measuring the lever arm’s position, factory-calibrated for accuracy, eliminating the need for direct airflow measurement devices and reducing contamination risks [66]. A digital volume control system with High Precision Drive chips is programmable for sensitive and precise volume flow and pressure control, essential for critical environments like hospitals and clean rooms. The actuator enables damper blade movement within 3 s for a 90° turn, with a precision of 0.1°, significantly improving airflow control accuracy. The digital interface between the controller and the actuator supports two-way communication, allowing the monitoring of the operating status, the damper position, and error messages from the connect software. This reduces manual maintenance and enhances system efficiency. Features include low power consumption, reduced wear, plug-and-play functionality, and compatibility with protocols such as LonWorks, Modbus, and BACnet, ensuring seamless integration with building management systems [67].

4.1.2. Radiology Department: Multi-Functional and Switchable Space Utilization Strategy

The Radiology Department carries out critical functions in diagnosing and monitoring diseases, particularly during epidemics or pandemics when rapid and accurate imaging is crucial for assessing patient conditions, tracking disease progression, and guiding treatment plans. This department supports high patient throughput and is integral to managing infectious disease outbreaks through timely and precise imaging services. Pedestrian traffic in the Radiology Department is significantly higher than in ward areas or operating rooms due to its central role in diagnostics, requiring frequent access by various personnel and patients. Imaging procedures involve staff movement for equipment operation, patient preparation, and image analysis, contributing to increased traffic compared to more restricted areas like operating rooms.

Typically, the department includes essential dressing rooms, while other areas, such as film reading offices and warehouses, are planned as flexible auxiliary spaces (see Figure 4a). During an epidemic, functional rooms are restructured to align with infection control requirements, including designated areas for dressing, entry, buffering, first and second gown removal, and exit buffering. These modifications ensure compliance with airflow organization and zoning management standards, meeting isolation and protection needs. In epidemic scenarios, non-essential sanitary passage areas are repurposed as medical protection spaces to enhance infection control measures (see Figure 4b). During normal operations, these areas revert to auxiliary functions. This flexible design optimizes the department’s usability, ensuring efficient doctor-patient flow while maintaining stringent isolation and safety protocols during outbreaks.

4.1.3. Simulation of Airflow in Negative-Pressure Operating Rooms

The surgical area is divided into three zones, a negative-pressure surgical zone, a positive-pressure surgical zone, and an interventional surgical zone, to achieve regional divisions for different periods and different usage situations. The upper part of Figure 5 shows that patients with non-respiratory infectious diseases are admitted during routine operations. During the epidemic, only the negative-pressure operating area is open to admit patients with respiratory infectious diseases (see the lower part of Figure 5).

4.2. Parameters and Calibrations

This study implemented an SVM algorithm to analyze pedestrian traffic data in the Radiology Department, particularly during epidemics or pandemics. The application began with collecting data from surveillance videos to monitor pedestrian movement. Advanced computer vision techniques, such as object detection and tracking, extracted features like walking speeds, distances traveled, and time spent in specific areas. These processed data formed the input for SVM, enabling detailed behavioral analysis. Then, features were extracted and classified to categorize pedestrian behaviors into predefined patterns. These patterns included routine movements, congestion hotspots, and violations of infection control protocols, such as lingering in restricted zones. The SVM models effectively differentiated between normal and abnormal flow patterns in various scenarios by analyzing features like velocity vectors, trajectory shapes, and time-series data. The iTCM Datasets [69], Room Climate Datasets [70], and Indoor Temperature Data Collection for Machine Learning Climate Control datasets [71] were used to train the SVM and LSTM models. The noise filtering method of outlier removal was performed. The [69] datasets were collected from tropical regions like Singapore, which has potential geographical, demographic, and seasonal biases. The [70] datasets have potential occupancy, activity, and sensor placement biases. The [66] datasets have potential sensor calibration, temporal, and environmental biases. Using them together can diversify data collection and enhance generalizability. The Root Mean Square Error (RMSE) of the SVM predictions was 0.654. The RMSE of the LSTM predictions was 0.654. See Figure A1 in Appendix A for further performance details of SVM and LSTM predictions. The analysis suggests that this combined SVM–LSTM approach could be valuable for real-time HVAC control and planning in hospital environments [68].

Figure 6 and

Table 6. Route comparison of the Radiology Department (P is ‘patient’ and MS is ‘medical staff’).

Routine				Epidemic
	Distance (m)	Speed (m/s)	Duration (s)		Distance (m)	Speed (m/s)	Duration (s)
P	96	1.25	77	P	45	1.25	36
MS	89	1.48	60	MS	55	1.48	37

explain how integrating SVM with pedestrian traffic and airflow simulation models enhances the decision-making process. Simulation tools can model the impact of design changes such as repositioning dressing rooms or adding buffer zones. The SVM, in turn, validates these simulations by comparing them to real-world data, ensuring that the proposed adjustments are practical and effective. Table 6 compares the parameters for infection control such as average walking speeds, density in specific areas, and peak congestion times. The estimation and analysis of pedestrian traffic inform decisions about implementing one-way traffic systems, adjusting room capacities, or optimizing the placement of functional spaces to minimize risk. Particularly, by analyzing patterns classified by SVM and integrating them with simulation results, real-time decision-making is enhanced. For example, during an epidemic, the SVM can identify overcrowding near imaging rooms or bottlenecks in critical areas, prompting timely interventions. Additionally, SVM models can continuously learn from new data, adapting to evolving traffic patterns to ensure the long-term efficiency of infection control protocols. This combined approach of the SVM and simulation provides a robust, data-driven framework to optimize the Radiology Department’s layout and operational processes. It enhances the safety of patients and staff while maintaining efficient departmental workflows.

Implementing multi-functional and switchable spaces significantly influences the efficiency and safety of hospital workflows during epidemics (see Table 6 for quantitative impacts). By optimizing the flow line, the distance patients need to travel is reduced from 96 m to 45 m, a reduction of 53.13%. This minimizes physical exertion for patients and reduces the risk of cross-contamination by limiting movement through high-traffic areas. Similarly, the distance traveled by medical team members is reduced by 38.20%, allowing staff to allocate more time to critical patient care and reduce their exposure to potential hazards. In addition to spatial efficiency, the duration of travel is also significantly shortened. Patients experience a 53.24% reduction in travel time, which enhances their overall experience and reduces waiting times. For medical staff, a 38.33% reduction in travel duration improves workflow efficiency, enabling quicker response times and better resource management during peak demand periods. These results demonstrate how reconfigurable spaces and optimized flow lines enhance operational efficiency and improved safety while minimizing the risk of infection transmission for patients and medical staff.

Figure 7 shows the evacuation simulations of the surgical floor space whereas Figure 7a shows the evacuation simulation results of the 5F surgical floor during routine operations. The total number of people was 184, including 10 surgical patients and 174 family members and medical staff. The total time of evacuation was 166.3 s. Figure 7b shows the evacuation simulation results during an epidemic. The total number of people was 220, including 16 surgical patients and 204 family members and medical staff. The total time of evacuation was 178.8 s. The congestion areas are labeled in circles. The evacuation congestion areas are different in different epidemic situations. It is necessary to add evacuation signs on the ground and guide people to evacuate to different fire stairs to avoid congestion and trampling.

4.3. Model Verification with Field Test Results

Figure 8 presents the airflow simulation analysis of negative-pressure wards under two scenarios. Figure 8a depicts airflow velocity during routine operations, where the relative velocity in the patient room ranges from 0.000353 m/s to 0.323 m/s. Less than 5% of the room’s cell area exhibits a relative velocity at the upper range of 0.323 m/s.

In contrast, Figure 8b illustrates airflow velocity during epidemic-control conditions, with a velocity range of 8.66 × 10⁻⁵ m/s to 0.431 m/s. Over 10% of the cell area in this scenario demonstrates relative velocities of 0.4 m/s or higher, indicating a more dynamic airflow environment. Figure 8c shows the airflows during routine operations, and Figure 8d shows the airflows under epidemic control. The airflow patterns of both scenarios are similar. In both cases, airflow from the supply port rapidly descends to the ground and disperses laterally, forming prominent vortices on the left and right sides of the room. These vortices eventually direct the airflow toward the exhaust port, completing the circulation pattern. Notably, beds situated near the sides of the air supply port experience greater impact from the airflow, likely due to their proximity to the high-velocity regions immediately downstream of the supply port. These findings highlight critical considerations for airflow design in negative-pressure environments to ensure effective containment and minimize potential risks to occupants.

The Radiology Department operates negative-pressure examination rooms that function during both routine and epidemic conditions, ensuring appropriate ventilation and infection control. The airflow dynamics in these rooms are characterized by variations in cell-relative velocity magnitudes, as depicted in Figure 9. During routine operations (Figure 9a), the cell-relative velocity magnitude ranges from 0.000422 m/s to 0.129 m/s, indicating a relatively stable airflow with minor variations. During epidemic operations (Figure 9b), the range increases from 0.000739 m/s to 0.295 m/s, reflecting a higher airflow rate designed to meet stricter infection control requirements. Despite differences in flow rates, the indoor airflow directionality remains consistent under both routine and epidemic conditions (see Figure 9c,d). This suggests that the ventilation design ensures reliable airflow patterns irrespective of the operational context. The slightly higher negative pressure observed during epidemics enhances the room’s capacity to contain airborne pathogens, thereby reducing the risk of cross-contamination.

The simulation of airflow in a negative-pressure operating room is shown in Figure 10. According to the characteristics of routine time (Figure 10a,c) and epidemic (Figure 10b,d), the ventilation frequency is adjusted to meet the various treatment tasks. The cell-relative velocity magnitude ranges from 2.46 × 10⁻⁶ m/s to 0.0298 m/s, indicating a relatively stable airflow with minor variations. During epidemic operations (Figure 10b), the range increases from 9.9 × 10⁻⁶ m/s to 0.0532 m/s, reflecting a higher airflow rate designed to meet stricter infection control requirements.

In normal times, the ventilation frequency is 3 times/h; in epidemic times, it is 12 times/h. The exhaust volume should be greater than the supply volume. The airflow in the operating room flows out from the air supply port, passes through the operating table, and then spreads out in all directions. When the airflow spreads, a vortex is formed between it and the wall and finally flows out from the exhaust port.

Table 7. Airflow features.

Features	Wards	Radiology Dept.	Operating Room
Airflow velocity in routine time	0.000353 to 0.323 m/s	0.000422 to 0.129 m/s	2.46 × 10−6 m/s to 0.0298 m/s
Airflow velocity in epidemic	8.66 × 10−5 to 0.431 m/s	0.000739 to 0.295 m/s	9.9 × 10−6 m/s to 0.0532 m/s

presents a comparative analysis of airflow characteristics in three hospital areas, the wards, the Radiology Department, and the operating rooms, during routine and epidemic-control conditions. Significant differences exist in airflow velocities and the proportion of strong airflow areas, critical for maintaining clean and pathogen-free environments. In the wards, the increase in velocity and high-airflow zones enhances the removal of airborne contaminants, contributing to improved infection control during epidemic periods. The system’s ability to adapt to these conditions ensures effective pathogen containment and circulation throughout the wards.

The Radiology Department exhibits a lower velocity range during routine operations and epidemic control. This suggests that the airflow management system in this department focuses on maintaining stable, moderate circulation suitable for radiology equipment and procedural requirements. However, the lower percentage of strong airflow zones may necessitate additional strategies, such as localized ventilation, to ensure adequate pathogen removal during epidemics.

The operating rooms demonstrate significantly lower airflow velocities during routine operations and epidemic control. Despite the low velocities, more than 10% of the area consistently exhibits strong airflow in both scenarios. This design likely reflects the critical need for laminar flow and controlled circulation to minimize airborne contamination during surgical procedures. The consistent presence of high-velocity zones ensures that contaminants are swiftly removed, maintaining a sterile environment essential for patient safety.

5. Discussion

Three distinctive departments were selected in this study to examine the horizontal and vertical traffic flow simulations. In the route comparison analysis, the reduction in travel distance was in the range of 38.20% to 53.13%, and the reduction in travel time was in the range of 38.33% to 53.24%, which improved the operation efficiency. The negative pressure simulations showed that increased airflow velocities were achieved to enhance the removal of airborne contaminants by ensuring that pathogens were rapidly transported toward the exhaust port, where they were effectively expelled. Additionally, the prominent vortices observed on both sides of the room in Figure 8c,d further aided in dispersing airborne particles, preventing localized stagnation zones where pathogens could accumulate. The similarity in airflow patterns under both scenarios indicates that the system was consistently designed to prioritize effective circulation. Further results confirmed that beds near the air supply port were more exposed to high-velocity airflow, which reduced the local pathogen load around these critical areas. However, this also necessitated careful design to balance air distribution across the room and ensure uniform protection for all occupants. Increasing the number of high-velocity regions during epidemic-control operations and maintaining consistent airflow circulation patterns ensures a significant reduction in pathogen concentrations, leading to a cleaner and safer hospital environment. These findings underscore the importance of dynamic airflow management in infection control strategies.

Regarding airflow velocity, the wards experienced the highest velocity range during epidemic control (up to 0.431 m/s), followed by the Radiology Department (up to 0.295 m/s) and the operating room (up to 0.0532 m/s). These variations reflect the different functional priorities and design requirements of each space. Strong airflow zones exist because the wards and operating rooms exhibited increased strong airflow areas during epidemic conditions (>10%), which can enhance pathogen control.

The adaptability of the BIM–digital-twin framework in more complex hospital layouts lies in its ability to integrate real-time data, predictive modeling, and simulation-based optimization across diverse functional areas. Hospitals with intricate layouts, such as multi-wing facilities or high-rise medical centers, present additional challenges in airflow management, energy distribution, and patient flow. The proposed framework can address these complexities by leveraging multi-scale BIM modeling, which allows for the hierarchical decomposition of spatial data at the building, system, and equipment levels. By integrating reinforcement learning and dynamic simulations, the framework can adapt to varying occupancy levels, infection control requirements, and emergency scenarios. Moreover, sensor-driven feedback loops enable real-time adjustments to HVAC and ventilation systems, ensuring that airflow dynamics remain optimized despite structural variations. By utilizing machine learning algorithms such as LSTM and the SVM, the framework can predict congestion points and adjust pathways accordingly, improving operational efficiency and patient safety. This adaptability makes the BIM–digital-twin framework scalable and applicable to a wide range of hospital configurations, from small regional clinics to large metropolitan healthcare complexes.

6. Synthesis with Other Literature

This paper leverages real-time sensor data and ML to enable adaptive HVAC control for operational efficiency and energy management. Compared to other BIM indoor-environmental models [27,42], this BIM–digital-twin model minimizes manual intervention and supports decisions based on automatically updated sensor data, enhancing HVAC system performance. The model adapts to dynamic objects or evolving layouts (e.g., between routine and epidemic modes), offering actionable strategies for optimizing healthcare environments. Additionally, the combination of simulation-based resource planning and predictive modeling introduces a new standard for designing strategically aligned healthcare facilities.

The research bridges technological innovation with practical applications in public health infrastructure. In a recent study, Wang et al. [64] employed metrics such as Euclidean distance, the Dice coefficient, and a force-directed graph algorithm for layout selection and optimization and precise building region segmentation [64]. In comparison, the framework in this paper merges ML and CFD to address real-time factors like airflow, energy efficiency, and layout optimization. By incorporating reinforcement learning and dynamic simulations, the approach provides a more adaptive and holistic solution to hospital design challenges. Compared to traditional rule-based and simulation-driven methods (e.g., [35,42,64]), this framework improves efficiency by automating adaptive adjustments based on real-time data, accuracy by leveraging deep learning models to optimize HVAC performance and airflow predictions, and adaptability by incorporating dynamic simulations that respond to fluctuating conditions within healthcare environments. Technology implementations such as the SVM and LSTM algorithms for patient flow optimization and ventilation modeling emphasize on the framework’s precision and sustainability.

This work underscores the importance of a forward-thinking vision for modern healthcare infrastructure, highlighting scale, quality, innovation, and strategic planning [26,28,57]. It advocates for multidisciplinary collaboration to build resilient systems capable of addressing global healthcare needs. In this paper, BIM has been leveraged to incorporate complex hospital functional areas by integrating detailed spatial, mechanical, and operational data at multiple levels—building, system, and equipment. At the building level, BIM provides a comprehensive digital representation of the hospital layout, including operating rooms, wards, and Radiology Departments, for spatial coordination, regulatory compliance, and accessibility. At the system level, BIM supports the integration of critical hospital infrastructure such as HVAC systems, enabling detailed mapping of airflow dynamics, thermal zoning, and pressure differentials, which are crucial in sensitive environments like operating rooms and isolation wards. At the equipment level, BIM facilitates precise layout, equipment positioning, and sensor placement to optimize energy efficiency and environmental control. It enhances the design and management of air ducts, ventilation shafts, filtration units, and control systems to ensure optimal air circulation, sterility, and comfort. Additionally, the parametric modeling and simulation features of BIM can support performance analysis, enabling dynamic real-time adjustments based on occupancy, airflow demands, and temperature variations.

7. Conclusions

The findings underscore the importance of adaptive airflow management systems for different hospital zones. Wards and operating rooms prioritize dynamic airflow to enhance pathogen removal during epidemics while the Radiology Department maintains moderate and stable airflow. These adaptive strategies demonstrate the critical role of airflow design in ensuring infection control and maintaining a safe hospital environment. This integration of digital twins, advanced ML algorithms, and BIM technologies offers a transformative approach to improving hospital HVAC systems. Additionally, digital-twin-based parametric modeling streamlines the design and analysis and enhances the adaptability of hospital layouts to routine and emergency needs. The optimization of the framework’s utility further requires the smart platform to support real-time feedback and simulation adjustments, enabling responsive operations and seamless transitions between routine and epidemic combination modes.

The proposed framework can dynamically control temperature and airflow for better efficiency of HVAC systems. By incorporating AI and IoT-based monitoring systems, the approach enables predictive maintenance and real-time environmental adjustments in buildings and ensures energy-efficient operations while improving system reliability. This approach offers strategies to reduce operational costs while meeting public health requirements. The framework’s resilience equips hospitals to adapt to changing demands, maintaining responsiveness during routine operations and crises. Ultimately, the optimized hospital environments can improve healthcare delivery, create better working conditions, and minimize environmental impacts.

This paper tackles key challenges in public health, economics, environmental sustainability, and emergency response. Optimizing HVAC systems reduces disease transmission, enhances patient outcomes, and safeguards healthcare workers. Energy-efficient operations support sustainability goals, cut costs, and alleviate the financial burden associated with extended hospital stays. Environmentally, the framework lowers carbon footprints, fostering eco-friendly healthcare practices. Hospital resilience is strengthened by facilitating real-time airflow adjustments, ensuring patient safety, and setting a standard for sustainable and adaptive building design.

Future work should focus on adaptive optimization frameworks that integrate advanced computational models, real-time data, and emerging technologies. Research could explore AI and IoT applications for predictive maintenance and real-time system adjustments, as well as multi-objective optimization methods to balance thermal comfort, energy efficiency, and cost control. Environmental impact modeling should target carbon neutrality by leveraging renewable energy sources and sustainable materials. Additionally, scalability and resilience testing across diverse hospital environments is vital to address challenges like pandemics. These advancements will further enhance HVAC optimization, connecting engineering innovation with societal and environmental benefits. While this paper primarily focuse on hospital environments, testing other building types in future work can further improve generalizability.