Solving patient scheduling problems in hospitals using a dedicated GenAI system

Abstract

Objectives

Patient scheduling is a vital yet complex task that strongly influences patient satisfaction and optimizes healthcare efficiency. Recent studies have emphasized the importance of devising innovative approaches and developing new scheduling frameworks. Therefore, this study establishes a dedicated generative artificial intelligence (GenAI) system for patient scheduling.

Methods

The proposed system first imports scheduling data to formulate the default patient scheduling problem. Subsequently, users enter their scheduling requirements using natural language via the system interface, which are parsed using a deep neural network to establish the corresponding extended three-field notations. A customized genetic algorithm is automatically generated to solve the customized patient scheduling problem.

Results

The dedicated GenAI system was applied to a real-world case obtained from the literature, involving 12 anesthesiologists, surgeons, and anesthesia resuscitation doctors; 12 operating rooms; and 15 patients undergoing three types of surgeries, each consisting of three operations. The experimental results reveal that the difference in the optimal fitness achieved using this system and branch-and-bound was less than 1% on average, demonstrating that the proposed methodology is effective. In addition, the most complex customized patient scheduling problem could be automatically modeled and solved in 20 s. Furthermore, the scheduling performance achieved using this system was significantly higher (α = 0.05) than those achieved using two current practices. Moreover, customized patient scheduling problems are often substantially more complex than problems addressed using traditional methods reported in previous studies.

Conclusions

Applying this dedicated GenAI system improved the effectiveness of patient scheduling. This is expected to considerably enhance patient satisfaction and overall healthcare efficiency.

Keywords

patient scheduling genetic algorithm generative artificial intelligence hospital

1. Introduction

In hospital outpatient settings, patient scheduling is a particularly prominent issue; moreover, effective scheduling systems can significantly impact patient satisfaction and overall healthcare efficiency. Recent comprehensive reviews have highlighted the complexity of patient scheduling in hospitals and indicated that innovative solutions and new scheduling models are necessary to address this.^1,2 However, reviewing recent literature reveals that the use of general-purpose generative artificial intelligence (GenAI) is unsuitable for handling the diverse scheduling requirements of medical professionals.

GenAI learns underlying patterns and structures present in old data to generate new data that fulfills users’ requirements present in natural-language prompts.^3–5 GenAI is typically based on transformer-based deep neural networks,⁶ especially large language models (LLMs) (Radford et al., 2019; Goyal et al., 2024; Wang et al., 2025).^4,7,8 Moreover, GenAI enables users unfamiliar with machine learning or deep learning to leverage these advanced artificial intelligence technologies and has achieved widespread and effective application results.^9–11

Furthermore, GenAI has been widely used in hospitals and clinics for clinical administration (including healthcare-associated infection surveillance and clinical text structuring),^12,13 decision support,¹² care process improvement and planning,^14,15 workflow automation,¹² and other functions. GenAI systems can automatically populate available time slots for healthcare professionals. However, patients may compete for limited or even single medical resources, such as doctors, nurses, and robots, in the same time period.^16,17 Therefore, applying GenAI in patient scheduling, which must be optimized overall, remains rare, mainly owing to the poor capabilities of existing GenAI systems in formulating and/or solving complex optimization problems,¹⁸ such as those in the field of patient scheduling.¹⁹

Another issue pertains to the inference result obtained using a GenAI system, which may be incorrect, rendering it difficult to verify owing to the complex reasoning mechanism underlying the GenAI system.¹⁰ This issue becomes critical if the scheduling plan generated by a GenAI system is infeasible and not optimal. This study aims to address these research gaps by establishing a dedicated GenAI system for patient scheduling in hospitals or clinics. In the literature, most patient scheduling problems in hospitals or clinics have been formulated as job shop or flexible job shop problems.^20–24 Therefore, the proposed system aims to solve such scheduling problems.

In the dedicated GenAI system, the scheduling data are imported, rather than being entered with prompts. Subsequently, planners enter their scheduling requirements in natural language via the system interface. Next, a deep neural network (DNN) is constructed and trained to extract the information regarding keywords to establish an extended three-field representation (ETFN) for the scheduling problem.¹⁰ The high similarity of user backgrounds and the uniformity of system functions ensure that constructing a DNN can more conveniently and efficiently parse users’ natural-language queries than constructing an LLM.¹⁰ Based on the ETFN, the corresponding optimization model is formulated and solved using a customized genetic algorithm (GA), automatically generated by modifying the default GA program. Thus, the dedicated GenAI system is flexible and can handle highly complex or special objectives or constraints, such as optimizing a special function of cycle times, thereby restricting the processing of patients in a specific operating or testing room. In contrast, Wang et al. (2025) constructed multi-agent LLMs that combined GAs with various scheduling heuristics for similar purposes.²⁵ However, both the description and data related to the scheduling problem needed to be provided with prompts. In addition, as the GAs had not been customized, only default scheduling objective functions could be optimized. In addition, Chen et al. (2025) established decision rules to parse users’ scheduling requirements,¹⁰ whereas this study constructs a DNN parser. Furthermore, distinct from the problems solved in previous studies, dedicated resources (examination room or doctor) are required when performing surgery on patients, yielding flexible job shop scheduling problems under resource constraints.

The proposed methodology and some existing methods are compared in Table 1. The proposed methodology is evaluated using a real-world day surgery case obtained from the literature, involving the scheduling of patients on the verge of undergoing biliary surgery at a large public hospital in China.

Table 1.

Differences between the proposed methodology and some existing methods.

Method	Domain	System type	Data input	Query parser	Optimizer	Objectives	Constraints
Wang et al. (2025)²⁵	Manufacturing	Single machine, flow shop, job shop, resource-constrained project	With prompts	LLM	GA	Default	Default
Chen et al. (2025)¹⁰	Manufacturing	Flexible job shop	Imported	Decision rules	Customized GA	Customized	Customized
Chen et al. (2025)⁴	Manufacturing	Flexible manufacturing system	Imported	AdaBoost	Customized GA	Customized	Customized
Proposed methodology	Medical	Flexible job shop with resource constraints	Imported	DNN	Customized GA	Customized	Customized

The remainder of this paper is organized as follows. The relevant literature is reviewed in Section 2, and the architecture and implementation procedure of the dedicated GenAI system are presented in Section 3. The effectiveness of the dedicated GenAI system is evaluated by applying it to a real-world day surgery case obtained from the literature in Section 4. Section 5 discusses the experimental results. Finally, Section 6 concludes this study and suggests some topics that can be investigated in the future.

2. Literature review

Relevant references on modeling and solving flexible job shop scheduling problems in hospitals or clinics are reviewed as follows.

Zhao et al. (2013) investigated a patient scheduling problem for a clinic.²⁰ Patients were assigned unequal priorities, and subsequently, the scheduling objective was to minimize the weighted sum of completion times of all patients. A mixed integer-linear programming (MILP) problem was formulated, which was non-deterministic polynomial-time (NP)-hard. However, the authors applied the existing optimization software, Cplex, to derive the (local) optimal solution, rather than applying a bio-inspired algorithm to determine the global optimal solution.

Xiang et al. (2015) solved an operating room surgery scheduling problem, which required determining the start time of existing surgeries and allocating necessary resources to the scheduled surgeries.²¹ Considering resource availability and the expertise and qualifications of human resources, the authors constructed a flexible job shop scheduling problem with multiple resource constraints, characterized by a mixed integer-nonlinear programming (MINLP) model, the objective function of which minimizes the makespan. An ant colony algorithm was proposed to solve this operating room surgery scheduling problem.

Abderrabi et al. (2021) formulated an MILP model to schedule the functions related to meal preparation in hospital catering to enable collective catering.²² This was an NP-hard flexible job shop scheduling problem, aiming to minimize the sum of the completion times of all preparations. The authors applied a GA enhanced by iterated local search to identify the optimal solution.²⁶

To reduce wait times for patients at a counter, Sarfaraj et al. (2021) treated patients as jobs and counters as machines, subsequently formulating and solving an MILP problem, which was an NP-hard flexible job shop scheduling problem.²³ The proposed methodology shortened the total completion time (i.e., makespan) by 13% relative to the current method of manual scheduling.

Lingkon et al. (2025) solved a multi-objective flexible job shop problem with resource-limited constraints in hospitals, in which three objectives were optimized.²⁴ These included minimizing the makespan, minimizing the maximum machine workload, and maximizing the sum of the machine workloads.²⁵ Owing to limited resources, certain patients could only be examined in certain testing rooms and not in all the testing rooms. The multi-objective flexible job shop problem was an MINLP problem, and the weighted sum of the three objectives was optimized instead, for which a hybrid discrete firefly algorithm was proposed. In addition, sensitivity analysis was conducted to ensure that the weights of the objective functions were appropriate.

Several scheduling problems in hospitals and clinics were modeled as job shop scheduling problems, such as day surgery scheduling,²⁷ green patient flow optimization,²⁸ and operating room scheduling,²⁹ which represent a special case of flexible job shop scheduling problems. Therefore, the methods proposed in the reviewed references are also applicable.

3. Methodology

The dedicated GenAI system is composed of five main parts (refer to Figure 1), namely, the system interface, DNN, optimization model formulator, GA, and system database.

Figure 1.

Main parts of the dedicated GenAI system.

3.1. Generating the default patient scheduling model

Based on the imported scheduling data, the default patient scheduling model is automatically generated as a flow shop scheduling model with resource constraints, as provided in the Appendix.

The default patient scheduling problem is an MINLP problem, with an ETFN of $F J m / r_{j} / C_{\max}$ , which is NP-hard.³⁰ Deriving the global optimal solution to an MINLP problem is difficult; moreover, formulating an MINLP problem is challenging for most people responsible for patient scheduling in hospitals or clinics. Specifically, user-imposed objectives or constraints add or modify the objective and/or constraints of the default patient scheduling problem.

3.2. Parsing user queries

The dedicated GenAI system permits users to enter queries in natural language. Subsequently, information regarding possible keywords is extracted from each query, including the number of keywords and their relative positions.¹⁰

To facilitate the extraction process, users are instructed to divide their queries into segments using semicolons.³¹ Subsequently, each segment is preprocessed using lowercase conversion, stemming and lemmatisation, and entity linking (refer to Figure 2).

Figure 2.

Preprocessing of segments.

The DNN for parsing users’ queries is a feed-forward, fully connected neural network for classification³² with the following configuration (refer to Figure 3):

● Input layer: There are $\sum_{ϱ = 1}^{ζ} φ$ inputs to the DNN, where $φ$ indicates the maximum number of times keyword $ϱ$ may appear. The $ς$ -th input to the DNN is the relative position of the $τ$ -th occurrence of keyword $ϱ$ (refer to Figure 4), where

ς = \sum_{ξ = 1}^{ϱ - 1} φ + τ,

(16)

● Hidden layers: The DNN has ten hidden layers. The activation functions on the nodes of the first nine layers include a rectified linear unit (ReLU), hyperbolic tangent (tanh), and sigmoid functions to minimize cross-entropy loss.

● Output ( $o_{u}$ ): The classification result (an integer) of scheduling requirement u is determined using the softmax activation function that aggregates O _t,c as follows:

o_{t, c} = \frac{e^{{w_{r_{H}, c}^{(o)} h}_{t, r_{H}}^{(H)}}}{\sum_{r_{H} = 1}^{R_{H}} e^{{w_{r_{H}, c}^{(o)} h}_{t, r_{H}}^{(H)}}},

(17)

Figure 3.

DNN parser.

Then,

o_{t} = \arg \max_{c} o_{t, c},

(18)

● Actual value ( $a_{u}$ ): This is the index of the modification to the default patient scheduling model.

● Training algorithm: The Broyden-Fletcher-Goldfarb-Shanno quasi-Newton algorithm (LBFGS)³³ is applied to train the DNN, and the classification performance is determined in terms of hit rate (HR).

H R = \sum_{u = 1}^{U} I_{u} / U,

(19)

where

I_{u} = {\begin{cases} 1 & i f & o_{u} = a_{u} \\ 0 & o t h e r w i s e \end{cases} .

(20)

The identified modifications associated with all segments are applied to the ETFN of the default patient scheduling problem to generate the ETFN of the scheduling problem that addresses the user’s requirements. For example, after the modification presented in Figure 4, the ETFN changes from $F J m / r_{j} / C_{\max}$ to $F J m / r_{j}, X_{3, *, 2} + X_{3, *, 5} = 1 / C_{\max}$ .

Figure 4.

Inputs to the DNN.

3.3. Formulating the customized optimization model

The modifications to the ETFN correspond to the changes required in the optimization model of the default patient scheduling problem. Table 2 lists some common examples of such changes relevant to hospitals.

Table 2.

Examples of changes to the default optimization model.

Modification	Meaning	Objective	Constraints
$* / Y_{j_{1} k_{1} j_{2} k_{2} = 1} / *$	Operation $k_{1}$ of patient $j_{1}$ should be conducted before operation $k_{1}$ of patient $j_{1}$	-	• $Y_{j_{1} k_{1} j_{2} k_{2} = 1}$
$* / * / \bar{F} + 3 σ_{f}$	Minimizing the average cycle time plus 3 times cycle time standard deviation	• Removed: Min $Z = C_{\max}$	$• σ_{F} = \sqrt{\frac{\sum_{j = 1}^{n} {(F_{j} - \bar{F})}^{2}}{n - 1}}$
$* / * / \bar{F} + 3 σ_{f}$		• Added: Min $Z = \bar{F} + 3 σ_{F}$
$* / C_{j} \leq τ / *$	All operations of patient j should be completed before $τ$ .	-	$• C_{j} \leq τ$
$* / s_{* k} \leq τ / *$	Operation k should be started before $τ$ for all patients	-	$• s_{j k} \geq τ$
$* / d_{*} = τ / N_{T}$	Minimize the number of patients whose surgery completion times are later than $τ$	• Removed: Min $Z = C_{\max}$	$• d_{j} = τ$ $\forall j$
$* / d_{*} = τ / N_{T}$		• Added: Min $Z = N_{T}$	$• d_{j} = τ$ $\forall j$
$* / X_{j k i} = 1 / *$	Operation k of patient j should be performed by doctor i	-	$• X_{j k i} = 1$
$* / X_{j k i_{1}} + X_{j k i_{2}} = 1 / *$	Operation k of patient j should be performed by doctor $i_{1}$ or $i_{2}$	-	$• X_{j k i_{1}} + X_{j k i_{2}} = 1$
$* / Z_{j k l} = 0 / *$	Resource l is unusable for performing operation k on patient j	-	$• Z_{j k l} = 0$
$* / \sum_{j = 1}^{n} \sum_{k = 1}^{K} Z_{j k l_{1}} \geq \sum_{j = 1}^{n} \sum_{k = 1}^{K} Z_{j k l_{2}} \forall l_{2} \neq l_{1} / *$	Use resource $l_{1}$ more than other resources	-	$• \sum_{j = 1}^{n} \sum_{k = 1}^{K} Z_{j k l_{1}} \geq \sum_{j = 1}^{n} \sum_{k = 1}^{K} Z_{j k l_{2}} \forall l_{2} \neq l_{1}$

3.4. GA for solving the customized scheduling problem

In this study, the GA is automatically customized to satisfy users’ job scheduling requirements in terms of feasibility checks and/or fitness assessments. To this end, codes were pre-written to solve the default patient scheduling problem and implement all possible modifications. After extracting a user’s scheduling requirements using the DNN parser, codes for the corresponding modifications replace or are inserted into the code for the default patient scheduling problem.

The codes for the corresponding modifications, considering the data in Table 2 as an example, are listed in Table 3.

Table 3.

Codes for the corresponding modifications in Table 2 (in MATLAB).

Objective	Constraints	Customized code
-	$• Y_{j_{1} k_{1} j_{2} k_{2} = 1}$	infeasibility=0; for c1=1:population_size for c2=c1+1:population_size if current_chromosome(c1, 1)==j2 && current_chromosome(c1, 2)==k2 && current_chromosome(c2, 1)==j1 && current_chromosome(c2, 2)==k1 infeasibility=1; end end end
• Removed: Min $Z = C_{\max}$	$• σ_{F} = \sqrt{\frac{\sum_{j = 1}^{n} {(F_{j} - \bar{F})}^{2}}{n - 1}}$	cycle_time_stdev=sqrt(sum(cycle_time-mean(cycle time))/(number_of_patients-1)); … fitness=mean(cycle_time)+3*cycle_time_stdev;
• Added: Min $Z = \bar{F} + 3 σ_{F}$
-	$• C_{j} \leq τ$	infeasibility=0; for j=1:number_of_patients if completion_time(j, 1)> $τ$ infeasibility=1; end end
-	$• s_{j k} \leq τ$	start_time(patient, operation)=max(max(doctor_available_time(doctor, 1), resource_available_time(resource, 1)), completion_time(patient, operation-1)); … infeasibility=0; if start_time(patient, operation)> $τ$ infeasibility=1; end
• Removed: Min $Z = C_{\max}$	$• d_{j} = τ \forall j$	due_date(:,1)= $τ;$ … number_of_tardy_patients=sum(due_date<completion_time); … fitness=number_of_tardy_patients;
• Added: Min $Z = N_{T}$	$• d_{j} = τ \forall j$
-	$• X_{j k i} = 1$	patient=current_chromosome(:, 1); operation=current_chromosome(:, 2) … if patient==j && operation==k doctor=i; end current_chromosome(:, 3)=doctor;
-	$• X_{j k i_{1}} + X_{j k i_{2}} = 1$	patient=current_chromosome(:, 1); operation=current_chromosome(:, 2) … if patient==j && operation==k if rand>0.5 doctor=i1; else doctor=i2; end end current_chromosome(:, 3)=doctor;
-	$• Z_{j k l} = 0$	infeasibility=0; … patient=current_chromosome(:, 1); operation=current_chromosome(:, 2) doctor=current_chromosome(:, 3) … if patient==j && operation==k && resource=l infeasibility=1; end
-	$• \sum_{j = 1}^{n} \sum_{k = 1}^{K} Z_{j k l_{1}} \geq \sum_{j = 1}^{n} \sum_{k = 1}^{K} Z_{j k l_{2}} \forall l_{2} \neq l_{1}$	infeasibility=0; … for l=1:number_of_resources resource_load(l, 1)=sum(current_chromosome(:, 4)==l); end … if resource_load(i1, 1)<resource_load(l2, 1) infeasibility=1; end

3.5. Statistical analysis

Statistical methods applied in this study include a system usability scale (SUS) test and a paired t-test. The former was employed to examine the usability of the dedicated GenAI system, whereas the latter was employed to examine whether the scheduling performance of the dedicated GenAI system is significantly better than those of current practices.

4. Case study

4.1. Materials and method

The dedicated GenAI system was applied to a real-world day surgery case obtained from the literature, involving the scheduling of patients on the verge of undergoing biliary surgery at a large public hospital in China.²⁷ In this case, three types of surgeries were to be performed on patients, namely, laparoscopic cholecystectomy, cholecystectomy, and common bile duct exploration T-tube drainage, each involving preoperative anesthesia, surgery, and anesthesia resuscitation. To perform these surgeries, three anesthesiologists, six surgeons, and three anesthesia resuscitation doctors were present. Four operating rooms were available for each type of surgery, which constituted the resource constraint in this case, and there were no missing data.

A customized GA was employed in the dedicated GenAI system to solve patient scheduling problems. The primary endpoint of this GA was whether the patient scheduling performance had satisfied the user’s expectations, which was used to determine the convergence target of the GA evolution.

4.2. Application of the proposed methodology

First, the scheduling data are imported, as listed in Table 4.

Table 4.

Scheduling data. (a) Operation time, (b) resource (operating room) constraints, (c) patient data.

Surgery type	Operation time by anesthesiologist (min)			Operation time by surgeon (om)						Operation time by anesthesia resuscitation doctor (min)
Surgery type	#1	#2	#3	#1	#2	#3	#4	#5	#6	#1	#2	#3
1	30	40	35	80	80	70	90	-	60	40	40	40
2	60	80	70	-	140	130	110	150	-	50	50	50
3	75	70	65	210	190	160	150	190	-	60	60	60

Surgery type	Operating room
1	#1∼#4
2	#5∼#8
3	#9∼#12

Patient #	Surgery type
1	1
2	1
3	1
4	3
5	3
6	3
7	2
8	2
9	1
10	1
11	2
12	2
13	3
14	1
15	3

Based on the scheduling data, the default scheduling problem was automatically identified as $F J 12 / Z_{j k l} / C_{\max}$ . Moreover, theoretically, the number of feasible solutions amounted to $12^{45} \cdot {(45!)}^{12}$ .

Users enter their scheduling requirements in natural language via the user interface, as displayed in Figure 5. The requirements were segmented, from which all possible keywords, the corresponding relative positions, and auxiliary information were automatically extracted (refer to Figure 6).

Figure 5.

User interface of the dedicated GenAI system.

Figure 6.

Parsing process of natural-language queries.

The extracted keywords and their relative positions were input to a DNN to determine the modifications required to the ETFN. The DNN had 63 inputs and 10 hidden layers, and the activation functions employed by the hidden and output layers were ReLU and softmax, respectively. The output was the index of the required modification (without auxiliary information) to the ETFN.

One hundred and one users input their scheduling requirements, which are decomposed into five hundred segments, as listed in Table 5. These segments were used to train the DNN using the LBFGS algorithm.³³ Four-fold cross-validation was followed when training the DNN, which was completed in under 5 s on a PC with AMD Ryzen 5 3400G with Radeon Vega Graphics 3.70 GHz and 16.0 GB RAM. The maximum number of iterations was 1000, and the average HRs of the training and test data were 99.6% and 98%, respectively. The confusion plot of the test data is displayed in Figure 7. As some scheduling requirements involve due dates, the due date of each patient is set at four times the expected total treatment time, as listed in Table 6. Therefore, the default scheduling problem changed to

F J 12 / Z_{j k l}, d_{j} / C_{\max}

Table 5.

Users’ scheduling requirements for training the DNN parser.

#	Scheduling requirement
1	Treatments for all patients should be completed as soon as possible
2	Patient #3 should be treated by Anesthesiologist #2
3	Anesthesiologist #3 is available after 234
4	Every patient should complete the required treatment as soon as possible
5	Set a target to complete the treatment on each patient
6	The number of patients that did not reach their targets should be less than 3
…
498	Set a due date for each patient
499	Assign Patient #4 to Surgeon #2 or #3
500	Minimize the number of patients who do not reach their targets

Figure 7.

Confusion plot of the test data.

Table 6.

Due dates of the patients.

Patient #	Surgery type	Total treatment time (expected) (min)	Due date (min)
1	1	151	302
2	1	151	302
3	1	151	302
4	3	310	620
5	3	310	620
6	3	310	620
7	2	252.5	505
8	2	252.5	505
9	1	151	302
10	1	151	302
11	2	252.5	505
12	2	252.5	505
13	3	310	620
14	1	151	302
15	3	310	620

Subsequently, the DNN parser assisted users in formulating and solving scheduling problems that addressed their requirements. For example, the first user set the following scheduling requirements: “Treatments on all patients should be completed as soon as possible; patient #3 should be treated by Anesthesiologist #2; Anesthesiologist #3 is only available after 234.” The required modifications were determined using the DNN parser as “// $C_{\max}$ ,” “/ $X_{3, *, 2} = 1$ /,” and “/ $a_{3}^{(M)} = 234$ /.” Therefore, the customized scheduling problem to be solved was identified as $F J 12 / Z_{j k l}, d_{j}, X_{3, *, 2} = 1, a_{3}^{(M)} = 234 / C_{\max}$ . A customized code for a GA was automatically generated to solve this problem, in which each population contained 100 chromosomes. When 200 populations were generated or the optimal fitness was improved by less than 0.1, the evolution process was terminated. The selection method was the roulette wheel method, and single-point crossover was employed. A chromosome was mutated by changing the doctor to treat the patient at the mutation point, and the crossover and mutation probabilities were 0.2 and 0.1, respectively. A line chart is prepared to observe the fluctuation of fitness during the evolution process, as displayed in Figure 8. Obviously, the maximum fitness converged after approximately 60 populations to 8.20 E-4.

Figure 8.

Line chart for observing fitness fluctuations during the evolution process.

Figure 9.

Gantt chart.

The experimental results revealed the following:

● All patients’ operations were completed within 1220 (min).

● The scheduling results are presented as a Gantt chart to the user, as displayed in Figure 9.

● The operations assigned to the 12 operating rooms are listed in Table 7

Table 7.

Operations assigned to the 12 operating rooms.

Operating room #	Operation #
1	O21, O131, O41, O81, O32, O61, O122, O152
2	O11, O31, O82, O63, O53, O121, O51, O151
3	O111, O112, O91, O101
…
12	O92, O73, O101

5. Discussion

The experimental results indicate the following:

● Allowing users to use natural language as input enabled them to express their scheduling requirements in as much detail as possible. In this case, the scheduling problem became increasingly constrained, shrinking the feasible region. Therefore, the customized GA could converge with a reduced number of iterations or populations. In this case, only approximately 110 iterations or populations were sufficient for the evolving process to converge in the case of the maximum fitness.

● Patient scheduling in hospitals or clinics differs from job scheduling in manufacturing. Hospitals prioritize fairness in scheduling results, ensuring that each patient spends the same duration in the hospital, whereas factories usually optimize overall performance, instead of focusing on individual jobs. This difference becomes highly apparent when using GenAI to facilitate users in expressing their scheduling requirements.

● To further investigate the effectiveness of the dedicated GenAI system, a branch-and-bound (B&B) algorithm³⁴ was proposed to solve the customized patient scheduling problem written in Lingo on the same platform. Manually creating and refining the customized patient scheduling problem (if it was found to be infeasible) required approximately 3 h. The two methods found the same optimal solution, proving the effectiveness of the dedicated GenAI system.

● The scheduling results obtained using the proposed methodology are compared with those obtained using the most-work-remaining (MWR) rule, considering resource constraints such as doctor and operating room availability, manual scheduling, and B&B in Table 8. The scheduling performance obtained using the proposed methodology, in terms of fitness, was 56% and 29% higher than those obtained using manual scheduling and MWR, substantially close to that obtained using B&B. This outcome benefited both the planner and patients, as all scheduling requirements were satisfied and patients received treatments with improved effectiveness. In addition, the proposed methodology enhanced the ease of implementation, efficiency, and flexibility of patient scheduling.

Table 8.

Comparison of the scheduling results obtained using the proposed methodology and the current practice.

Method	Execution time (s)	Ease of implementation	Optimal scheduling performance
Manual scheduling	1645	Moderate	6.37E-3
MWR	246	Difficult	7.67E-3
B&B	379	Extremely difficult	1.01E-3
Proposed methodology	21	Very easy	9.91E-4

● A usability test is conducted using SUS based on feedback from 23 physicians and nurses,³⁵ the results of which are listed in Table 9. These results revealed that the usability of the dedicated GenAI system was at an average-to-good level.

Table 9.

Usability test results.

SUS item	Result (average/Std. Dev.)	Usability
1. I would like to use this system frequently.	3.7 (0.56)	Average
2. I found the system unnecessarily complex.	1.26 (0.45)	Good
3. I thought the system was easy to use.	4.04 (0.37)	Average
4. I will require support from a technical person to be able to use this system.	1.7 (0.47)	Average
5. I found the various functions in this system to be well-integrated.	3.74 (0.81)	Average
6. I thought that substantial inconsistencies are present in this system.	1.96 (0.56)	Average
7. I expect that most people will learn to use this system very quickly.	4.65 (0.49)	Good
8. I found the system very cumbersome to use.	1.48 (0.51)	Good
9. I felt very confident using the system.	3.74 (0.81)	Average
10. I needed to learn many things before I could begin using this system.	1.91 (0.6)	Average

● The results of the SUS questionnaire showed that the usability of the dedicated GenAI system was particularly satisfactory in reducing perceived complexity or cumbersomeness, thereby accelerating the expected learning time., which was due to enabling users to express their scheduling requirements in natural language and the ability of the dedicated GenAI system to interpret these queries with high accuracy. Nevertheless, traditional patient scheduling systems offer users a wide range of scheduling options. With the dedicated GenAI system, users must enter their appointment requirements, which may not be easy for users with little scheduling experiences, making them feel that they need to learn many things.

● The SUS total score of the dedicated GenAI system was compared with those of other published systems for similar purposes in Table 10. While some published systems achieved higher SUS total scores, the usability of the dedicated GenAI system also surpassed that of some other published systems, reaching a level between good and excellent.

Table 10.

Total SUS scores of various instruments.

Instrument	SUS total score
Muezzin and Khatib scheduling app³⁷	85.6
Hospital online scheduling system³⁸	79.5
Web-based scheduling system for resource management³⁹	73.5
Flexible production scheduling system⁴⁰	65.5
Dedicated GenAI system	78.9

● Table 11 lists the scheduling performances corresponding to all users of the dedicated GenAI system. The data in this table highlights the diversity of user scheduling requirements and the resulting customized patient scheduling problems. Clearly, the most significant advantage of the dedicated GenAI system was the ability to formulate and solve various customized patient scheduling problems for users in the target hospital. Furthermore, B&B validation revealed that the dedicated GenAI system was highly effective, with an average discrepancy of under 1%.

Table 11.

Scheduling performances corresponding to all users achieved using the dedicated GenAI system.

User #	Scheduling problem	Solution time (s)	Max fitness	Average fitness	Max fitness (B&B)	Difference (%)
1	$F J 12 / Z_{j k l}, d_{j}, X_{3, *, 2} = 1, a_{3}^{(M)} = 234 / C_{\max}$	18.4	8.20 E-4	5.54 E-4	8.20 E-4	0%
2	$F J 12 / Z_{j k l}, d_{j} / \bar{F}$	19.6	1.09 E-3	8.43 E-4	1.09 E-3	0.5%
3	$F J 12 / Z_{j k l}, d_{j}, N_{T} \leq 3 / C_{\max}$	20.0	1.08 E-3	5.77 E-4	1.11 E-03	2.7%
…
101	$F J 12 / Z_{j k l} {, d}_{j}, X_{4, , 2} + X_{4, , 3} = 1 / N_{T}$	17.4	0.5	0.42	0.5	0%

● A paired t-test was conducted to examine the following hypotheses to ascertain whether the advantage of the dedicated GenAI system over current practices was significant.

H₀: The scheduling performance in terms of the maximum fitness achieved using this system would be the same as those achieved using the current practices.

H₁: The scheduling performance in terms of the maximum fitness achieved using this system would be higher than those achieved using the current practices.

Table 12 lists the results obtained from the paired t-test. Specifically, H₀ was rejected if α > p(T ≤ t) in a one-tailed test. The proposed methodology outperformed two current practices, namely, manual scheduling and MWR. The scheduling performance achieved using the dedicated GenAI system was significantly higher (α = 0.05) than those achieved using both current practices.

● However, the proposed methodology possesses certain limitations. First, eliminating the requirement for users to construct and encode optimization algorithms is a black-box approach,³⁶ which may reduce users’ understanding and trust in the dedicated GenAI system.³⁷ Furthermore, allowing users to input scheduling requirements in natural language may cause multiple scheduling objectives to be inadvertently included, which requires the use of multi-objective GAs to handle.³⁸

● Another bio-inspired algorithm, such as ant colony optimization, particle swarm optimization, agent system, or artificial bee colony, can also be applied to solve the patient scheduling problem.^39–42

Table 12.

Results of the paired t-test.

	Manual scheduling	MWR	Dedicated GenAI system
Mean	0.095	0.096	0.288
Variation	0.023	0.023	0.180
Observations	14	14	14
Pearson correlation coefficient	0.681	0.681	NA
Degree of freedom	13	13	NA
t statistic	-2.495	-2.113	NA
p(T ≤ t) one-tail	0.013	0.027	NA
t critical one-tail	1.771	1.771	NA
p(T ≤ t) two-tail	0.027	0.054	NA
t critical two-tail	2.160	2.160	NA

6. Conclusions

GenAI systems have rarely been applied to patient scheduling in hospitals or clinics. Such applications are particularly urgent because people responsible for patient scheduling usually lack the relevant expertise, distinct from engineers responsible for job scheduling in factories. In addition, patient scheduling results affect each patient, whereas operation scheduling results are only meaningful to the production controller. Therefore, a dedicated GenAI system for patient scheduling in hospitals or clinics is established in this study. First, this system imports the scheduling data to formulate the default patient scheduling problem. Users enter their scheduling requirements in natural language via the system interface to formulate their customized patient scheduling problems, for which a DNN parser extracts information regarding keywords to establish the corresponding ETFNs. Subsequently, a customized GA is automatically generated to solve the customized patient scheduling problem.

The dedicated GenAI system was applied to a real-world case obtained from the literature, involving 12 anesthesiologists, surgeons, or anesthesia resuscitation doctors, 12 operating rooms, and 15 patients undergoing three types of surgeries, each consisting of three operations. The experimental results yielded the following conclusions:

● The DNN parser achieved very high classification performance for queries in the training data and effectively classified unlearned scheduling requirements.

● Given the scheduling requirements of users, the complexity of customized patient scheduling problems was higher than those of problems addressed using traditional methods in previous studies. Therefore, applying existing general-purpose GenAI systems was difficult, as minimal historical cases are available for learning.

● Increasing the number of incorporated scheduling requirements increased the number of constraints in the customized patient scheduling problem, thereby shrinking the feasible region and enabling the customized GA to converge within a reduced number of iterations or populations.

● The optimal solution obtained using the dedicated GenAI system was verified using the B&B method. The difference in the optimal fitness achieved using the two methods was under 1% on average, demonstrating the effectiveness of the proposed methodology.

● The most complex customized patient scheduling problem could be automatically modeled and solved within 20 s.

One straightforward means of monitoring and verifying a scheduling plan generated by the dedicated GenAI system is to export the scheduling results to a project management system to check the correctness of the optimal scheduling performance, identify any resource conflicts, and monitor activities as the plan progresses. In addition, system updates are accomplished by updating the information contained in the data input to the dedicated GenAI system.

This proposed system, however, poses a few limitations. Scaling the dedicated GenAI system to large hospitals handling numerous patients will be challenging. Furthermore, this study only considered two resources, namely, doctors and operating rooms, to constrain the patient scheduling problem. Considering additionally resources, such as machines or robots in rehabilitation and other healthcare scenarios,^16,17,43 will significantly increase the scale of the patient scheduling problem and expand the feasible region owing to increased dimensionality. Determining whether the dedicated GenAI system will remain effective under such circumstances requires further investigation, constituting some directions for future research.

Footnotes

ORCID iD

Tin-Chih Toly Chen

Author contributions

All authors contributed equally to the writing of this paper.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

Data will be made available on request.*

Appendix

(A1)

Min Z = Cmax

s.t. (A2)

Cmax \geq C_{j}; j = 1 \sim n

(A3)

C_{j} = C_{j K}; j = 1 \sim n

(A4)

c_{j k} = s_{j k} + \sum_{i = 1}^{m} X_{j k i} Z_{j k l} p_{j k i l}; j = 1 \sim n, k = 1 \sim K

(A5)

\sum_{i = 1}^{m} X_{j k i} = 1; j = 1 \sim n, k = 1 \sim K

(A6)

\sum_{l = 1}^{L} Z_{j k l} = 1; j = 1 \sim n, k = 1 \sim K

(A7)

s_{j 1} \geq r_{j}; j = 1 \sim n

(A8)

s_{j k} \geq c_{j, k - 1} \forall k = 2 \sim K; j = 1 \sim n

(A9)

s_{j_{2} k_{2}} \geq \sum_{i = 1}^{m} X_{j_{1} k_{1} i} X_{j_{2} k_{2} i} Y_{j_{1} k_{1} j_{2} k_{2}} c_{j_{1} k_{1}}; \forall k_{1}, k_{2} = 1 \sim K; j_{1}, j_{2} = 1 \sim n; j_{1} \neq j_{2}

(A10)

s_{j_{2} k_{2}} \geq \sum_{l = 1}^{L} Z_{j_{1} k_{1} l} Z_{j_{2} k_{2} l} Y_{j_{1} k_{1} j_{2} k_{2}} c_{j_{1} k_{1}}; \forall k_{1}, k_{2} = 1 \sim K; j_{1}, j_{2} = 1 \sim n; j_{1} \neq j_{2}

(A11)

Y_{j_{1} k_{1} j_{2} k_{2}} + Y_{j_{2} k_{2} j_{1} k_{1}} = 1; \forall k_{1}, k_{2} = 1 \sim K; j_{1}, j_{2} = 1 \sim n; j_{1} \neq j_{2}

(A12)

X_{j k i}, Y_{j_{1} k_{1} j_{2} k_{2}}, Z_{j k l} \in {0, 1}; \forall {k, k}_{1}, k_{2} = 1 \sim K; {j, j}_{1}, j_{2} = 1 \sim n; j_{1} \neq j_{2}; i = 1 \sim m; l = 1 \sim L

Other variables ∈ R⁺where $C_{\max}$ is the total completion time (makespan). $C_{j}$ is the completion time of patient j. $s_{j k}$ and $c_{j k}$ denote the start time and completion time of the k-th operation on patient j. If the k-th operation of patient j is performed by doctor i, $X_{j k i} = 1$ ; otherwise, $X_{j k i} = 0$ . If the l-th resource (e.g., testing room, operating room, etc.) is used in performing the k-th operation of patient j, $Z_{j k l} = 1$ ; otherwise, $Z_{j k l} = 0$ . $p_{j k i l}$ is the time required for the k-th operation on patient j by doctor i with resource l. $r_{j}$ is the arrival time of patient j. $Y_{j_{1} k_{1} j_{2} k_{2}} = 1$ if the $k_{1}$ -th operation on patient $j_{1}$ precedes the $k_{2}$ -th operation on patient $j_{2}$ by the same doctor or using the same resource; $Y_{j_{1} k_{1} j_{2} k_{2}} = 0$ if otherwise. The default objective function is to minimize the makespan, calculated according to Constraints (A2)-(A4). Constraints (A5) and (A6) are to ensure that each operation can be performed by only a single doctor and using a single resource, respectively. Constraints (A7)-(A8) mean that an operation can start only when the patient has arrived to the hospital and the previous operation has been completed. Constraints (A9) and (A10) are to sequence any two operations by the same doctor or using the same resource, respectively.

In addition, the cycle time of patient j can be derived as (A13)

F_{j} = C_{j} - r_{j}

Therefore, the average cycle time and cycle time standard deviation are calculated as (A14)

\bar{F} = \sum_{j = 1}^{n} F_{j} / n

(A15)

σ_{F} = \sqrt{\frac{\sum_{j = 1}^{n} {(F_{j} - \bar{F})}^{2}}{n - 1}}

respectively. When scheduling patients, it is especially important to ensure that all patients finish their consultations earlier or later by roughly the same amount, which requires minimizing cycle time standard deviation. Furthermore, the available time of doctor i is indicated with $a_{i}^{(M)}$ ; the available time of resource l is denoted by $a_{l}^{(R)}$ .

Although there is no specific due date for a patient, the patient does have an expected end time for his/her consultation, which is similar to the concept of a due date. Schedulers can set the due date for each patient to be the patient’s arrival time plus the historical average cycle time. Schedulers can also set the due date for each patient to be the closing time of the hospital/clinic. Then, then the lateness and tardiness of the patient can be derived as (A16)

L_{j} = C_{j} - d_{j}

(A17)

T_{j} = \max (L_{j}, 0)

respectively. Several scheduling objectives related to them can be optimized: (A18)

Maximum lateness : L_{\max} = \max_{j} L_{j}

(A19)

Total tardiness : \bar{T} = \sum_{j = 1}^{n} T_{j}

(A20)

Number of tardy patients : N_{T} = \sum_{j = 1}^{n} l_{j}

where (A21)

l_{j} = {\begin{cases} 1 & i f & C_{j} \geq d_{j} \\ 0 & o t h e r w i s e \end{cases}

References

Abdalkareem

Amir

Al-Betar

, et al. Healthcare scheduling in optimization context: a review. Health and Technology 2021; 11(3): 445–469. https://doi.org/10.1007/s12553-021-00547-5

Musa

Rafique

Mathew

. Hospital scheduling systems current challenges and emerging solutions. 2024 Global Digital Health Knowledge Exchange & Empowerment Conference, 2024, Abu Dhabi, United Arab Emirates, pp. 1–9.

Chiu

. The impact of Generative AI (GenAI) on practices, policies and research direction in education: A case of ChatGPT and Midjourney. Interactive Learning Environments 2023; 32: 1–17. https://doi.org/10.1080/10494820.2023.2253861

Chen

TCT

Chiu

Lin

. Customizable GenAI system for assisting job scheduling in a flexible manufacturing system consisting of multiple factories. International Journal of Advanced Manufacturing Technology 2025; 141: 5721–5734. https://doi.org/10.1007/s00170-025-16958-x

Kido

Takadama

. The challenges for GenAI in social and individual well-being. Proceedings of the AAAI Symposium Series 2024; 3(1): 365–367. https://doi.org/10.1609/aaaiss.v3i1.31236

Bengesi

El-Sayed

Sarker

, et al. Advancements in generative AI: A comprehensive review of GANs, GPT, autoencoders, diffusion model, and transformers. IEEE Access 2024; 12: 69812–69837. https://doi.org/10.1109/access.2024.3397775

Radford

Child

, et al. Language models are unsupervised multitask learners. OpenAI Blog 2019; 1(8): 9.

Goyal

Rastogi

Rajagopal

, et al. Healai: A healthcare llm for effective medical documentation. Proceedings of the 17th ACM International Conference on Web Search and Data Mining, 2024, Mérida, Mexico, pp. 1167–1168.

Law

. Application of generative artificial intelligence (GenAI) in language teaching and learning: A scoping literature review. Computers and Education Open 2024; 6: 100174. https://doi.org/10.1016/j.caeo.2024.100174

10.

Chen

Chiu

. GenAI job scheduling system for solving a flexible job shop scheduling problem. AI EDAM 2025; 39: e27. https://doi.org/10.1017/s0890060425100152

11.

Yuan

Rastogi

Naik

, et al. A continued pretrained llm approach for automatic medical note generation. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies 2024; 2: 565–571.

12.

Frid

Bassegoda

Mahamud

MAC

, et al. Bridging generative AI and healthcare practice: insights from the GenAI Health Hackathon at Hospital Clínic de Barcelona. BMJ Health & Care Informatics 2025; 32(1): e101640. https://doi.org/10.1136/bmjhci-2025-101640

13.

Morgan

Goodman

Branch-Elliman

, et al. Using generative AI for healthcare-associated infection surveillance. Clinical Infectious Diseases 2025; 82(3): 473–479.

14.

Cognetta-Rieke

Rocchio

Seys

, et al. Embracing innovation and GenAI in nursing: Early lessons from the field. Nursing Outlook 2025; 73(5): 102519. https://doi.org/10.1016/j.outlook.2025.102519

15.

Tischendorf

Hinsche

Hasseler

, et al. GenAI in nursing and clinical practice: A rapid review of applications and challenges. Journal of Public Health 2025: 1–11. https://doi.org/10.1007/s10389-025-02523-z

16.

Tamantini

Umbrico

Orlandini

. Automated planning and scheduling in robot-aided rehabilitation: A review. Journal of NeuroEngineering and Rehabilitation 2025; 22(1): 180. https://doi.org/10.1186/s12984-025-01710-z

17.

Al-Refaie

Chen

Judeh

. Optimal operating room scheduling for normal and unexpected events in a smart hospital. Operational Research 2018; 18(3): 579–602. https://doi.org/10.1007/s12351-016-0244-y

18.

Singh

. 10 best ChatGPT prompts to solve complex math problems, 2024. https://www.geeksforgeeks.org/chatgpt-prompts/#9-optimization-problems

19.

NGH

Wang

. Effective prompting with ChatGPT for problem formulation in engineering optimization. Engineering Optimization 2025; 57: 1–18. https://doi.org/10.1080/0305215x.2025.2450686

20.

Zhao

Zhu

, et al. Modeling the clinic patient scheduling as a flexible job-shop problem: A case study. IISE Annual Conference Proceedings, New Orleans, USA, p. 2952.

21.

Xiang

Yin

Lim

. An ant colony optimization approach for solving an operating room surgery scheduling problem. Computers & Industrial Engineering 2015; 85: 335–345. https://doi.org/10.1016/j.cie.2015.04.010

22.

Abderrabi

Godichaud

Yalaoui

, et al. Flexible job shop scheduling problem with sequence dependent setup time and job splitting: Hospital catering case study. Applied Sciences 2021; 11(4): 1504. https://doi.org/10.3390/app11041504

23.

Sarfaraj

Lingkon

Zahan

. Applying flexible job shop scheduling in patients management to optimize processing time in hospitals. International Journal of Research in Industrial Engineering 2021; 10(1): 46–55.

24.

Lingkon

MLR

Dash

. Multi-objective flexible job-shop scheduling with limited resource constraints in hospitals using hybrid discrete firefly algorithm. Journal of The Institution of Engineers (India): Series C 2025; 106(1): 403–423. https://doi.org/10.1007/s40032-024-01160-x

25.

Wang

Chu

. Multi-agent large language models as evolutionary optimizers for scheduling optimization. Computers & Industrial Engineering 2025; 206: 111197. https://doi.org/10.1016/j.cie.2025.111197

26.

Wang

Chen

. Adapted techniques of explainable artificial intelligence for explaining genetic algorithms on the example of job scheduling. Expert Systems with Applications 2024; 237: 121369. https://doi.org/10.1016/j.eswa.2023.121369

27.

Bai

Zhang

Luo

, et al. Day surgery scheduling and optimization in large public hospitals in China: A three‐station job shop scheduling problem. Journal of Healthcare Engineering 2022; 2022(1): 1149657–1149716. https://doi.org/10.1155/2022/1149657

28.

Vali

Salimifard

Gandomi

, et al. Application of job shop scheduling approach in green patient flow optimization using a hybrid swarm intelligence. Computers & Industrial Engineering 2022; 172: 108603. https://doi.org/10.1016/j.cie.2022.108603

29.

Ahmed

Redwan

Al Sohan

MFA

, et al. Optimizing hospital operating room scheduling: A comparative analysis of metaheuristic algorithms for job shop scheduling. 2025 International Conference on Electrical, Computer and Communication Engineering, 2025, Bangladesh, pp. 1–6.

30.

Chen

TCT

Lin

Wang

. Visualization XAI techniques for facilitating schedulers’ comprehension and interpretation of GA applications in job scheduling. International Journal of Human–Computer Interaction 2026; 42(1): 638–652. https://doi.org/10.1080/10447318.2025.2504186

31.

Naili

Chaibi

Ghezala

HHB

. Comparative study of word embedding methods in topic segmentation. Procedia Computer Science 2017; 112: 340–349. https://doi.org/10.1016/j.procs.2017.08.009

32.

Chiu

Chen

TCT

Wang

. Flexible and objective diagnosis of type II diabetes by using a fuzzy deep learning ensemble approach. Complex & Intelligent Systems 2025; 11(6): 268. https://doi.org/10.1007/s40747-025-01894-w

33.

Saputro

DRS

Handayani

Widyaningsih

. Parameter estimation of Gumbel distribution using Quasi-Newton Broyden Fletcher Goldfarb Shanno (BFGS) method and its application on data of daily precipitation in Purworejo regency. Journal of Physics: Conference Series 2019; 1188(1): 012020. https://doi.org/10.1088/1742-6596/1188/1/012020

34.

Chiu

Chen

TCT

. Enhancing the understandability and interpretability of type-II diabetes diagnosis using a deep fuzzy neural and XAI approach. 2025 IEEE 15th Symposium on Computer Applications & Industrial Electronics, 2025, Malaysia, pp. 12–17.

35.

Lewis

. The system usability scale: past, present, and future. International Journal of Human–Computer Interaction 2018; 34(7): 577–590. https://doi.org/10.1080/10447318.2018.1455307

36.

Chen

TCT

. Explainable artificial intelligence (XAI) in manufacturing. In: Explainable Artificial Intelligence (XAI) in Manufacturing: Methodology, Tools, and Applications. Springer, March, 2023. 1–11.

37.

Chen

TCT

. Explainable and Customizable Job Sequencing and Scheduling: Advancing Production Control and Management with XAI. Springer Nature, 2025.

38.

Mahmed

Kahar

MNM

. Dynamic patient scheduling in hospitals using variable length non-dominated sorting genetic algorithm III. Informatica 2024; 48(16): 155–164.

39.

Chen

TCT

Chiu

Lin

. Novel visualization XAI techniques for assisting schedulers’ comprehension and interpretation of ACO applications in job scheduling. International Journal of Human–Computer Interaction 2026; 1–34. https://doi.org/10.1080/10447318.2025.2610447

40.

Kanaga

EGM

Valarmathi

. Multi-agent based patient scheduling using particle swarm optimization. Procedia Engineering 2012; 30: 386–393. https://doi.org/10.1016/j.proeng.2012.01.876

41.

Chen

Wang

. An agent-based fuzzy collaborative intelligence approach for precise and accurate semiconductor yield forecasting. IEEE Transactions on Fuzzy Systems 2013; 22(1): 201–211.

42.

Bolaji

ALA

Bamigbola

Adewole

, et al. A room-oriented artificial bee colony algorithm for optimizing the patient admission scheduling problem. Computers in Biology and Medicine 2022; 148: 105850. https://doi.org/10.1016/j.compbiomed.2022.105850

43.

Lin

Chen

TCT

. An intelligent system for assisting personalized COVID-19 vaccination location selection: Taiwan as an example. Digital Health 2022; 8: 20552076221109062. https://doi.org/10.1177/20552076221109062