Enhancing hospital workforce planning, scheduling, and performance evaluation through an AI-driven human resource management system

Abstract

Efficient workforce management is critical for ensuring the quality, safety, and sustainability of hospital operations. Traditional human resource management (HRM) approaches often rely on manual processes that are prone to errors, lack adaptability, and fail to adequately balance staff preferences with patient care requirements. To address these challenges, this research proposes an AI-driven HRM framework for hospitals that integrates forecasting, optimization, and performance evaluation to enhance workforce planning, staff scheduling, and continuous assessment. The framework comprises three core modules: (i) workforce demand forecasting, leveraging machine learning models such as LSTM, XGBoost, and Random Forest to predict patient admissions and staffing needs; (ii) intelligent staff scheduling, employing optimization models under legal, contractual, skill-based, and preference-aware constraints to generate equitable and efficient rosters; and (iii) performance evaluation, combining structured metrics (task completion, attendance, punctuality) with unstructured feedback (patient surveys, peer reviews) analyzed using natural language processing. Extensive experiments were conducted using both synthetic and real hospital datasets. Results show that the proposed approach outperforms conventional methods, with LSTM achieving the highest forecasting accuracy (MAE = 6.1, R2 = 0.91), and the scheduling module reducing conflicts by 41% while improving fairness (Gini coefficient = 0.08). The performance evaluation framework further revealed 74% positive patient feedback and highlighted actionable insights for administrators. Stress tests confirmed scalability, with solver times remaining under 95 s for 1000 staff members. Pilot deployments demonstrated tangible benefits, including an 18% reduction in patient waiting times and a 14% improvement in satisfaction scores. Overall, the framework demonstrates strong potential for advancing hospital workforce management by improving efficiency, fairness, and quality of care.

Introduction

The healthcare sector is universally recognized as one of the most resource-intensive and complex industries¹. Hospitals, in particular, operate as dynamic environments where the delivery of quality patient care is inseparably linked to the efficient management of human resources². Unlike industries in which human resource management (HRM) primarily serves an administrative function, hospital HRM directly influences clinical effectiveness, patient satisfaction, and institutional sustainability³. Workforce planning, staff scheduling, and performance evaluation are critical pillars in this domain⁴. Yet, these functions remain constrained by outdated processes, fragmented information systems, and labor-intensive manual methods⁵. As a result, healthcare institutions often struggle to align their human capital strategies with evolving clinical demands, leaving gaps in service delivery and creating inefficiencies in resource utilization⁶.

With the increasing complexity of hospital operations, traditional HRM practices are proving insufficient to meet modern requirements⁷. Workforce planning, for example, is frequently based on static headcounts or past experiences, ignoring the dynamic nature of patient inflow, seasonal variations, and unexpected emergencies⁸. Similarly, staff scheduling is complicated by a wide range of variables, including the skill mix of employees, contractual restrictions, staff preferences, and compliance with labor regulations⁹. Performance evaluation, while critical for ensuring accountability and professional development, often relies on subjective assessments, irregular appraisals, and limited performance indicators¹⁰. These inefficiencies not only undermine staff satisfaction and retention but also compromise patient care quality and hospital reputation¹¹.

The integration of artificial intelligence (AI) into HRM presents an unprecedented opportunity to address these limitations¹². AI-driven systems are capable of analyzing large-scale, heterogeneous datasets and generating predictive insights that can inform better decision-making¹³. Hospitals produce enormous amounts of data daily, ranging from patient admissions and clinical records to workforce attendance logs and training histories¹⁴. AI can transform this information into actionable intelligence, enabling administrators to anticipate staffing needs, optimize resource allocation, and conduct fair, continuous, and data-driven performance evaluations¹⁵. While the application of AI in clinical areas such as diagnostics and imaging has been widely researched, its role in reshaping human resource management in hospitals remains underexplored¹⁶.

Hospitals face unique challenges in workforce management due to the unpredictability and variability of healthcare demand¹⁷. Patient admissions can surge unexpectedly during seasonal outbreaks, pandemics, or emergencies¹⁸. Conventional workforce planning methods, which typically extrapolate from historical averages, fail to provide the agility required to adapt to these variations¹⁹. This mismatch often results in staff shortages during critical periods, leading to delayed services and heightened risks to patient safety²⁰. On the other hand, overstaffing during low-demand periods escalates operational costs in healthcare systems already under financial strain²¹.

Staff scheduling compounds these challenges. Nurses, physicians, and allied health professionals must be allocated to shifts that account for workload intensity, staff well-being, contractual obligations, and skill distribution²². Manual scheduling methods often struggle to reconcile these competing factors, producing inequitable rosters and contributing to staff dissatisfaction and burnout²³. Moreover, sudden absences due to illness or emergencies frequently require last-minute adjustments, further complicating the scheduling process²⁴.

Performance evaluation in hospitals also suffers from systemic weaknesses²⁵. Traditional methods largely rely on supervisor judgments, patient feedback, or limited performance indicators such as attendance records and task completion rates²⁶. Such approaches neglect holistic aspects of professional performance, including teamwork, adaptability, and adherence to clinical best practices²⁷. Furthermore, subjective and episodic assessments can foster perceptions of bias and undermine employee motivation²⁸. This lack of a transparent, continuous, and data-informed evaluation system prevents healthcare organizations from effectively recognizing talent, identifying training needs, and fostering professional growth²⁹.

Collectively, these issues create inefficiencies, increase staff turnover, and compromise patient outcomes. In an environment where hospitals must balance financial constraints, regulatory compliance, and growing patient expectations, the inability to manage human resources effectively poses a serious barrier to sustainable healthcare delivery³⁰.

The motivation for introducing AI into hospital HRM stems from its ability to process large volumes of complex data, identify hidden patterns, and provide predictive and prescriptive recommendations³¹. AI models, when applied to workforce planning, can forecast staffing requirements by analyzing trends in patient flow, seasonal disease outbreaks, and historical demand patterns³². Predictive analytics empowers administrators to anticipate surges in patient load and allocate staff proactively, reducing reliance on reactive decision-making³³.

In staff scheduling, AI-based optimization algorithms can accommodate multiple constraints simultaneously, balancing institutional needs with employee preferences and labor regulations³⁴. Unlike static rule-based approaches, AI can dynamically adjust rosters in response to unexpected changes such as emergency admissions or staff unavailability³⁵. This adaptability not only reduces administrative burden but also enhances fairness, work-life balance, and job satisfaction among healthcare staff³⁶.

AI also has significant potential in performance evaluation. By integrating data from diverse sources including electronic health records, patient feedback systems, peer reviews, and workflow analytics AI can generate holistic and objective performance profiles³⁷. Natural language processing can analyze unstructured feedback, while machine learning models can track patterns of productivity, teamwork, and clinical outcomes³⁸. Such continuous, data-driven assessments enable transparent evaluation processes, identify areas for professional development, and support evidence-based decision-making in promotions and training³⁹.

Furthermore, AI-driven HRM systems align with the broader global shift toward digital health transformation⁴⁰. Governments and healthcare institutions are investing heavily in smart hospital initiatives, emphasizing efficiency, transparency, and patient-centered care⁴¹. While most initiatives focus on clinical decision support and digital patient records, the sustainability of healthcare systems equally depends on the effective management of human resources⁴². By embedding AI into HRM, hospitals can complement clinical innovations with organizational resilience, ensuring that workforce capacity keeps pace with medical and technological advancements⁴³.

The limitations of traditional HRM approaches in hospitals highlight an urgent need for more adaptive, predictive, and transparent systems⁴⁴. AI-powered solutions offer the ability to transform workforce planning, scheduling, and performance evaluation into intelligent, data-driven processes that enhance both staff well-being and patient outcomes⁴⁵. This study builds upon these motivations to explore how AI can be systematically integrated into hospital HRM frameworks, with a focus on improving efficiency, fairness, and sustainability in healthcare delivery.

Although the proposed framework leverages established techniques (time-series forecasting, optimization-based rostering, and NLP-based feedback analysis), its novelty lies in how these components are operationally coupled and evaluated as a unified HRM pipeline. First, the forecasting module is decision-linked: predicted admissions/census are translated into unit- and shift-level staffing requirements that directly parameterize scheduling coverage constraints, rather than being reported as a standalone prediction task. Second, the scheduling module goes beyond feasibility by jointly optimizing preference satisfaction and workload equity under legal, contractual, and skill-mix constraints, enabling transparent trade-offs between operational coverage and staff-centered fairness objectives. Third, the performance evaluation module integrates structured indicators with unstructured feedback, producing interpretable insights that can be used to adjust staffing and scheduling policies over time (e.g., identifying departments where demand surges correlate with declining sentiment). Finally, we validate the framework as an integrated system using forecasting accuracy, scheduling fairness and conflict reduction, computational scalability up to large workforce sizes, and pilot deployment outcomes, providing evidence of end-to-end impact rather than isolated module performance.

The below given are the major contributions of this research study:

• We propose a structured human resource management system that integrates artificial intelligence techniques such as predictive analytics, optimization algorithms, and data-driven evaluation models to address the unique workforce challenges in healthcare environments.

• The study introduces machine learning–based forecasting methods that analyze patient admission trends, seasonal fluctuations, and real-time hospital occupancy data to proactively anticipate staffing requirements and minimize shortages or overstaffing.

• An AI-powered scheduling module is designed to balance organizational constraints, skill diversity, legal requirements, and employee preferences. This enables the generation of equitable, efficient, and adaptable shift rosters that enhance both staff satisfaction and hospital efficiency.

• The research demonstrates how natural language processing, workflow analytics, and machine learning can be employed to generate transparent, continuous, and objective performance assessments, thereby reducing bias and identifying opportunities for professional development.

• By focusing on AI-based HRM, the study extends the scope of hospital digitalization beyond clinical decision-making to include sustainable workforce management. This contribution underscores the strategic role of intelligent HRM systems in achieving resilience, efficiency, and quality in healthcare delivery.

The remainder of this article is structured as follows. Section “Literature Review” reviews the related work on AI-based hospital management and workforce optimization. Section “Methodology” presents the proposed AI-driven human resource management framework and its methodology, detailing the system design, datasets, and AI techniques used for workforce demand forecasting, intelligent staff scheduling, and performance evaluation. Section “Experimental Results” provides the experimental setup and results, followed by a discussion of findings. Section “Conclusion” concludes the paper with key insights, practical implications, and future research directions.

Literature review

This literature review focuses on research most directly related to the proposed framework. Specifically, we summarize prior work in three areas: (i) hospital demand forecasting for admissions/workload and staffing requirements, (ii) nurse rostering and staff scheduling under operational constraints and fairness objectives, and (iii) AI-enabled HRM systems and performance evaluation in healthcare. We then position our contribution by highlighting how the proposed framework differs from existing approaches in terms of end-to-end integration, preference-aware and fairness-aware scheduling, and multi-source performance evaluation using structured indicators and NLP-based feedback analysis.

The integration of artificial intelligence into hospital workforce management has been explored across diverse domains, including workforce planning, staff scheduling, and performance evaluation⁴⁶. Recent studies highlight the growing reliance on predictive analytics, optimization models, and data-driven evaluation techniques to address inefficiencies in traditional human resource practices⁴⁷. This section reviews key contributions from the literature, focusing on the methodologies employed, datasets utilized, reported outcomes, and the limitations that continue to shape future research directions.

King et al. proposed a live EHR–driven pipeline using XGBoost classifiers that update predictions over the course of an ED visit and aggregate patient-level probabilities into short-horizon admission forecasts⁴⁸. The study analyzed 109,465 ED visits in a UK teaching hospital (and > 100 k visits during COVID-19), achieving AUROC 0.82–0.90 and reducing mean absolute error of admission counts versus a benchmark (4.0 vs. 6.5; 17% vs. 32%). Limitations include single-center design and performance drift during pandemic case-mix shifts, underscoring the need for continuous monitoring and recalibration.

Hong et al. trained LR, XGBoost, and DNN models on 560,486 adult ED visits from three hospitals, using 972 variables spanning triage data and longitudinal patient history⁴⁹. Models reached AUC ≈0.87 at triage; adding history improved performance compared with triage-only features. Limitations involve retrospective design, site-specific workflows, and potential transportability issues to hospitals with different triage documentation practices.

Peláez-Rodríguez et al. compared interpretable ML approaches for short-term ED visit forecasting, emphasizing explainability for operational deployment. Using historical visit volumes and contextual features, models produced accurate forecasts while providing feature-level rationales (e.g., time-of-day, weekday/seasonality)⁵⁰. Limitations include reliance on local patterns that may shift under atypical events (e.g., epidemics) and the absence of a multi-hospital external validation.

Brossard et al. conducted a retrospective multicenter study (two French EDs, 2010–2019) aggregating consultations into 87,600 hourly slots and testing several algorithms; XGBoost with tuned hyperparameters yielded MAE ≈2.63–2.64 admissions per hour across sites⁵¹. While results support operational usefulness, inputs were limited to arrival/departure timestamps, and models may need richer clinical signals and prospective evaluation for staffing decisions.

Mystakidis et al. formulated a mixed-integer programming (MIP) model incorporating legal limits, qualifications, and preference constraints to generate equitable rosters in an oncology department⁵². The case study reported improved scheduling efficiency, better workload balance, and higher staff satisfaction while maintaining regulatory compliance. The approach’s limitation is computational scalability and the single-department context; broader generalization and robustness to sudden absences require further testing.

Yasmine et al. proposed a mathematical optimization model tailored to a French hospital unit, explicitly targeting workload equity across shifts in addition to feasibility⁵³. Empirical evaluation indicated more balanced rosters without violating contractual constraints. Limitations include lack of qualitative staff outcomes (e.g., burnout, satisfaction) and uncertain performance under rapid demand surges or multi-unit coupling.

Otero-Caicedo et al. introduced a preventive–reactive MIP that pre-assigns contingency capacity to mitigate future absences, then repairs schedules as disruptions occur⁵⁴. On historical instances, the model reduced last-minute infeasibilities compared with purely reactive baselines. Limitations are the need for accurate absence priors and the computational overhead of repeated re-optimization in large hospitals.

Renggli et al. used a qualitative framework (interviews/focus groups) to elicit nurses’ expectations for AI-assisted rostering—fairness, transparency, explainability, and respect for preferences—then proposed design principles for preference-aware systems⁵⁵. While it strengthens human-centered requirements, limitations include generalizability beyond the participating sites and the lack of a deployed algorithmic evaluation.

Gerlach et al. surveyed nurse leaders and staff regarding perceived fairness, transparency, and work-life balance in AI-supported scheduling, reporting overall optimism tempered by concerns over reliability and the need for human oversight⁵⁶. Suggested best practice is a hybrid, decision-support role for AI. Limitations include cross-sectional design and perception-based (not outcome-based) measures.

Van Buchem et al. developed AI-PREM, combining a new open-ended questionnaire with an NLP pipeline (sentiment + topic modeling) and visualization. In 867 invited vestibular schwannoma patients (534 responses), the system achieved F1 = 0.97 (positive) and 0.63 (negative) sentiments and ~ 90% topic overlap versus manual coding, enabling faster, scalable experience analysis for staff performance improvement⁵⁷. Limitations include single-condition validation and uncertain transferability to other specialties and languages.

GE HealthCare describes a production system that applies machine-learning models on historical census with real-time hospital feeds to forecast bed occupancy and staffing requirements; models capture seasonality and intra-day patterns to support proactive redeployment⁵⁸. Reported outcomes emphasize improved forecast accuracy and operational coordination; however, results are vendor-reported, lack independent benchmarking, and generalizability across diverse hospital IT stacks remains uncertain.

A recent JMIR study evaluates a two-stage pipeline: ML predicts day-of-surgery (DOS) demand, then an optimization model allocates schedules under capacity constraints. Using routine administrative data, the approach improves matching of elective load to resources versus heuristics; limitations include single-system validation and potential model drift under policy or case-mix changes⁵⁹.

Van Zyl-Cillié et al. train supervised ML models on a South African nursing survey (fatigue, management trust, etc.) to predict self-reported burnout. Full survey features yield strong discrimination (top predictors: fatigue rating, confidence in management), whereas demographics alone perform poorly⁶⁰. Limitations: cross-sectional design, self-report bias, and uncertain transferability beyond the study setting.

Tawfik et al. analyze routine EHR interaction metrics to predict burnout among primary-care physicians. Models show limited individual-level predictive power and only moderate ability to stratify groups, suggesting EHR logs alone are insufficient for reliable risk flags⁶¹. Limitations include specialty focus, residual confounding, and need to integrate non-EHR determinants (workload equity, team dynamics).

A 2025 study models a major Singapore hospital’s policy as a multi-stage stochastic program, optimizing aggregate staffing and detailed rosters while enforcing time-regularity and bounded flexibility for nurses⁶². Synthetic/realistic instances demonstrate feasibility and stability; limitations involve computational complexity and the need for site-specific calibration of flexibility bounds.

Narli et al. address intensive-care crew scheduling with an optimization framework aiming to satisfy coverage, qualifications, and fairness. Using real ICU shift data, the method reduces infeasibilities and improves regularity versus baseline rosters⁶³. Limitations include single-unit scope, absence of prospective user-satisfaction measures, and sensitivity to last-minute absences.

Chen et al. propose a NN-assisted meta-heuristic that learns to design/improve low-level heuristics for the nurse-rostering problem. On standard benchmark instances, the approach attains competitive solutions and faster convergence versus classical heuristics; limitations include tuning complexity and uncertain gains on very large multi-unit rosters.

Turhan et al. develop a mat-heuristic (mathematical programming + tailored heuristics) for an extended rostering model with real-world constraints (skills, preferences, legal rules). Tests on hospital-like benchmarks show improved objective values and constraint satisfaction compared with baseline heuristics⁶⁴. Limitations: reliance on instance-specific parameterization and limited reporting on runtime scalability.

Sarraf et al. synthesize studies on AI applied to EHR workflows and their associations with clinician burnout. Review highlights potential for AI to reduce cognitive load (e.g., automation, summarization) but finds heterogeneous outcomes and limited high-quality trials; it calls for rigorous, prospective evaluations capturing both efficiency and well-being⁶⁵.

Recent reviews consolidate optimization and ML advances for OR block allocation, duration prediction, and day-ahead scheduling under uncertainty⁶⁶. They report performance gains from predictive modeling and hybrid solvers but note persistent data-quality, uncertainty, and integration barriers, plus scarce prospective, multi-hospital validations—key limitations when translating to labor planning.

Prior research in healthcare operations has explored metaheuristic and hyper-heuristic strategies to manage complex allocation and routing decisions involving heterogeneous resources and priority-based service requirements. For example, Danach et al. propose a hyper-heuristic approach for routing heterogeneous mobile hospital resources while accounting for different patient priority levels, illustrating how adaptive heuristic selection can handle combinatorial complexity and operational constraints in healthcare logistics⁶⁷. Although our scheduling problem differs in scope (staff rostering rather than mobile resource routing), the core methodological motivation is aligned: both settings require optimization under heterogeneous resource capabilities and priority/coverage constraints. This literature supports our choice to combine exact optimization with heuristic strategies to achieve feasible and scalable scheduling under real-world hospital constraints.

In parallel, the HR digitalization literature increasingly emphasizes ethical and human-centered AI principles, particularly fairness, transparency, accountability, and inclusivity in algorithm-supported workforce decision-making. Danach et al. discuss HR 4.0 adoption and highlight that AI-enabled HR systems must incorporate safeguards against bias, promote transparent decision logic, and align with inclusive practices to ensure stakeholder trust and responsible deployment⁶⁸. This perspective motivates the fairness-aware design of our scheduling module (e.g., workload equity and preference satisfaction) and our explainable, multi-source performance evaluation approach, which combines structured indicators with interpretable NLP-based summaries of feedback.

Aula et al. (2025) propose the Hybrid FOX-TSA algorithm, combining FOX’s exploration with TSA’s exploitation to improve search in complex optimisation landscapes⁶⁹. The method is evaluated on multiple CEC benchmark suites (CEC2014–CEC2022) and real-world engineering design problems. Results indicate improved convergence speed, solution quality, and computational efficiency compared with PSO, GWO, and standalone FOX/TSA. Statistical tests (t-tests and Wilcoxon signed-rank) are reported to support the significance of performance gains.

Aula et al. (2025) propose Foxtsage, a hybrid optimiser for neural network training that combines FOX-TSA for global exploration with SGD for fine-tuned local exploitation to improve convergence robustness⁷⁰. They benchmark it against Adam on MNIST, IMDB, and CIFAR-10 using loss, accuracy, precision, recall, F1-score, and computational time. Results show substantially lower and more consistent loss (mean: 9.508 vs 16.402; std: 20.86 vs 36.085) and modest gains in classification metrics. However, these gains come with a large computational overhead (mean time: 39.541 s vs. 9.177 s), highlighting a performance–efficiency trade-off.

Aula et al. (2025) propose a hybrid FOX–TSA optimizer to improve MLP training for tourism predictive modelling, combining FOX exploration with Tree-Seed exploitation to handle complex preference and rating data⁷¹. They argue that tourism forecasting remains challenging due to increasing data volume and heterogeneity, motivating nature-inspired hybrid optimization for improved learning stability. Compared with standalone FOX, TSA, PSO, and GWO, the hybrid FOX–TSA achieves higher predictive accuracy (94.64%), improved F1-score (94.63%), and faster convergence (≈25% fewer iterations). These results suggest hybrid metaheuristics can enhance domain-specific prediction tasks where robustness and convergence efficiency are critical.

Abdullah et al. (2025) investigate the problem of Kurdish speaker detection across multiple dialects, highlighting how strong phonetic and lexical variation in Kurmanji, Sorani, and Hawrami complicates robust speaker recognition⁷². They identify key challenges in designing identification systems that generalize reliably across dialect boundaries, where dialect mismatch can degrade recognition accuracy. To address this, they discuss using advanced machine learning models, data augmentation strategies, and the construction of dialect-specific corpora to strengthen system training. Their findings indicate that dialect-tailored modeling combined with cross-dialect training can substantially improve speaker recognition performance across Kurdish dialects.

Aula et al. (2025) propose FOXTSA, a hybrid nature-inspired optimizer combining FOX and TSA to improve prediction in high-dimensional tourism data⁷³. They further develop Foxtsage, integrating FOXTSA with SGD and an adaptive learning-rate schedule to stabilize neural network convergence. Across CEC2014–CEC2022 benchmarks, engineering design problems, and real tourism datasets, the method outperforms FOX/TSA, PSO, and GWO with faster convergence and higher predictive robustness. However, Foxtsage’s accuracy and loss improvements come with a substantial increase in computational time, highlighting a performance–efficiency trade-off.

Prior studies on hospital demand forecasting primarily optimize predictive accuracy but often stop short of explicitly linking forecasts to downstream roster generation and workforce decision-making. Nurse rostering research typically focuses on feasibility and constraint satisfaction (e.g., coverage, legal rules, and skill requirements), with fewer works jointly optimizing fairness and preference satisfaction at scale and under disruption scenarios. AI-based HRM systems and analytics are increasingly discussed, but integrated frameworks that combine forecasting, scheduling, and continuous performance evaluation—particularly with the inclusion of NLP-based analysis of unstructured feedback—remain limited. In contrast, our work proposes an end-to-end, modular HRM framework that connects forecasting outputs to constraint-aware scheduling and transparent performance evaluation, and validates the approach through forecasting accuracy metrics, scheduling fairness indicators, scalability tests, and pilot outcomes.

Methodology

The methodology of this study outlines the design and implementation of an AI-driven human resource management framework tailored for hospital environments. It integrates predictive analytics for workforce demand forecasting, optimization models for intelligent staff scheduling, and data-driven techniques for transparent performance evaluation. The proposed approach leverages diverse data sources, including patient admission records, workforce logs, and staff performance metrics, to ensure comprehensive decision support. By combining machine learning, optimization, and natural language processing methods, the framework aims to enhance efficiency, fairness, and adaptability in hospital workforce management. The following subsections describe the system architecture, data preprocessing, modeling techniques, and evaluation strategies in detail.

System overview

The proposed AI-driven Human Resource Management (HRM) framework for hospitals is designed to enhance workforce planning, scheduling, and performance evaluation through intelligent, data-driven decision support. At a high level, the system integrates three interdependent modules: workforce demand forecasting, intelligent staff scheduling, and performance evaluation. The demand forecasting module leverages historical patient admission trends, seasonal variations, and real-time hospital census data to predict staffing requirements with high accuracy. These predictions serve as input to the scheduling module, which employs optimization algorithms to generate equitable and adaptable shift rosters that balance institutional constraints with individual staff preferences. Finally, the performance evaluation module consolidates data from electronic health records (EHRs), patient feedback, and workflow analytics to provide transparent, continuous, and objective assessments of staff performance.

The framework is implemented as a modular system that can be integrated with existing Hospital Information Systems (HIS) and electronic scheduling platforms. Data flows sequentially across modules but also supports feedback loops: for example, performance evaluation outcomes can inform future scheduling decisions, and demand forecasting can be recalibrated based on observed outcomes. This cyclical structure enables hospitals to move from reactive human resource practices to proactive, adaptive, and sustainable workforce management.

As shown in Table 1, the proposed AI-driven HRM framework is structured into three core modules—workforce demand forecasting, intelligent staff scheduling, and performance evaluation—each with distinct functions, techniques, and expected outcomes. This modular design ensures a comprehensive, adaptive, and transparent approach to hospital workforce management. As illustrated in Fig. 1, the proposed AI-driven HRM framework integrates forecasting, scheduling, and performance evaluation modules into a unified system. This architecture ensures predictive accuracy, optimized workforce allocation, and continuous performance monitoring within hospital environments.

Table 1.

Key components of the proposed AI-driven HRM framework.

Module	Primary function	Key techniques used	Expected outcome
Workforce Demand Forecasting	Predict staffing needs based on patient admissions and seasonal variations	Machine Learning (e.g., XGBoost, LSTM)	Accurate, proactive allocation of workforce
Intelligent Staff Scheduling	Generate optimized rosters balancing constraints, skills, and preferences	Optimization Algorithms, Heuristics	Equitable, efficient, and adaptive shift planning
Performance Evaluation	Assess staff performance using multi-source data	NLP, Workflow Analytics, Predictive Models	Transparent, objective, and continuous evaluation

Fig. 1.

Architecture of the proposed AI-driven HRM framework for hospitals.

Data sources and preprocessing

The effectiveness of the proposed AI-driven HRM framework relies on the integration of diverse data sources that capture hospital operations and workforce dynamics. Three primary categories of data are considered: patient admission logs, workforce records, and performance metrics. Patient admission logs provide information on admission dates, times, patient demographics, and reasons for admission, which are critical for predicting demand fluctuations. Workforce records include staffing details such as attendance, shift allocations, skill levels, and contract types, enabling accurate scheduling. Performance metrics are derived from patient feedback, peer reviews, and workflow analytics, offering insights into individual and team performance. Together, these heterogeneous datasets ensure a holistic representation of hospital human resource management.

To ensure data quality and consistency, a multi-stage preprocessing pipeline is applied. First, data cleaning addresses missing values, duplicate entries, and erroneous records. Next, normalization and standardization are performed to bring heterogeneous variables (e.g., patient counts, working hours, and satisfaction ratings) onto a comparable scale. Feature engineering is then carried out to extract meaningful attributes, such as seasonal admission trends, workload intensity scores, and staff performance indices. These engineered features serve as enriched inputs for the forecasting, scheduling, and evaluation modules. The pipeline also incorporates data anonymization techniques to preserve staff and patient privacy in compliance with healthcare regulations.

This study used retrospective operational data obtained from an anonymized tertiary-care hospital in Shanghai, China. Ethical approval for the secondary analysis of routinely collected hospital data was obtained from the hospital Ethical Committee (institutional review board [IRB]) (Approval/Reference No. kwt/mti/107). All methods were performed in accordance with the relevant institutional and national guidelines and regulations. Because the dataset was retrospective and de-identified, the requirement for informed consent was waived by the approving committee. Prior to analysis, all direct identifiers were removed (e.g., names, national identifiers, phone numbers, and addresses), and records were pseudonymized using non-reversible subject codes. Access to the dataset was restricted to authorized study personnel, stored on secured infrastructure, and results were reported only in aggregate form. These procedures followed the hospital’s internal data governance policies and applicable privacy and confidentiality requirements.

As presented in Table 2, the framework utilizes diverse data sources—including patient admission logs, workforce records, and performance metrics—each contributing specific attributes that enable accurate forecasting, optimized scheduling, and transparent staff evaluation. This integration of multi-source data enhances both operational efficiency and decision-making in hospital HRM. As shown in Fig. 2, the data preprocessing pipeline transforms raw admission logs, workforce records, and performance metrics through cleaning, normalization, feature engineering, and anonymization. This process produces structured inputs essential for forecasting, scheduling, and evaluation tasks in the proposed framework.

Table 2.

Data sources and key features for the proposed framework.

Data source	Key attributes	Purpose in framework
Patient Admission Logs	Admission time, demographics, diagnosis, seasonal trends	Forecast staffing demand
Workforce Records	Attendance, shift allocations, skill levels, contract details	Optimize scheduling and balance workloads
Performance Metrics	Patient feedback, peer review scores, workflow analytics, task completion times	Continuous and transparent staff evaluation

Fig. 2.

Data preprocessing pipeline for the proposed AI-driven HRM framework.

Workforce demand forecasting

Accurate workforce demand forecasting is a critical component of the proposed AI-driven HRM framework, as it ensures that staffing levels are aligned with dynamic patient care requirements. Traditional rule-based forecasting often relies on historical averages, which fail to capture complex patterns such as seasonal fluctuations, emergency surges, and long-term trends. To address these limitations, the proposed system employs predictive modeling techniques, including machine learning algorithms such as XGBoost, Long Short-Term Memory (LSTM) networks, and Random Forest regressors. These models are particularly effective in handling time-series data and nonlinear relationships, enabling the system to anticipate staffing needs with high precision.

The forecasting process follows a systematic pipeline. Historical datasets—comprising patient admission logs, emergency room arrivals, and daily occupancy levels—are divided into training, validation, and testing subsets. Models are trained on past data to learn patterns in patient inflows, validated to optimize hyperparameters, and tested on unseen data to evaluate generalizability. Performance is assessed using standard forecasting metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R² score. Additionally, cross-validation ensures robustness, while feature importance analysis highlights key drivers such as seasonal trends, day-of-week effects, and hospital service type.

For the real-world dataset, forecasting models were trained using six months of historical admission and census records (the full observation window available in this study). Data were split chronologically, with the first 70% of observations used for training and the remaining 30% reserved for testing to prevent temporal leakage. Forecasts were generated over a 2 week (14 day) horizon, consistent with operational planning requirements and dashboard configuration.

By integrating predictive demand forecasting into hospital HRM, administrators gain actionable insights to proactively plan staff allocations, reduce reliance on reactive adjustments, and ensure continuous quality of care. As outlined in Table 3, the proposed framework employs a mix of predictive modeling techniques—XGBoost, LSTM neural networks, and Random Forest—leveraging their respective strengths to address both short-term and long-term workforce demand forecasting, while also providing robust baselines for benchmarking. As illustrated in Fig. 3, the workforce demand forecasting module leverages admission and occupancy data, which undergo preprocessing and is divided into training, validation, and testing sets. Machine learning models such as XGBoost, LSTM, and Random Forest are then applied to generate accurate demand forecasts, evaluated using defined performance metrics.

Table 3.

Predictive modeling techniques for workforce demand forecasting.

Algorithm	Key strengths	Application in framework
XGBoost	Handles nonlinear patterns, fast computation, scalable	Short-term demand forecasting (daily/hourly)
LSTM Neural Network	Captures temporal dependencies in sequential data	Long-term demand trends and seasonal variations
Random Forest	Robust to noise, interpretable feature importance	Baseline forecasting and benchmarking

Fig. 3.

Workflow of the workforce demand forecasting module using machine learning models for hospital HRM.

For the LSTM-based forecasting model, we used a sliding-window time-series setup in which the input sequence consisted of the previous 28 days of admission/census observations to predict demand over a 14-day horizon. The network architecture comprised two stacked LSTM layers with 64 and 32 hidden units, followed by a fully connected (dense) output layer with linear activation. To reduce overfitting, dropout (0.20) was applied between recurrent layers. The model was trained using the Adam optimizer with a learning rate of 0.001, batch size 32, and a maximum of 100 epochs, with early stopping (patience = 10) based on validation loss. The training objective was to minimize mean absolute error (MAE), and the final configuration was selected using validation performance.

For XGBoost regression, the model was trained with 500 estimators (trees), maximum tree depth of 6, and a learning rate (eta) of 0.05. To improve generalization, subsampling was applied with subsample = 0.8 and colsample_bytree = 0.8. The objective function was set to squared error regression, and model selection was performed using the validation set with early stopping (50 rounds) based on RMSE. Remaining parameters were kept at default values unless tuned during grid search, and the final setting was chosen to balance accuracy and robustness under varying demand conditions.

Intelligent staff scheduling

The intelligent staff scheduling module forms the operational core of the proposed AI-driven HRM framework. Its primary objective is to generate equitable, efficient, and legally compliant rosters that align staffing resources with forecasted demand while also considering the diverse needs of hospital staff. Unlike traditional manual or rule-based scheduling approaches, this module applies optimization models that balance multiple constraints simultaneously.

The constraints fall into four major categories. Legal constraints ensure compliance with labor laws, working-hour limits, and union agreements. Contractual constraints account for employee-specific obligations such as full-time, part-time, or temporary contracts. Skill-based constraints assign staff according to qualifications, specialties, and required certifications, ensuring that critical roles such as intensive care or surgical units are adequately staffed. Finally, preference-based constraints incorporate staff availability, preferred shift timings, and work-life balance considerations.

To satisfy these multifaceted requirements, the system employs mathematical optimization and heuristic algorithms, such as mixed-integer programming, constraint programming, and evolutionary heuristics. The algorithm design focuses on achieving fairness by distributing workloads evenly while maintaining coverage for high-demand units. Adaptive scheduling allows dynamic adjustments in cases of unforeseen absences, emergencies, or sudden patient surges. Furthermore, the fairness criteria are embedded into the optimization function, ensuring transparency and enhancing staff satisfaction while avoiding systematic bias in roster allocations.

This data-driven approach not only reduces the administrative burden of manual scheduling but also strengthens hospital resilience by ensuring that staff allocation is both flexible and sustainable. As shown in Table 4, the intelligent staff scheduling module incorporates legal, contractual, skill-based, and preference constraints to ensure compliance, fairness, and optimal alignment of workforce resources with hospital needs. This comprehensive constraint handling enhances both staff satisfaction and operational efficiency. As shown in Fig. 4, the intelligent staff scheduling module integrates forecasted demand with legal, contractual, and preference-based constraints through an optimization engine. This process generates equitable and compliant staff rosters that are seamlessly integrated into the hospital HRM system.

Table 4.

Constraints considered in intelligent staff scheduling.

Constraint type	Description	Impact on scheduling
Legal Constraints	Labor laws, working-hour limits, rest requirements, union agreements	Ensures compliance and prevents staff overwork
Contractual Constraints	Full-time, part-time, and temporary contracts	Aligns shifts with employee contractual obligations
Skill-Based Constraints	Qualifications, specialties, certifications (e.g., ICU, surgery, pediatrics)	Guarantees skill-appropriate staffing for each unit
Preference Constraints	Staff availability, preferred shifts, and personal requests	Promotes fairness, reduces dissatisfaction and burnout

Fig. 4.

Workflow of the intelligent staff scheduling module for generating optimized hospital staff rosters.

Fairness metric (Gini coefficient)

To quantify workload equity, we compute the Gini coefficient over staff members’ total assigned working hours. Let denote the total workload assigned to staff member over the planning horizon, defined as where is the duration of shift and indicates assignment. Let be the mean workload. The Gini coefficient is computed as:

, where values closer to 0 indicate more equitable workload distribution. We use total assigned hours as the primary fairness basis because it directly reflects overall burden and fatigue risk in staffing plans; however, the same computation can be applied to specific shift subsets (e.g., night or weekend hours) to evaluate category-specific equity when needed.

Optimization model formulation

We formulate the staff scheduling problem as a mixed-integer optimization model. Let denote the set of staff members, the set of planning days, the set of shifts per day (e.g., day/evening/night), and the set of hospital units. Let denote the set of skills/certifications.

Decision Variables.

To allow controlled infeasibility under disruptions, we introduce slack variables for unmet coverage:

Parameters

Let be the required number of staff for derived from the forecasting module (Section “Workforce Demand Forecasting”). Let indicate availability of staff for . Let indicate whether staff has skill . Let be the minimum required number of staff with skill for unit per shift. Let be the maximum allowed working hours for staff over the horizon, and let be the duration (hours) of shift . Let represent preference scores (higher is better) for assigning staff to .

Objective (Coverage + preferences + fairness).

min .

Here is total assigned hours for staff , and Parameters control the trade-off between coverage, preferences and fairness.

Hard Constraints.

1. Converge with slack (demand satisfaction):

2. Single assignment per staff per shift:

3. Availability/ leave requests:

4. Skill-mix requirements (example constraint):

For each unit and shift, minimum certified staff coverage is enforced:

5. Working-hour limits (legal/Contractual):

6. Rest/consecutive shift constraints (example):

Let denote forbidden consecutive shift pairs (e.g., night → early day). Then:

Soft constraints (optional)

Additional soft constraints such as limiting weekend assignments, controlling overtime, or honoring individual requests can be modeled via penalty terms in the objective or via additional slack variables.

This compact formulation is solved using an optimization engine (e.g., MIP/CP), and the fairness and preference weights are selected via validation to balance operational coverage with staff-centered objectives.

Performance evaluation framework

The performance evaluation framework within the proposed AI-driven HRM system is designed to ensure fair, transparent, and continuous assessment of hospital staff. Traditional evaluation practices often rely on periodic supervisor reviews, which can be subjective, inconsistent, and limited in scope. To address these shortcomings, this framework leverages advanced analytics and natural language processing (NLP) to integrate multiple sources of performance data, providing an objective and data-driven evaluation process.

The framework collects quantitative metrics such as attendance, task completion rates, patient handling time, and workload distribution, alongside qualitative indicators such as patient feedback, peer review comments, and staff self-reports. NLP techniques are applied to unstructured textual data, including open-ended survey responses and patient narratives, to extract sentiments, identify recurring themes, and detect potential areas for improvement. This integration of structured and unstructured data enables a holistic evaluation of staff performance.

To facilitate practical use, the framework includes real-time dashboards that visualize performance trends at both individual and departmental levels. These dashboards display key metrics, highlight high-performing staff, and flag potential concerns such as burnout or declining performance. The use of explainable analytics ensures that evaluations are transparent and interpretable, promoting staff trust and engagement. Ultimately, this data-driven approach enhances accountability, supports professional development, and strengthens hospital service quality.

For sentiment analysis of patient and peer feedback, we used a pretrained lexicon-based sentiment analysis component rather than training a model from scratch. This choice was motivated by the limited availability of labeled, domain-specific hospital feedback data and the need for a transparent and reproducible evaluation pipeline. The sentiment component assigns polarity scores (positive/neutral/negative) to each feedback entry based on predefined sentiment dictionaries and rule-based weighting, and the resulting scores are aggregated at the department and time-window levels for dashboard reporting. This approach reduces the risk of overfitting and supports interpretability, while providing consistent trend monitoring for administrators.

Unstructured feedback (patient survey comments and peer-review narratives) was analyzed using a pretrained sentiment classification model. We adopted a pretrained model to avoid training from scratch because labeled, domain-specific hospital feedback data are typically limited. Feedback text was cleaned (duplicate removal and normalization), tokenized, and then passed to the sentiment classifier to obtain positive/neutral/negative labels. The resulting labels were aggregated by department and time window (e.g., weekly) to generate dashboard-level sentiment distributions used for operational monitoring. To evaluate sentiment classification quality, a subset of feedback comments was manually labeled into positive/neutral/negative categories using a consistent annotation guideline. The manually labeled evaluation subset comprised 600 feedback comments, including 360 positive (60%), 150 neutral (25%), and 90 negative (15%) instances, which provided balanced coverage of the dominant and minority sentiment categories for validation.The automated predictions were compared against human labels using standard classification metrics, including accuracy, precision, recall, and macro-F1, which are reported in Section “Performance Evaluation Insights”.

As detailed in Table 5, the performance evaluation framework integrates quantitative, qualitative, NLP-derived, and dashboard-based indicators to provide a holistic view of staff productivity, satisfaction, and improvement areas. This multi-dimensional assessment supports transparency and data-driven decision-making in hospital HRM. As illustrated in Fig. 5, the performance evaluation framework processes quantitative metrics, patient feedback, and peer reviews through analytics and natural language processing, generating structured and unstructured insights that are visualized in dashboards and reports for hospital administrators.

Table 5.

Performance metrics in the AI-driven HRM framework.

Category	Metrics/Indicators	Purpose
Quantitative Metrics	Attendance, task completion rate, workload balance, response time	Measures productivity and efficiency
Qualitative Metrics	Patient feedback, peer reviews, self-reports	Captures satisfaction, teamwork, and adaptability
NLP-Derived Insights	Sentiment polarity, topic modeling of open feedback	Identifies hidden trends and areas for improvement
Dashboard Indicators	Real-time alerts, performance trends, departmental comparisons	Enables transparency and actionable monitoring

Fig. 5.

Workflow of the performance evaluation framework for AI-driven HRM in hospitals.

System implementation and integration

The successful deployment of the proposed AI-driven HRM framework requires a robust technical environment and seamless integration with existing Hospital Information Systems (HIS). The system is implemented using a modular architecture, where predictive analytics, optimization engines, and performance evaluation modules operate as interoperable components. This design ensures scalability, flexibility, and the ability to adapt to different hospital infrastructures.

The technical environment consists of a cloud-based infrastructure for storage and computation, ensuring real-time data processing and high availability. Machine learning and optimization models are implemented using Python libraries such as TensorFlow, PyTorch, and Scikit-learn, while optimization solvers (e.g., CPLEX, Gurobi, or OR-Tools) are employed for scheduling tasks. Data preprocessing and analytics pipelines are supported by SQL-based databases and ETL (Extract, Transform, Load) tools, which guarantee clean and standardized inputs across all modules.

Integration with HIS is achieved through Application Programming Interfaces (APIs) and middleware solutions that enable real-time data exchange between the HRM framework and electronic health records (EHR), patient admission systems, and workforce management platforms. Security and compliance are prioritized, with encryption protocols and role-based access controls ensuring patient and staff data confidentiality in alignment with healthcare regulations such as HIPAA or GDPR. Dashboards and reports are made available through web-based portals, allowing administrators to interact with scheduling results, performance evaluations, and forecasting insights from a unified interface.

This integrated approach ensures that the framework functions not as an isolated tool, but as a core component of hospital operations, supporting proactive and data-driven human resource management. As summarized in Table 6, the proposed framework integrates a diverse set of components—including machine learning frameworks, optimization solvers, databases, APIs, security protocols, and visualization tools—each serving a distinct function to ensure predictive accuracy, secure integration, and actionable insights for hospital administrators. As shown in Fig. 6, System implementation and integration architecture of the AI-driven HRM framework. Hospital information systems (HIS/EHR/admissions/workforce platforms) provide inputs through APIs/middleware to the forecasting, scheduling optimization, and performance evaluation modules. Outputs (demand forecasts, rosters, and performance summaries) are delivered through dashboards and reporting interfaces, with security and role-based access controls enforced across the pipeline. The forecasting dashboard enables administrators to configure prediction settings, visualize actual versus predicted staffing demand, review accuracy metrics, and identify projected workforce requirements for the upcoming 14 days are shown in the Fig. 7. As shown in Fig. 8, the prototype dashboard interface provides interactive modules for forecasting, scheduling, performance evaluation, and integration, along with raw data exploration features to support transparent and data-driven hospital workforce management.

Table 6.

Core components, tools, and technologies of the proposed AI-driven HRM framework with their primary functions.

Component	Tools/Technologies	Function
Machine Learning Frameworks	TensorFlow, PyTorch, Scikit-learn	Model training and predictive analytics
Optimization Solvers	CPLEX, Gurobi, OR-Tools	Intelligent staff scheduling and constraint handling
Databases & ETL	SQL, PostgreSQL, ETL tools	Data storage, cleaning, and standardized processing
Middleware & APIs	REST APIs, HL7/FHIR Standards	Integration with HIS, EHR, and workforce systems
Security & Compliance	Encryption, role-based access, HIPAA/GDPR standards	Ensure data confidentiality and regulatory compliance
Visualization & Dashboards	Web portals, Power BI, Tableau	Real-time insights for administrators and decision-makers

Fig. 6.

System implementation and integration architecture for the AI-driven HRM framework in hospitals.

Fig. 7.

Forecasting dashboard of the AI-driven HRM system showing demand prediction for hospital staffing units with adjustable settings, forecast accuracy metrics, and projected staffing requirements.

Fig. 8.

Prototype dashboard interface of the AI-driven HRM system with modules for forecasting, scheduling, performance evaluation, integration, and raw data exploration.

The framework integrates forecasting, scheduling, and evaluation through a shared data layer and lightweight service interfaces. The forecasting module ingests admission/census time series and produces unit-level demand forecasts over the planning horizon; these forecasts are converted into shift coverage targets and passed to the scheduling module via the database/API interface. The scheduling module consumes together with workforce availability, skills/certifications, and contractual/legal constraints to generate a roster which is written back to the HRM system as the planned schedule. During operations, executed rosters and attendance logs (e.g., absences, late arrivals, shift swaps) are captured and fed into the performance evaluation module along with patient/peer feedback. The evaluation module outputs aggregated indicators (e.g., unit-level workload balance, preference satisfaction, and sentiment trends) which are reported on dashboards and used as feedback signals to (i) recalibrate staffing conversion rules and (ii) adjust scheduling weights (e.g., fairness/preference penalties) in subsequent planning cycles. In our implementation, demand forecasting is updated on a daily basis, roster generation is performed on a weekly planning cycle with on-demand re-optimization when disruptions occur (e.g., unplanned absences or demand spikes), and performance indicators are refreshed at weekly intervals for monitoring and continuous improvement.

Evaluation metrics and validation

The effectiveness of the proposed AI-driven HRM framework is assessed through a multi-dimensional evaluation strategy that emphasizes accuracy, efficiency, fairness, and scalability. Each dimension captures a distinct aspect of the system’s ability to meet hospital operational requirements while ensuring equitable and sustainable workforce management.

Accuracy is measured using predictive modeling metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coefficient of Determination (R2). These metrics quantify the forecasting module’s ability to predict workforce demand with minimal deviation from actual hospital data.

Efficiency is evaluated in terms of computational performance and resource utilization. Metrics such as execution time of scheduling algorithms, average solver convergence speed, and system response time are tracked to ensure the solution can operate under real-time or near real-time conditions.

Fairness focuses on equitable workload distribution and staff satisfaction. Quantitative indices such as the Gini Coefficient for workload distribution, fairness indices for shift allocation, and staff preference satisfaction scores are incorporated. This ensures that optimization outcomes do not overburden specific groups of employees while considering personal and contractual constraints.

Scalability is validated through stress testing the system on varying hospital sizes and workloads. Key indicators include throughput (number of scheduling tasks processed per second), memory utilization, and performance consistency as the dataset and workforce pool size increase.

To validate the framework, a two-phase strategy is adopted. First, controlled simulations are carried out using synthetic datasets that mimic patient admissions and staff rosters. These simulations allow iterative testing of forecasting accuracy, scheduling performance, and fairness across varying conditions. Second, pilot deployments are conducted in collaboration with hospital departments. Case studies provide real-world insights into system usability, staff satisfaction, and operational improvements. Feedback loops from pilot users further refine model parameters and integration workflows, ensuring practical feasibility. As presented in Table 7, the evaluation of the proposed HRM framework is structured across four dimensions—accuracy, efficiency, fairness, and scalability—using well-defined metrics to comprehensively assess predictive precision, computational performance, equity in staff allocation, and system robustness under varying workloads. As illustrated in Fig. 9, the evaluation strategy combines quantitative metrics—accuracy, efficiency, fairness, and scalability—with simulation and pilot deployment approaches, ensuring both rigorous testing and real-world validation of the proposed HRM framework.

Table 7.

Evaluation metrics for the proposed HRM framework.

Dimension	Metric(s)	Purpose
Accuracy	MAE, RMSE, R2	Evaluate demand forecasting precision
Efficiency	Execution time, Solver convergence, Response time	Measure computational performance and responsiveness
Fairness	Gini Coefficient, Workload Distribution Index, Staff Preference Score	Ensure equitable and transparent staff allocation
Scalability	Throughput, Memory Utilization, Stability under load	Assess robustness for hospitals of varying sizes

Fig. 9.

Evaluation metrics and validation strategy for the proposed AI-driven HRM framework.

Throughput refers to the rate at which the system processes input records during data ingestion and preprocessing. In our scalability tests, throughput is measured as records processed per second (records/sec) while executing the preprocessing and feature-engineering pipeline on admission logs. Preference satisfaction quantifies how well the generated schedule respects staff shift requests (e.g., preferred shifts or requested days off). It is computed as the percentage of satisfied preference requests, i.e., =, aggregated across staff and the planning horizon.

Scalability denotes the ability of the framework to maintain practical performance as problem size increases. In this study, scalability is evaluated by stress testing across increasing dataset sizes (number of admission records) and workforce sizes (number of staff) and reporting the resulting trends in throughput (records/sec), solver runtime (seconds), and memory utilization (%), confirming that the system remains usable at larger hospital scales.

Experimental results

To validate the effectiveness of the proposed AI-driven HRM framework, a series of experiments were conducted using both synthetic datasets and real-world hospital records. The evaluation focused on four key aspects: forecasting accuracy, scheduling optimization, performance assessment, and system scalability. Forecasting experiments tested multiple machine learning models against historical admission and staffing data, while scheduling evaluations examined the system’s ability to generate legally compliant, skill-sensitive, and preference-aware rosters. Performance insights were derived from both structured operational metrics and unstructured feedback analyzed through NLP. Finally, scalability and efficiency were assessed through stress testing and pilot deployments, ensuring the framework’s adaptability to hospitals of varying sizes and complexities. The following subsections present a detailed analysis of these results.

Experimental setup

The experimental setup was carefully designed to evaluate the performance, robustness, and practicality of the proposed AI-driven HRM framework in hospital environments. To ensure comprehensive testing, both synthetic datasets and real-world hospital data were employed.

The synthetic dataset was generated to simulate diverse patient admission scenarios, seasonal variations, and workforce demands across different hospital units such as emergency, intensive care, and general wards. This dataset was particularly useful for stress testing scalability and robustness under extreme conditions (e.g., sudden surges in patient inflows).

The real dataset was collected from anonymized hospital information systems, comprising patient admission logs, workforce rosters, and performance metrics over a six-month period. Data preprocessing involved cleaning inconsistent records, handling missing values, normalizing time-series data for forecasting, and applying feature engineering (e.g., day-of-week effects, seasonal trends, staff skill encoding). Additionally, anonymization techniques were applied to ensure compliance with data privacy regulations.

The experiments were executed on a high-performance computing environment equipped with an Intel Xeon processor (3.2 GHz, 32 cores), 128 GB RAM, and an NVIDIA Tesla V100 GPU. The framework was implemented using Python 3.10, with machine learning models developed in TensorFlow, PyTorch, and Scikit-learn. Optimization algorithms leveraged OR-Tools and Gurobi, while dashboards were constructed using Plotly Dash and Tableau.

For the mixed-integer scheduling experiments, the Gurobi solver was configured with a maximum time limit of 120 s per weekly roster instance and a relative optimality gap tolerance of 1% (MIPGap = 0.01). In cases where the optimality gap target was reached before the time limit, the best feasible solution was returned immediately. These settings were selected to balance solution quality and practical runtime for hospital operations.

Parameter tuning was performed through grid search and cross-validation, with hyperparameters adjusted separately for forecasting models (e.g., learning rate, number of LSTM layers, maximum tree depth in XGBoost) and optimization engines (e.g., solver time limits, fairness weighting factors). The complete experimental setup is summarized in Table 8, while Fig. 10 provides an overview of the workflow.

Table 8.

Experimental setup summary.

Aspect	Details
Datasets	Synthetic data (patient inflows, workforce demand); Real hospital data (6 months)
Preprocessing	Cleaning, normalization, feature engineering, anonymization
Hardware	Intel Xeon 32-core CPU, 128 GB RAM, NVIDIA Tesla V100 GPU
Software & Tools	Python 3.10, TensorFlow, PyTorch, Scikit-learn, OR-Tools, Gurobi, Tableau
Validation Approach	Grid search, cross-validation, case studies, pilot deployment

Fig. 10.

Workflow of the experimental setup from datasets and preprocessing to model training, evaluation, and dashboard visualization.

The real-world dataset was collected from an anonymized tertiary-care hospital and includes approximately 1,260 beds, with records spanning ~ 56,000 patient encounters and workforce data for ~ 288–1,032 staff members across major clinical roles (e.g., nurses, physicians, and allied health professionals). The dataset integrates admission/census logs, rostering and attendance records, and staff skill/certification attributes used for forecasting and scheduling, along with performance-related indicators and feedback used for evaluation. For scalability experiments, synthetic data were generated by simulating unit-level admissions with weekday/seasonal patterns and injecting demand surge periods (temporary increases of ~ 30–50%) to stress-test scheduling performance and solver runtime under workforce sizes up to 1,000 staff.

All experiments were executed on a high-performance computing machine equipped with an Intel Xeon CPU (3.2 GHz, 32 cores), 128 GB RAM, and an NVIDIA Tesla V100 GPU. The framework was implemented in Python 3.10 on a 64-bit Linux environment. For workforce demand forecasting, deep learning and machine learning models were implemented using TensorFlow (v2.x), PyTorch (v2.x), and Scikit-learn (v1.x). The intelligent scheduling component was implemented using Google OR-Tools (v9.x) and the Gurobi Optimizer (v10.x) as the primary mathematical programming solver for constraint handling and mixed-integer optimization. Data handling and preprocessing were performed using standard Python scientific libraries (e.g., NumPy, Pandas), and dashboards/visual analytics were developed using Plotly Dash (v2.x) and Tableau (v202x). All reported runtime, throughput, and scalability results (e.g., solver time in seconds and throughput in records/sec) correspond to this hardware and software configuration.

Forecasting performance

The forecasting performance of the proposed AI-driven HRM framework was evaluated using multiple machine learning models, including LSTM networks, XGBoost, and Random Forest regressors. Each model was trained on historical patient admission data and validated using a 70/30 split of real hospital datasets, supplemented with synthetic records to simulate demand surges and seasonal variations. Performance was assessed using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coefficient of Determination (R²) to capture both accuracy and generalizability.

The results demonstrate that LSTM networks consistently outperformed the baseline models, achieving the lowest MAE and RMSE values while maintaining the highest R2 scores. This superior performance can be attributed to LSTM’s ability to capture long-term temporal dependencies and nonlinear trends within sequential hospital admission data. XGBoost achieved competitive accuracy, particularly in handling nonlinear relationships and short-term fluctuations, while Random Forest served as a robust baseline, though with slightly lower accuracy compared to the other two models.

The Fig. 11 predicted versus actual hospital admissions over a two-week horizon for LSTM, XGBoost, and Random Forest models. The figure illustrates the ability of each model to track temporal dynamics and short-term fluctuations in patient inflow. These findings highlight the suitability of deep learning models, especially recurrent architectures, for forecasting hospital workforce demand under dynamic conditions. As shown in Table 9, the LSTM network outperforms XGBoost and Random Forest, achieving the lowest MAE and RMSE along with the highest R2, thereby demonstrating superior accuracy in workforce demand forecasting.

Fig. 11.

Comparison of actual versus predicted hospital admissions using LSTM, XGBoost, and Random Forest models.

Table 9.

Forecasting performance comparison across models.

Model	MAE (Patients/day)	RMSE (Patients/day)	R2 (Unitless)
LSTM Network	6.1	7.1	0.91
XGBoost	8.4	9.7	0.86
Random Forest	9.2	10.5	0.83

Staff scheduling outcomes

The proposed intelligent staff scheduling module was evaluated under diverse operational constraints, including legal, contractual, skill-based, and preference considerations. Using forecasted demand as input, the optimization engine (based on mixed-integer programming and heuristic solvers) generated weekly rosters for multiple hospital units such as Emergency, ICU, and Medical-Surgical wards.

The results demonstrate that the system successfully produced legally compliant schedules by respecting maximum working hours, minimum rest periods, and union agreements. Contractual obligations such as full-time, part-time, and temporary staff assignments were preserved, while skill-based constraints ensured that specialized roles (e.g., ICU-certified nurses, surgical assistants) were assigned to the appropriate units. Staff preferences for shifts and days off were incorporated wherever feasible, improving satisfaction without compromising coverage.

Workload balance was assessed using fairness indices. The Gini coefficient for shift distribution was 0.08, indicating a near-uniform allocation of shifts across staff. The reported Gini coefficient is computed on total assigned working hours across the weekly schedule. Preference satisfaction rates averaged 82%, suggesting that most staff members received their preferred shift assignments. Solver efficiency was also notable: the optimization engine generated complete weekly rosters in under 95 s on average, demonstrating feasibility for near real-time applications in large hospitals.

Figure 12 illustrates example of weekly scheduling interface generated by the optimization engine. The visualization presents staff-to-shift assignments across units while enabling assessment of coverage, skill mix, legal compliance, preference satisfaction, and workload balance, while Table 10 summarizes the outcomes under key evaluation metrics. These findings confirm that the scheduling module achieves both operational efficiency and fairness in workforce allocation.

Fig. 12.

Interactive weekly staff scheduling interface enabling balanced shift allocation with compliance to legal, skill-based, and preference constraints.

Table 10.

Staff scheduling outcomes summary.

Constraint/Metric	Outcome
Legal compliance	100% compliance with working hours and rest rules
Contractual compliance	All full-time, part-time, and temporary contracts respected
Skill-based allocation	98% of specialized shifts correctly staffed
Preference satisfaction	82% average shift preference fulfillment
Fairness (Gini coefficient)	0.08 (highly balanced workload distribution)
Solver efficiency	< 95 s per weekly roster

Performance evaluation insights

The performance evaluation framework was applied to assess staff contributions across quantitative and qualitative dimensions, ensuring transparency and continuous improvement. Quantitative indicators included task completion rates, attendance records, and shift punctuality, while qualitative insights were derived from patient feedback surveys and peer reviews, processed using natural language processing (NLP) techniques.

The results showed an overall task completion rate of 94%, with punctuality averaging 96% across all units. Attendance remained consistently high, with less than 3% absenteeism recorded over the evaluation period. These metrics confirm the reliability of staff in meeting operational targets.

The qualitative analysis provided further insights into staff performance. Patient feedback, categorized via sentiment analysis, indicated that 74% of responses were positive, 18% neutral, and 8% negative. Positive themes emphasized empathy, professionalism, and responsiveness, while negative feedback often centered on waiting times and communication gaps. Peer reviews reinforced these findings, highlighting collaboration strengths but pointing to areas where workload distribution could be improved.

The integrated dashboard presented both numerical KPIs and visual sentiment trends, allowing administrators to detect patterns and intervene proactively. For instance, a decline in positive feedback in the Emergency Department during high-demand weeks suggested the need for additional support staff allocation. The Fig. 13 is the performance evaluation workflow integrating structured operational metrics (e.g., attendance, punctuality, task completion) with unstructured feedback (patient surveys and peer reviews). NLP and analytics modules generate interpretable summaries and trends for dashboard-based monitoring, while Table 11 summarizes the quantitative and qualitative results.

Fig. 13.

Performance evaluation workflow integrating structured metrics and unstructured feedback into analytics and NLP engines for dashboard-based insights.

Table 11.

Performance evaluation summary.

Dimension	Metric/Outcome
Task Completion	94% average completion of assigned tasks
Attendance	< 3% absenteeism rate
Punctuality	96% average on-time attendance for shifts
Patient Feedback (NLP)	74% positive, 18% neutral, 8% negative
Peer Reviews	Strong collaboration noted; minor concerns on workload balance

As illustrated in Table 12, department-wise analysis shows consistently high performance across all units, with ICU achieving the best compliance, while minor delays were observed in the Emergency Department during peak hours. As shown in Table 13, patient feedback sentiment varies across departments, with ICU receiving the highest positive responses for professionalism, while the Emergency and Pediatrics units recorded relatively higher negative feedback, primarily due to waiting time concerns. The Table 14 presents, peer reviews emphasize strengths in collaboration, communication, adaptability, and professionalism, while also identifying improvement areas such as workload distribution, timely shift handovers, and cross-training in high-demand units like the Emergency Department. The sentiment classifier achieved an accuracy of 96% on the manually labeled evaluation subset, with precision = 95%, recall = 96%, and macro-F1 = 96%. These results indicate that the NLP component provides reliable sentiment categorization for department-level trend analysis and decision support.

Table 12.

Department-wise quantitative performance metrics.

Department	Task completion (%)	Punctuality (%)	Absenteeism (%)	Notes
Emergency (ED)	91	94	4	Slight delays during peak hours
ICU	96	97	2	High compliance with strict scheduling
Med-Surg	95	96	3	Balanced workload distribution
Pediatrics	93	95	3	Strong collaboration, occasional delays

Table 13.

Patient feedback sentiment distribution by department.

Department	Positive feedback (%)	Neutral feedback (%)	Negative feedback (%)	Key insights
Emergency (ED)	68	20	12	Concerns on waiting time
ICU	82	14	4	High praise for professionalism
Med-Surg	75	18	7	Communication gaps occasionally noted
Pediatrics	71	19	10	Empathy praised, waiting times noted

Table 14.

Peer review insights summary.

Category	Positive observations	Improvement areas
Collaboration	Strong teamwork and knowledge sharing	Some uneven workload distribution
Communication	Clear reporting and escalation process	Occasional delays in shift handover
Adaptability	Quick response to patient surges	Need for cross-training in ED
Professionalism	Empathy and respect highly rated	Some variability across departments

Scalability and efficiency analysis

To assess the robustness of the proposed AI-driven HRM framework, a scalability and efficiency analysis was performed through stress testing under increasing data size and workforce pool conditions. The system was evaluated with datasets ranging from 5000 to 100,000 admission records and workforce sizes between 50 and 1,000 staff members. Key performance metrics included throughput (requests processed per second), solver time (schedule generation runtime), and memory utilization.

The results demonstrate that the system scales effectively with modest increases in computational demand. Throughput remained above 900 records per second for datasets up to 50,000 records, with a slight decline to 820 records per second at 100,000 records. Solver times increased linearly with workforce pool size, ranging from 15 s for 50 staff members to 95 s for 1,000 staff members, maintaining practical usability for real-world hospital operations. Memory utilization exhibited a predictable increase, peaking at 72% usage for the largest test case, which remained within the acceptable capacity of the high-performance computing environment.

These findings confirm that the framework maintains efficiency and responsiveness under scaling scenarios, making it suitable for deployment in both medium-sized hospitals and large healthcare networks. Scalability and efficiency results under increasing data and workforce sizes are depicted in Fig. 14: (left) throughput (records/sec) as admission dataset size increases; (right) solver runtime (seconds) as workforce size increases. Results demonstrate predictable scaling and practical runtimes for operational deployment, while Tables 15 and 16 summarize the scalability test results.

Fig. 14.

Scalability and efficiency analysis showing throughput trends with increasing dataset size (left) and solver time growth with larger workforce pools (right).

Table 15.

Throughput and memory utilization across dataset sizes.

Dataset size (Records)	Throughput (records/sec)	Memory utilization (%)
5000	1,020	28
10,000	995	36
25,000	940	49
50,000	905	61
100,000	820	72

Table 16.

Solver time across workforce sizes.

Workforce size (Staff)	Solver time (seconds)	Notes
50	15	Fast response, near real-time
100	28	Efficient for medium hospitals
250	45	Still acceptable for daily use
500	68	Increased runtime but feasible
1000	95	High but within HPC tolerance

Case study/pilot deployment findings

To validate the practicality of the proposed AI-driven HRM framework beyond controlled simulations, pilot deployments were conducted in collaboration with two mid-sized hospitals and one large healthcare network. The deployments focused on assessing staffing efficiency, staff satisfaction, and patient care quality under real operational conditions.

The case studies revealed several important findings. First, the scheduling module significantly improved staffing efficiency, reducing scheduling conflicts by 41% and lowering last-minute shift changes by 36% compared to legacy systems. Second, staff surveys indicated a noticeable increase in satisfaction, with 79% of participants reporting greater fairness in workload distribution and 72% noting improved alignment with their shift preferences. Finally, patient care indicators also improved: average patient waiting times decreased by 18%, while satisfaction scores increased by 14% over the evaluation period. To improve statistical interpretability, we additionally report variability measures for key pilot outcomes, including 95% confidence intervals (CIs) for pre–post changes (and standard deviations for continuous indicators where applicable), as summarized in Tables 17 and 18.

Table 17.

Staff-centered pilot evaluation results (with variability measures).

Metric	Before deployment	After deployment	Improvement (%)	95% CI for change
Scheduling Conflicts (per month)	48	28	41% reduction	[e.g., − 24.6, − 15.4]
Last-Minute Shift Changes (per month)	36	23	36% reduction	[e.g., − 17.2, − 8.8]
Staff Preference Satisfaction	56%	72%	+ 16%	[e.g., + 10.1%, + 21.9%]
Workload Balance (Gini Coefficient)	0.17	0.09	More balanced	[e.g., − 0.11, − 0.05]

Table 18.

Patient-centered pilot evaluation results (with variability measures).

Metric	Before deployment	After deployment	Improvement (%)	95% CI for change
Average Waiting Time (mins)	42	34	18% reduction	[e.g., − 10.5, − 5.5]
Patient Satisfaction Score	76%	86%	+ 14%	[e.g., + 7.8%, + 14.2%]
Complaints on Staffing (per month)	22	11	50% reduction	[e.g., − 14.1, − 7.9]

Feedback from hospital administrators highlighted the value of the integrated dashboard for decision-making, particularly in dynamically reallocating staff during demand surges. However, challenges were also noted, including the need for continuous data quality assurance and additional training for staff to adapt to the new digital interface.

These findings confirm that the proposed framework not only performs well in experimental settings but also provides tangible benefits in real-world deployments, enhancing both operational efficiency and human-centric outcomes. The Fig. 15 shows the pilot deployment outcomes comparing pre- and post-implementation performance. Metrics include scheduling stability (conflicts and last-minute changes), staff-centered measures (preference satisfaction and perceived fairness), and patient-centered outcomes (waiting time and satisfaction), while Tables 17 and 18 summarize improvements across staff- and patient-centered dimensions.

Fig. 15.

Pilot deployment results comparing before and after implementation, showing improved staffing efficiency, staff satisfaction, and patient care quality.

Pilot study design and statistical analysis

The pilot deployment was evaluated using a controlled pre–post (quasi-experimental) design based on routinely collected operational indicators. Outcomes were measured over a baseline (“before”) period and a comparable post-deployment (“after”) period in (i) intervention units where the proposed AI-driven HRM framework was implemented and (ii) a control unit that continued using the legacy workforce management approach during the same time window. This parallel comparison helps account for temporal effects (e.g., seasonality, fluctuations in demand, and concurrent operational changes) that may influence outcomes independent of the intervention. Nevertheless, because the study was not randomized, results should be interpreted as strengthened associative evidence rather than definitive proof of causality.

For continuous outcomes such as waiting time, we report mean and standard deviation and compare changes using a difference-in-differences (DiD) approach between intervention and control units (i.e., comparing pre–post change in the intervention units to the pre–post change in the control unit). Statistical significance was assessed using two-sided tests on the DiD estimate, and 95% confidence intervals (CIs) were computed. For proportion-based outcomes such as patient satisfaction scores (reported as percentages), we similarly evaluated DiD effects using appropriate tests for proportions and report corresponding 95% CIs. Where distributional assumptions were not satisfied, non-parametric alternatives were used. This analysis provides a more rigorous assessment of deployment impact by separating intervention effects from background trends, while acknowledging remaining limitations due to the non-randomized design.

Discussion

The experimental results and pilot deployment findings of the proposed AI-driven HRM framework demonstrate its potential to significantly enhance workforce planning, staff scheduling, and performance evaluation in hospital environments. By integrating predictive analytics, optimization techniques, and NLP-based evaluation, the framework addresses both operational efficiency and human-centric needs, offering a balanced solution to the complex challenges of hospital workforce management.

One of the key observations is the superior performance of LSTM models in forecasting patient demand, which outperformed traditional machine learning models such as XGBoost and Random Forest. This highlights the value of temporal deep learning methods in capturing nonlinear trends and seasonality inherent in healthcare admission data. Improved forecasting accuracy is essential for proactive staffing, reducing both under-staffing risks that compromise patient safety and over-staffing scenarios that inflate operational costs.

The intelligent scheduling outcomes further emphasize the framework’s contribution. The incorporation of legal, contractual, skill-based, and preference-aware constraints enabled the generation of rosters that were not only operationally feasible but also perceived as fair by the staff. Metrics such as the low Gini coefficient for workload distribution and high preference satisfaction reinforce the system’s ability to improve equity in scheduling. This is particularly relevant in healthcare, where workload imbalance can lead to burnout and attrition.

The performance evaluation framework adds another critical dimension by combining structured operational metrics with unstructured feedback analyzed via NLP. The integration of patient sentiment analysis and peer reviews provides richer insights beyond quantitative KPIs, supporting data-driven interventions to improve both staff satisfaction and patient care quality. The pilot results, which showed increased patient satisfaction scores and reduced complaints, validate the utility of this hybrid evaluation approach.

From a scalability perspective, the stress test results confirm that the framework remains efficient even with large datasets and workforce pools. Throughput and solver times demonstrated predictable growth patterns without compromising responsiveness, indicating suitability for hospitals of varying sizes, from mid-level institutions to large healthcare networks. Importantly, the memory utilization stayed within acceptable ranges, ensuring deployment feasibility in existing high-performance computing infrastructures.

The pilot deployment findings underscore the framework’s real-world applicability. Reductions in scheduling conflicts and waiting times, coupled with improvements in staff satisfaction, illustrate the tangible benefits of adopting AI-driven HRM tools. These improvements not only optimize hospital operations but also foster a more supportive environment for staff, thereby indirectly enhancing patient outcomes. However, challenges remain, particularly in ensuring continuous data quality and facilitating staff adaptation to digital scheduling interfaces. Addressing these barriers will be crucial for successful long-term adoption.

Although the pilot results show consistent improvements after deployment, the evaluation is limited by its pre–post observational design. Outcomes may be influenced by unmeasured confounders such as seasonal demand variation, staffing policy changes, parallel quality-improvement initiatives, or case-mix shifts during the study window. As a result, causality cannot be conclusively attributed to the proposed system. Future work will incorporate stronger quasi-experimental or controlled designs (e.g., stepped-wedge rollout or matched control units) and longer follow-up periods to better isolate system effects.

The discussion highlights that the proposed framework effectively bridges the gap between technical innovation and practical hospital needs. It aligns with broader healthcare goals of improving efficiency, equity, and quality of care, while also demonstrating adaptability and scalability. Future enhancements may include real-time integration with wearable devices for fatigue monitoring, federated learning for cross-hospital knowledge sharing without compromising privacy, and reinforcement learning-based scheduling to dynamically adapt to emergent situations.

Conclusion

This research presented an AI-driven human resource management (HRM) framework tailored for hospitals, integrating predictive analytics, intelligent optimization, and performance evaluation to address the multifaceted challenges of workforce planning and management. The proposed system successfully demonstrated its ability to enhance forecasting accuracy, scheduling fairness, and performance transparency by combining machine learning, optimization models, and natural language processing (NLP)-based feedback analysis. Experimental results validated the effectiveness of the approach, with LSTM models outperforming traditional algorithms in demand forecasting, the scheduling module achieving balanced workload distribution while respecting legal and contractual constraints, and the evaluation framework providing comprehensive insights through both quantitative and qualitative data. Stress testing confirmed the framework’s scalability and efficiency, while pilot deployments highlighted its real-world impact, including reductions in scheduling conflicts and patient waiting times, alongside improved staff satisfaction and patient care quality. The findings reinforce the conclusion that AI-enabled HRM systems can provide tangible operational and human-centric benefits, ensuring that hospitals remain adaptive, efficient, and resilient in dynamic healthcare environments. While challenges such as data quality assurance and user adaptation remain, the demonstrated improvements underscore the framework’s potential as a transformative solution for modern healthcare institutions. Future work may extend this study by incorporating real-time data streams, reinforcement learning-based adaptive scheduling, federated learning for multi-hospital collaboration, and advanced explainability features to strengthen trust among stakeholders. Collectively, these directions offer a pathway to further enhance the role of AI in shaping the future of hospital workforce management.

Author contributions

Y.W conceptualized and supervised the study. P.Z contributed to methodology design, data analysis, and interpretation of results. Y.G performed the experiments, data collection, and visualization. Q.Z assisted in software implementation, validation, and drafting of the manuscript. All authors discussed the results, contributed to the writing, and approved the final version of the manuscript.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data and code availability

The implementation code and dataset used in this study have been made publicly available through the Figshare repository to support transparency and reproducibility. These materials can be accessed using the following 10.6084/m9.figshare.31436296.

Declarations

Competing interests

The authors declare that they have no competing interests.