‘Applied Sott Competing 103 (2021) 107181 Contents lists available at ScienceDirect Applied Soft Computing ELSEVIER journal homepage: vnww.elsevie AGRASP-based algorithm for solving the emergency room physician ® scheduling problem M. Cildoz’, F. Mallor*’, P.M. Mateo? 215 Publc Univers of Navarre, Campus Arai. Pamplona 31005, Spain University of Zaragsza, Pero Cerbun, Zaragoza 30009, pain ARTICLE INFO ABSTRACT ‘Arte ison ‘This paper addresses @ physician scheduling problem in an Emergency Room (ER) requiring a long- Received 8 June 2020 ‘erm work calendar co allocate work days and types of shift among all the doctors. The mathematical Received i esd for 5 October 2020 ‘Accepted 24 nary 2021 Avalble online 30 January 2021 ‘model i created without simplifications, using the real calendar, including holidays. This precludes the possibilty of yelic-type solutions, and involves numerous and Varied constraints (demand, workload, fergonomics, fairness, etc). An effective solution to this very dificult practical problem cannot be Feyworde ‘obtained, for large instances, with exact solution methods. We formulate a mathematical repre- Rina evies sentation of a real-world ER physician scheduling problem featuring a hybrid algorithm combining ras? continuous linear programming with a greedy randomized adaptive search procedure (GRASP) Linear Network flow optimization Physician seedling Emergency oom programming is used to model a general physiian-demand covering problem, where the solution is Used to guide the construction phase of the GRASP, to obtain inital full schedules for subsequent Improvement by iterative application of Variable Neighborhood Descent Search (VNDS) and Network. Flow Optimization (NFO). A computational study shows the superiority of our approach over the Integer Lineat Programming method in a set of instances of varying size and difficulty inspired by 3 real setting. The methodology is embeded in 3 software tool for generating one-year-ahead physician Schedules for a local ER. These solutions, which are now in use, outperform the manually-created Schedules used previously. {© 2021 The Authors. Published by Elsevier BV. This isan open acces article under the CCBY-NC-ND 1. Introduction The Emergency Room (ER) of a hospital is where medical andjor surgical care is given to patients arriving in need of im- ‘mediate attention. An ER is therefore a 24/7 service. Physicians are required to work night, day and weekend shifts, and to take fo different ER assignments. Complex constraints add tothe diff- culty of finding good and equitable schedules for the physicians. Examples of ergonomic constraints are described in Knauth [1], ‘while Gendreau et al. [2] offer an overview of other typical con- straints to classifying them into four categories: (1) supply and demand, (2) workload, (3) fairness and (4) ergonomics, based on five case studies performed in Canadian hospitals. This paper ad- dresses areal physician scheduling problem in which constraints ofall four categories are considered. ‘Although the physician scheduling problem shares many char~ acteristics with the nurse scheduling problem (and other work- force planning problems, see, for example, De Bruecker et al [3] and Van den Bergh et a. [4)), it has received much less attention * coresponding autor Emal addreser. mastacidaxounavarses(M, Cldor, rmalotinavartaes (F Mallon), mateoounicaes (PM, Mateo) Inepfdo.org/ 10.1016 2e0¢ 2021107181 license tpl /cretivecommons orglicenses(by-nc-nd/40)) in the literature. A review of the nurse rostering problem can be found in Burke etal [5] and Cheang et al. [6]. One can, of course, expect the type of techniques that work well in one problem to do just as well in another, but, despite their basic similarity, they also have differences that can condition the solution. A thorough analysis of such differences is provided in Erhard et al [7] which highlights the importance of modeling preferences and fairness, among other issues. Their conclusion is that its combined charac- Ceristies make the physician scheduling problem highly unique, and thus distinct from general personnel scheduling problems. A similar line of reasoning is given in Damet-Kurt et al. [8], where itis also pointed out that, in physician scheduling, the issue of staffing costs is not as relevant as that of minimizing deviations from the soft scheduling requirements. The same paper also reports on an analysis of over 5500 department schedules involv- ing a total of 57 medical specialties, concluding that the most complex physician scheduling problems arise in cases. where patient care coverage is provided 247 in variable settings such as Emergency Medicine departments. The physician scheduling problem addressed in this paper is complex because it addresses each and every detail of the real-life situation, including the real work calendar and a one- year planning horizon. Managing public holiday shifts remains 1568-4945(0 2021 The Autor. Published by Elsevier BV. This isan open acces article under the CC BY-NC.ND license (pfreatvecommonsonicensesby e-ndl40.los, F. Mollor and PAE. Meo 4 major problem because they preclude the possibility of using a cyclic schedule [9]. In addition to demand constraints, the ‘model considers all mandatory constraints, as well as staff hetero- ‘geneity, and personnel preferences. The objective function also pays attention to the fairness of the schedules among physicians, Which entails balancing the distribution of different types of shifts among physicians under a range of often conflicting criteria [2] Moreover, as stated in Bruni and Detti [10], a perfect workload balance would only be obtainable when considering long plan- ning horizons (with the above-mentioned public holiday shifts fairly distributed, or irregular daily demand for physicians). (Our study addresses a one-year planning horizon because twelve-month work calendars are a legal requirement in some countries, including Spain, where they are drawn up annually by the company (after consultation and a subsequent report to the workers’ representatives) and posted in a visible area ofthe work- place (Article 36 of Workers’ Statute, BOE-A-2015-11430, [11]). ‘his calendar must contain both the work schedule and annual Gistribution of working days and holidays. It may undergo mod- ification throughout the year due to changes affecting the staff family care leave, sickness, ete. In such cases, the manager has to meet staffing demand with minimum change to the original calendar. However, the operational management of the work calendar is a different problem and lies beyond the scope of this article. Because of the many factors taken into account when planning, schedules for ER physicians, it is not easy to ensure equity oF fairness in shift distribution, Some shifts are less desirable than others, such as those worked on holidays, on weekends and at nights. An unbalanced distribution of these shifts can affect the dynamics of the physicians’ group. their job satistaction, and the effectiveness of the healthcare received by patients [12]. Sched- ules failing at achieving a fair distribution of shifts can create feelings of injustice and the existence of favoritism in the work assignment. This can happen, for example, when some physicians have to work more weekends, and in a consecutive way, than other colleagues do. This lack of equity can create dissatisfac- tion with life-work balance and, ultimately, push physicians to burn-out [13] Manually created schedules at Hospital Compound of Navarre (HCN) in Spain, failed at getting a fair distribution of these less Gesired shifts, not because of faveritism but because of the com- plexity of the task and the difficulty to obtain a balanced distri- bution. For example, at HCN physicians should work an average of 25 weekends per year. the best manual solution gets ranges from 21 to 28 working weekends with many consecutive worked weekends. Similar inequities were found in the distribution of nights and holidays. This situation was perceived as unfair and ‘gave the scheduler a hard time to justify the distribution choices Compucer-implemented algorithms, as the one propased in this paper, can obtain better solutions (in the previous example, so- lutions with no physicians working two consecutive weekends). These schedules improve the equity in the distribution of shifts, and physicians perceive them as fairer and unbiased, improv- ing the group dynamics, the quality of physician life and the healthcare provided to patients. The physician scheduling problem is a combinatorial opti- ‘mization problem that falls into the category of NP-bard prob- lems [14], which are intractable for large instances. Metaheuris- fics are powerful algorithmic approaches, which have been ap- plied with great success to many difficult combinatorial optimiza- tion problems [15]. Good solutions can be obtained by designing heuristic algorithms, usually guided by metaheuristics, or by 2 combination of heuristics and exact methods (see Karp [16]). ‘The last type of algorithm is known as matheuristics [17], which integrates (metaheuristics and Mathematical Programming (MP) ‘Applied Soft Computing 103 (2021) 107151 strategies. The hybridization benefits the performance of the al- gorithm by exploiting the structure of the optimization problem to get better solutions (contribution of MP) while keeping a reasonable computation time to reach the solution (contribution of heuristics). We initially modeled the physician scheduling problem as an Integer Linear Programming (ILP) problem, but. after a real instance ofthis problem remains unsolved by CPLEX in one week, using a powerful computer, the need to solve the problem by using a different type of algorithms arises. We design a hybrid algorithm that combines the metaheuristic Greedy Randomized ‘Adaptive Search Procedure (GRASP), Variable Neighborhood De- scent (VNDS), and MP. The resulting algorithm falls in the cat- egory of matheuristic algorithms. In general, hybrid algorithms presents a “master-slave” structure, with one of the techniques guiding the other. In our case, the heuristic is the master and ‘he MP is the slave. The GRASP construction phase provides full schedules, which are subsequently improved through a VNDS type algorithm, in combination with Network Flow Optimization (NFO) models. Besides, the fitness function used in the GRASP algorithm depends on the result of a Linear Programming (LP) problem, which solves a general physicians’ demand-covering problem. The contribution of MP is double: on the one hand, the solution of a linear programming problem guides the constructive phase to promising solutions and, on the other hand, solutions of a series of small network flow problems build up the local search. The main practical contributions of this paper are, firstly, © present a mathematical mode! accounting for all types of con- straints and objectives considered in practice by a manager when creating a hospital ER physicians’ schedule for a one-year plan- ning horizon, and secondly, to provide a hybrid algorithm with the capacity to obtain near optimal solutions to large instances of areal physician scheduling problem within minutes. The main ‘methodological contributions of this paper are the design of a greedy constructive method with a randomized component de- pendent upon the exact solution to a general covering problem Which is solved by LP. This hybridization provides high quality solutions, in terms both of feasibility and of Objective Func- tion Value (OFV). The proposed VNDS method, in combination with NFO, is applied to repair feasibility when i is necessary. Once feasibility is achieved, NFO is used alone to explore large ‘neighborhoods to improve the OFV. The proposed methodology is tested on a real problem by solving the physician scheduling problem in a hospital ER with 42 physicians and a one-year planning horizon. From 2018 to 2020, the solution was directly used in practice, being deemed by the scheduler as sufficiently superior to replace the manually-created schedule, which was not able even to fulfill all hard ergonomic constraints The paper is organized as follows. Next section provides a revision of the related literature. In Section 3, the physician scheduling problem is defined and modeled as an ILP problem. Section 4 presents and explains the hybrid methodology with its four components: (1) the covering problem solved by continuous LP, (2) the construction of a full solution by a greedy random algorithm, and (3-4) the two local search procedures (VNDS and INFO), A computational study is carried out in Section 5, which also includes the case study, a sensitivity analysis ofthe algorithm parameters and an analysis of the contribution of each compo- nent of the algorithm in obtaining good solutions. The paper tends with some conclusions. All the notation is summarized in ‘Appendix B.los, F. Mollor and PAE. Meo 2. Related work The major solution approaches for solving the physician scheduling problem involve MP, metaheuristis, constraint pro- ¢gramming, and column generation (reviewed in Gendreau eta [2]. Similar results are presented by Carter and Lapierre [18]. who analyze the characteristics of the problem and scheduling techniques based on Linear Programming (LP) and metaheuristics (mainly Tabu Search). See Erhard et al. [7] for a recent review of G8 relevant papers addressing different types of physician Scheduling problem in hospitals. They are classified diversely as staing,rosteing, or re-planning problems. The majority, 61 papers, use MP models. They can be exactly solved for small in- stances or for problems that are not highly constrained [10.19.20]. In Topalogiu [21], resident physicians in a hospital's pulmonary uni¢ ae scheduled fora 6-month period. The author considers 29 instances with the numberof variables ranging from 486 1 1995 and the number of constraints ranging from 352 to 2907. Mest of the instances are exactly solved within seconds by commercial solvers, but notin all cases, even when the problems are small in size, In other cases, asin Beaulieu et al. [22], where ILP model could not be solved by a modified version of the branch and bound method, a heuristic approach based on a partial branch and bound was used, In fact, when the problem at hand is large (a large numberof physicians to be scheduled over a long planning horizon) and very detailed models are formulated, exact Solution approaches are usually impractical, being necessary to apply heuristic or heurstc-based hybrid algorithms. For example, Puente et al. [23] solved the physician rostering problem by using a genetic algorithm for a one-month planning horizon and a small/medium size ER with 16 physicians. In Carrasco [24], 2 simple heuristic i used to assign guard shifts over a one-year horizon. The problem is not highly constrained and the number of shifts assigned per day is small: two or three depending on the day type. The table in Appendix & compares several characteristics of physician rastering problems addressed by diferent studies. Their basic goal is to assign employees to work shift, taking into consideration organizational and regulatory rules, employee skills and preferences, staffing requirements, and. other problem-specifc issues. Stafing and re-planning. problems are hot taken into account The planning horizon considered in most published studies tends to be small, ranging from two to four weeks. In Brun- ner, Bard, and Kolisch [25], for example, the physicians in an anesthesia department are scheduled to cover a two-week plan- hing horizon, later extended to six weeks in Brunner, Bard, and Kolisch [26]. The need for long-term schedules is not exclusive to Spain (as reported in Carrasco [24]); it also occurs in countries Stich as Germany [27} Canada [18], and the United States |S, 28,29]. When a long-term schedule is obtained by means of exact solution methods, i (soften necessary to partition the full problem into a sequence of interlinked medium-term scheduling problems to be solved by ILP, as in Beaulieu et al. [22] and ‘Topaloglu [21], This approach is ertcized by some authrs, who claim that a good solution cannot be obtained by combining Partial solutions [24]. One-year-ahead planning is also consid- eed for a variety of staffing problems, as in Brunner, Bard, and Kolisch [25,26] and Brunner and Edenharter [30]; and 39-week schedules are obtained in Green etal. [31 Generally, the number of different shifts considered in the literature is small (< 10), Problems involving a one-year planning horizon, in particular, consider few shift types. Carrasco [24] assigns only on-call shifts, Bruni and Detti [10] two different shift types, and Cohn etal, [28] ive. Schoenfelder and Pfeferen's [32] is the only study that schedules a large number of shit eypes ‘Applied Soft Computing 103 (2021) 107151 Since few shift-assignment studies consider the real calendar, few direct implementations of the solution in real settings are reported. Sometimes, considerable additional manual scheduling is required to enable the use of the model solution. In Ferrand et al. [33] for example, after obtaining a yearly calendar by rolling out an 8-week cycle, public holiday shifts are manually assigned independently by the scheduler, which also requires the ‘manual adjustment of assigned shifts adjacent to the holidays to overcome incompatiblites. In addition to the size of the problem — in terms of the ‘number of physicians, the planning horizon, variability and het- erogeneity in the numberof shifts co be assigned each day, types of days (workdays, weekends, and holidays), shifts lengths, and other characteristics all add to the difficulty of obtaining balanced schedules, When shifts vary greatly in length, the equitable dis- tribution of the annual working hours among physicians becomes in itself a difficult task. This problem is related to the optimal ‘multi-way partitioning problem, which is one of the original 21 problems that Richard Karp proved NP-complete (34. Until now, only Schoenfelder and Pfefferlen’s [32] process plans monthly schedules taking into account the hours worked in the previous 22 weeks. The GRASP metaheuristic, introduced by Feo and Resende [35] and formally presented by Feo and Resende [36], is a multi- start method, with each iteration of the algorithm comprising ‘construction phase and a local search phase. The first phase leads to a complete solution, and the second is the improve- ‘ment phase, which continues until a locally optimal solution is reached. After several iterations of the construction phase and the local search procedure, the best overall solution is kept as the result. The construction phase is guided by a greedy function that measures the benefit of including each new element. The benefit of selecting each element changes at each step of the construction. The method is randomized by randomly choosing the next element from a list of candidates. GRASP can be easily hybridized with other approaches and optimization strategies such as Tabu Search, Simulated Annealing, Variable Neighborhood Seatch (VNS). and population-based heuristics [37] The VNS metaheuristic method. introduced by Mladenovi¢ and Hansen [38], is based on performing systematic changes of neigh borhoods during the search space exploration. The application of, NS is quite simple, requiring only the choice of a metric to mea- sure the distance among solutions in the solution space, which induces the neighborhood structure. A guide to the application of VNS to various classic problems can be found in Hansen and “Mladenovi¢ [39]. The basic principles of VNS have been extended to provide new versions of the algorithm, which have been suc- cessfully applied for solving hard optimization problems. One of the most relevant variants is VNDS which explores neighborhoods in a deterministic way [40} The choice of neighborhood structure is critical to the perfor- ‘mance of a local search algorithm. Basically, observation shows that the larger the neighborhood, the better the local optimal so- lutions, However, the larger the neighborhood, the longer it takes to explore, Thus, efficient search procedures are required to get the most out of exploring large neighborhoods. One useful option for exploring very large-scale neighborhoods is to use network Flow techniques, as discussed and applied in the context of the traveling salesman and routing problems by Ahuja et a. (41. The result of such a combination is a matheuristic algorithm. In this and other similar cases (see. for example, Punnen [42] and Dror and Levy [43)) the so-called related graph or improvement graph is a bipartite graph used to represent assignment and matching problems. The development of mathematics-based heuristics has focused. fon studies related to the Vehicle Routing Problem, in general,los, F. Mollor and PAE. Meo and to the home health care problem. in particular, in the con- text of health, where patients are assigned to worker-teams, patient health services are scheduled and routing decisions are made [44-46]. In a related field, Hernéndez-Leandro et al. [47] developed a matheuristic based on Lagrangian relaxation for the ‘multi-activty shift scheduling problem. Some matheuristic algo- rithms have been proposed for addressing the nurse rostering problem: A YNS to accelerate a column generation method is developed and also used in [48] and [49], respectively. San- tos et al. [50] proposed a Mixed Integer Programming (MIP) model and 2 heuristic to decomposed the problem to facil- {ate a local search procedure. The first phase builds a feasible solution by solving the problem considering only the hard con- straints, this solution is improved by a VNS with neighborhoods defined by fixing some decision variables of the incumbent solu- tion and optimizing the others by using a MIP solver. This idea has been adapted recently by Wikert et al. [51] to solve a physician scheduling problem for a month time period and thvee types of shits. 3. Definition and mathematical modeling of the scheduling problem ‘The solution to the physician scheduling problem lies in de- termining which physician will work in each shift of each day throughout the planning horizon. Shifts vary in type: there are day and night shifts, workday and holidays shifts, short and long shifts, etc. Even within these categories, there are differences in terms of the task requirement: from the triage area, to the resus- Citation room, to consultation for patients with milder symptoms, etc. There is also variation in the availability and annual working time of the physicians, such that they are not all able to work all types of shifts. Age or workylife balance issues may prevent cer- ‘ain physicians from working night shifts, for example. Physicians can therefore be grouped by availability and annual working time, such that all members of each group are able to work the same ‘number of hours and types of shift The objective of the problem is to obtain the fairest feasi- bile schedule. A fair schedule is one that is evenly distributed among physicians, with all members of a group working the same number of hours, public holidays, weekends, nights (unless exempt), and each type of shift etc. A balance between groups is also required: the ratio of worked to workable shifts for each physician should be kept proportional across the groups. This workload balancing idea is further developed in Section 4, To offer some idea of the magnitude of this problem, a ‘mediumlatge size public hospital might have approximately 40 physicians, and approximately 20 different shifts per day. Over a twelve-month planning horizon, this amounts to 365 x20 = 7300 assignments, each with 40 possibilities. The theoretical number of different assignments (40°*®) is considerably reduced when different types of constraints are included. However, the number of feasible solutions is still huge. The general formulation of this scheduling problem considers NN physicians groupable into M types with n- physicians of type G1 = 1,..-,M, and L types of shifts 5),j = 1,...,1, each defined by its duration d, (in hours), and other characteristics such as night shift, workday shift, the physician's location during the shift, and types of duties required, among others. There are in, shifts of type 5; in the planning period. Let T be the number of days for the planning horizon, ‘Applied Soft Computing 103 (2021) 107151 Each physician type G, r = 1,....M can work a maximum, of h = pH hours during the planning horizon (where H is the ‘number of working hours of a fulltime physician and py = 1), in a subset of shifts determined by binary indicators 7 1 ifa physician of type G, can work in shift type 5; lee (3 otherwise vr IME WSJ = Toco m Without loss of generality itis assumed that a subset of shifts S(t) € [5 = 1-1] needs tobe assigned each day and that the demand for each type of shit is one. This assumption reflects the high diversity of shifts in the ER, and places the definition of the problem ina worst case scenario. but the algorithm developed in this research can be straightforwardly adapted for a demand level greater than one. This physician scheduling problem can be ‘mathematically modeled as an ILP problem by using the following decision variables Xie Xu = [7 physician P, works S;on day ¢ = 10 otherwise Vix 1 NYS ESV VERT eT @) Feasible schedules need to cover all shifts, observe the maximum working hours of each physician, and comply with ergonomic constraints (especially those relating to the length of rest period after some types of shifts). Therefore, constraints are classified by {ype into (1) coverage. (2) ergonomic, and (3) work balance. ‘* Coverage constraints. The demand rules are the most basic compulsory requirements: each physician can be assigned a max- imum of one shift per day, and each shift must be assigned to @ single physician, Vote S1Vi= tn ¥e a @ se Dox = IY, S03 VE rT 4 «* Ergonomic constraints. ER Services are available at all hours of the day and night, every day of the year. Having to work long shifts at any part of the day without reasonably-spaced rest periods between shifts turns a poor work schedule into a potential health threat for physicians. To mitigate the effects of 2 chaotic work shift calendar, further constraints are added (both {o meet legal requirements and to accommodate suggestions from physicians) to enable physical and mental recovery as well as a normal social and family life. Specifically, these so-called ergonomic constraints are designed, among other purposes, to avoid consecutive night shifts, to program rest periods after a long ‘oF night shift, to plan weekends off, to avoid an excessive number of rest days between working days, to alternate shift lengths, ec. Ergonomic constraints are classified into three types according, to their purpose: to leave a time interval between shifts, to limit the number of shifts within a time window, and to limit ‘the number of consecutive working days. These constraints can be formulated for each shift type, for all shifts in general, or for subsets of De~ shifts with C— characteristics. For example, De ={night shifts worked on public holidays) contains all shifts with C— characteristcs=(night, public holiday). (G) Minimum days’ interval between shifts. There must be a ‘minimum interval of 8 days between two shifts belonging to the set De. Yo est and hats fare fae AyseeuNt ¥De (5a)los, F. Mollor and PAE. Meo For example, in the event of having to spread out night shifts by imposing a two-day interval between two worked night shifts (such that there can be only one night shift in a period of 3 days). 5 =2. This category of constraints includes a compulsory number of days off after certain types of shifts and is formulated as follows when 4, days’ rest are required after a shift 5; in a set De ht + 5S Xe 8.91 a vii, TH (5b) N; V5)e De (i) Meximum numberof shifts worked within a time windove. This constraint limits the maximum number of shifts in a set De assigned to physicians over a time window of w. days YL Le soar vi : 6) NV g=1e,...,T: ¥De ‘This type of constraint is used, say, to limit the number of public holidays worked within a certain period. Suppose that a physician cannot be assigned more than 5 public holiday shifts over a time window of 30 days. Then, vie = 5 and wie = 30. (ii) Maximum number of consecutive working days. Physicians cannot work more than wae consecutive days on any type of shift belonging to a set De Yo Dxa cue lei iO, Via be M VG a Here also, there may be constraints imposing a maximum on the ‘number of days’ gap between shifts, a minimum on the number of a certain type of shift chat can be assigned within a time window, and a minimum on che number of consecutive days on shifts belonging to a set D,. The formulation of these constraints is similar to that given in (5a), (6), and (7) Workload balancing constraints. These constraints are de- Signed to guarantee a fair distribution of the different types of shifts among all physicians. wae + Tess TWD (i) Fair distribution of working hours on shifts belonging to a set D- among all physicians. DY Lax < wt! vin nae YD Vbke = ot viet oo HY and HE are variables representing the maximum and, minimum number of hours worked on shifts with charac- tevaties in C, respectively. These constants could also be applied to 3 single typeof shits Sor to the entire set of shits IN (PE Gels WDe (8) .N(Pr€ Gr}: WDe (9) (ii) Fair distribution among all physicians of shifts in a set D. ke < ol Vim 1 MB eG WD. (10) ecarfeo YO Exes atv sN(RE Gh: WD. (11) aeirheot These constraints are similar tothe previous ones, bu are now aimed at balancing the number of shifts rather than the number of working hours, The variables J2 and f!, respectively, limit the maximum and minimum number of shifts worked by all physicians, ‘Applied Soft Computing 103 (2021) 107151 Fair distribution of shifts from a set D. among physicians in the same group. Constraints for balancing the number of shifts can be assigned to particular types of physicians. LD Vee ss vi .N(PL€ Gy): WDe (12) YS Week Viste NEG wD. (13) cirket The variables! and J. limit the maximum and minimum numberof shifts in set, worked by physicians Fin group Gut =, M, respectively The objective function is defined to reach the fairest distribu- tion of the workload among physicians by minimizing the range of the limiting variables HE and HY. J and! J and Je. Thus, the objective Function isthe minimization of the sum of all ranges ain Sa — HE) + So — Where #D is the number of sets of shifts D,, involved in the fair- ness constraints. Different weights may be used in the objective function to reflect the relative importance of the fairness of the shift distribution and working hours among physicians. Thus, the ILP model for the physician scheduling problem in- volves the minimization ofthe objective function (14) subject to 2 set of constraints (3)-(13), which is fully presented in Appendix C + uk (14) 4. The hybrid GRASP based algorithm This section explains the hybrid methodology. Section 4.1 provides a general overview of the algorithm, In Section 4.2 2 general covering problem, modeled as an LP problem, is solved to obiain the average number of shifts of each type that should be worked by physicians of each type. These averages are used in Section 4.3 by a greedy random algorithm to construct a full solution, Finally Section 4.4 presents two local search procedures {0 improve the solution obtained by the greedy algorithm, 4.1, General description of the algorithm The proposed heuristic algorithm comprises three stages: the first solves a global covering and balancing problem formulated as an LP model; the second is a construction phase, in which a full solution is obtained by applying a greedy randomized algorithm (guided by the solution of the first phase): and the third is an improvement stage, in which the solution provided by the previ- us stage is used as the input to a cyclic optimization alternating between VNDS and NFO which continues until a feasible solution js obtained; this solution is then improved by means of NFO alone. The first stage is executed only once, while the other two stages are iterated several times to define a multi-start procedure, as illustrated in Fig. 1. This hybrid GRASP-type algorithm will be identified as “Algorithm G++NO". The proposed methodology starts by determining the number of each type of shift that each physician should work over the entire planning horizon, in order to guarantee caverage of all shifts and a workload balance among physicians, based on a fait distribution of the diferent types of shifts (nights, weekends, holidays, etc.). This problem is formulated as a continuous LP problem, which, at a very low computational cost, provides the solution to be used in the next phase. The construction phase is the implementation of a GRASP algorithm to build a solution by assigning shifts to physicians sequentially. The procedure starts with the first day of the plan- ning horizon, assigning all the shifts for that day and progressinglos, F. Mollor and PAE. Meo day by day until a full assignment is obtained. The list of candi- dates for each shift assignment is first defined by the feasibility constraints and then by elitism based on a fitness function. This function takes into account the assignments made so far to all physicians and the theoretical number of shits of each type that each physician should work (obtained as the solution of the LP formulated in the first phase of the algorithm), The full scheduling obtained in the previous phase is improved by alternating VNDS to repair violations of the constraints (re- Quired if the constructive step provides an infeasible solution) With NEO to balance the distribution of shifts and working hours among the physicians. Once a feasible solution is obtained, im- provements to the fair distribution of the workload are sought using NFO only. In the following subsections, a detailed mathematical and algorithmic description is provided for the three components of the heuristic method, 4.2 A linear programming model to solve the general covering prob- lem The purpose of this optimization stp isto obtain the average number Zy of shits of type 5.) 1. lthat shouldbe worked by physicians of type G,, 1, M, in order to cover service demand within the regulatory working hous Variables 2 can be positive only if yy — 1, that i (1 — 7)2y = 0. In adliton, this enerl planing has (0 dstibate the shits among physica a evenly and fairly as possible or which the decison variables Zy must fll the following constrains: ‘ Demand Covering constraint Yonzy =m Vp L as) + Working hours constraint Yoazy such that Z) > 0} (26) | | Then, for each physician P, in the LoC(j,¢), the following recy funtion geht esate ee o) This greedy function measures the difference between the maximum proportion of shifts of type 5, already assigned to a Disician (51/2) and the ratio of shifts assigned eo-a particular Diysician. This vale is taen normalized to the target value forthe Whole planning horizon, Thus the greater the value of g (for Pisician Pte greater hisjher need fo work ths shift m order tomeet the reference valves By definition, this greedy function is a non-negative definite function, However, it could occur that 43 () =0 forall physicians inthe oC.) Enhancement of the greedy function. The greedy function was {defined based only on already assigned shifts of type 5) Neverthe- less, some shifts are important for the even distribution of other _generalshife characteristics among physicians. For example. ifthe shift that is being assigned is a weekend shift and all physicians hhave to work the same number of weekends within the planning horizon; thus, the greedy function must also take into account the consequences of the assignment for the even distribution of weekend shifts. For this purpose, for each set of shifts D. that has to be evenly distributed among physicians and S; € De, the following greedy function go, (i) is defined (om SOs (a Bl (28) where, 2m, = 1% and 25, (2) sm A normalized greedy function gy(i), which ranges in (0,1), is, defined as follows: eu) = —— 80 (20) (wmf ] -min ED) ‘The new enhanced greedy function g (i) is defined as: oO =a (0+ Yan) (at) where the summation is extended to all sets D, of shifts that need to be balanced and that include the shift Then, this greedy function balances the participation of each physician in all shifts and shift characteristics included in the objective function by assigning the shift to the physician who is farthest from meeting all the balancing conditions in which the shifts involved. The balancing assessment takes into account the theoretical values determined by the LP covering problem (Appendix ©). ‘Applied Soft Computing 103 (2021) 107151 43.3, Roulette wheel for the selection of « physician n In the construction procedure of the basic GRASP, the proba- bility pi) of selecting a physician P, © LoC\, ¢) would be chosen at random, with equal probability of being chosen. However, we propose to bias the selection toward the candidates that con- tribute the most to keep the solution balanced at that moment. ‘The idea of using probability distributions different from the uniform was proposed by Bresina [53]. In particular, we use a ower function that extends the polynomial function of order 1 [54] way Lhetacy.n (6 Observe that, if « = 0, we will have a random construction; if.@ = 1, the probability will be proportional to the greedy value. The greater the value of w is, the more elitist the selection ‘mechanism, If all physicians in the LoC G.t) have (i) = 0, then the probability of being chosen is equal amoig them. We recom ‘mend choosing large values for « (> 0.9) to obtain better inital solutions. Pa (32) 4.4, Improvement ofa solution The feasibility of a solution is improved by decreasing the ‘number of unfulfilled ergonomic constraints by means of a VNDS algorithm, which is followed by a NFO procedure to better ful- fil the balancing objectives. These two search mechanisms are applied iteratively (see Fig. 2) until a stop criterion is met (op- timization time or iterations with no improvement). The fol- lowing subsections offer a detailed description of each of these improvement steps. 44.1, Variable neighborhood descent search for repatrng infeasibi- ity Te construction phase is driven by the solution ofthe general covering problem and is particularly oriented toward construct- ing a feasible solution because the Lac in each shift assignment is first defined by physicians who full al constraints. However, in problems with itl slack for finding feasible solutions (too small 2 surplus with respect to the total demand for working hours and very tough ergonomic requirements), the construction phase could provide a solution that fails to meet certain constraints In this ease, the frst step of the improvement phase is a repair process, whereby a shift contributing to the infeasibility of one Physician's schedule is transferred to another physician. These shift transfers successively involve several physicians and are repeated several times. Fig. 3 represents the logic of these move- ments: shift 1, which causes the infeasibility of the sequence $1-S2 in physician P's schedule (after shift S1, there must be a day off), is transferred to physician Pay (causing infeasibility because, two days off are compulsory after shift $7) this requires transferving shift 7 to physician Py (again causing an infeasibil- ity) and this, in tum, results in the transfer of shift $3 to physician Py. After these transfers, the inital infeasibility of physician Preis solved without detriment tothe total numberof non-compliances ofthe remaining physicians, ‘The search for sequences of transfers leading to the improve- iment of the current schedule falls into the category of a VNDS algorithm with rationale as follows Let Xe the set of shifts assigned to physician Pn the incum- beent solution, that is, x, = [5, € (0) IXy = 1} and p{X,.X7) be the distance between solutions for scheduling a physician defined as li.) = KA (3)IM Cle, F. Moller and PAE Marco Applied Spt Computing 103 (2021) 107151 Improvement of X (greedy solu u) < mazter_vxB> ‘a ‘Apply VNDS to plement NFO solution [implement STC 10] X* to got Xr sl Fig. 2 Flowchart ofthe solution improvement phase Infeasible schedule Pryce Phoicin 28 Prosicin Proician 18 Fig. 3 Beample of shit (5°51 Py as Py Pig Ces)”, Cs ee ST) 57) wansfer among physicians (fasion 9, 14, 18,23) on deren days (=D Feasible schedule propos tos th S re 5). Ergonomic requirements fr the cient Iypes of shits: 57 must be followed by two days of Si, $5 must be followed by one day oft and 82, 3. S4 do not require the nextday to bea rest dy. ‘where [X,X/| represents the number of shifts that form part of scinedule X; but not of X/ and those which form part of schedule X} but not of Xi Let us observe that when a physician P, with schedule X; transfers a shift to another physician, the resulting schedule for P,, denoted by Xj. verifies p (X..X}) = 1. AA full schedule X is the aggregation of all the physicians’ schedules: X = (X1.X2.....Xn)e and then, p(X.X’) = Shh 1 (Xi.X;) represents the distance between two schedules for all physicians. The transfer ofa shift from one schedule X to obtain another schedule X’ is denoted by X' = fiX). The schedule solution X’ resulting fom a sequence of k transfers of shifts in which the transferee in one shift eransfer becomes the transferor i the next shift transfer is denoted by X" = K(X). The index p refers to the path p. which determines the transfers of shits between physicians. For example - Fig 3, the pathis Py 223Cfo: Moor and PRE Ato san Pry 24% py Pi. The length of a path isthe number of Wwansfers, in the case of Fis. 3 the length is 3. ‘A neighborhood of depth kis defined as (4) = (X18 pof length k such that X’ = F(X) Let us consider a certain typeof constrain that isnot fufiled by a solution X and thus requires repair. Let Q > 0 denote the ‘maximum number of unfulfilled constraints among all physicians and Po the set of physicians that reach this maximum number of rhon-fulfillments. sam y= {physicians with a number Q of non ~ flillments) ‘A recursive function enables fairly easy implementation ofthis, NDS procedure. In each step, each physician with an infeasible schedule tries to transfer a shift (which is problematic because it causes an infeasibility) to another physician, who is able to accept it, even if this results in an additional infeasibility, and then the infeasibility improvement problem is transferred to another physician, and the process is repeated. The steps of this VNDS algorithm are detailed in Algorithm 3, 4.4.2, A network flow optimization problem for balancing the distri- bution of shifts and working hours The goal ofthis optimization procedure isto transfer shifts of a certain type ftom physicians with surplus workload, who are working sgnfcantly more than average hous and 2 greater Dumber ofthat type of sft to physicans with sac that {ype of shi, who are working significantly fewer than average hours, The term "sgnicantly” is used in felation to a zone of indifference surrounding the average numberof hours worked, which is defined in order to stabilize the procedure as It pro- tresses. A physician P, is considered to have an acceptable (otal of working hours, Hi(X) = Sy, Dy-1 dX, in a schedule X ‘when 3s Tong as it belongs to-thisiterval of indifference. To formalize this ide, fr each iteration of his optimization proce- dle (1 UH (Pa (X) < LH) (PLH < HiX) < UH} The physicians in set Pys (X) can transfer shift, and those in set Pas (X) can receive shifts. Physicians in the balanced set Py (X) can play an intermediate role by both receiving and transferring shifts. This condition for transferring a shift i called the working hhours' condition (WHC). ‘A physician of type G, can transfer a shift of a certain type 5; when the number of assignments of this type exceeds the theo- retical number Z, determined in the pre-processing optimization phase; and, conversely, a physician can receive a shift of a certain {ype when the number of assignments of this type is below (34) (35) ‘Applied Soft Computing 103 (2021) 107151 (eosin, capacity) Fig. Example of work-flow neework Physicians 1,2 and 3 can transfer one Shifts physians 4,3 and 6 can receive and transfer ane shi. and physician 7 and Bean fee one shi this theoretical figure. In terms of the notation introduced in Section 4.2, a physician P is allowed to transfer a shift Sj, when Zp, > Zo, and a physician Ps allowed to receive a shift $, when, Zin, < Za, forall sets D- with relevance in the objective function and in which shift 5; participates. This shit transfer condition is rnamed the balancing shift condition( BSC), Building the network structure. The nodes represent physicians, and each ate (i,k) represents a possible transfer of a shift §, fiom physician P, to physician Pi, The physician P; belongs to set Ps (X), and Py belongs to set as (X) U Pa (X), oF P belongs to set Py (X), and Py belongs to set Ps (X). To plot an arc on the graph, both physicians, transferor and transferee, must meet the conditions WHC and BSC defined earlier and the transferee must be feasibly able to work this shift. When there exists more than one arc verifying the conditions between a pair of physi- cians, one of them is chosen at random (since it is the case that mote than one shift could feasibly be transferred from physician P; to physician Pi). Therefore, the network structure is built randomly and successive iterations of this procedure previde different networks. Assigning demands, capacities, and costs to the network. Nodes representing a physician in Pys (X) have a demand of —1, nodes representing a physician in Py; (X) have a demand of +1, and rnodes representing a physician in Py (X) have a demand of 0 (trans-shipment nodes). The network is expanded by unfolding each node in the set Pr (X), into two nodes that are connected by an arc. All arcs in the network have a maximum capacity of 1 and a ‘minimum capacity of 0. Costs: «= the ares between a physician in Pjs (X) and a physician in Pes (X) have a cost of ~2, ~ the ares between a physician in Pys (X) and a physician in Pp, (X), oF between a physician in Py (X) and a physician in Pes (X) have a cost of —1, « the arcs between nodes representing the same physician in Ph (X) have a cost of 0. Fig. 4 shows a simple example of a flow network with 3 physi- cians in set Pys (X), 3 physicians in set Phy (X), and 2 physicians in set Ps (0.los, F. Mollor and PAE. Meo Step Initialize Ney = 04K = Step L Neer = News +15 keke IHN > MAX op yp hen Ends ‘Applied Soft Computing 103 (2021) 107151 /* Start new iteration to find a new shift-transfer chain */ Compute set Po: if Po = 0 then Feasible solution, Ends (Choose randomly P; € Pg and shift S; € X; causing a constraint infeasibility; SHS k=0; Go 0 Step 2; Sup? k=k+1 Hk < maxeeyensearch then /* Begin the explor jon of the neighborhood of depth k*! if Physician Pr-| X}s = Xp U (Sy} does nor nerease the infeasibiiies of Py then |. Make definitive all temporal transfers and go to Step 4; elseif 3 PhysicianPy- |X}. ‘Transfer temporarily shift Sto Py Select shift Sj € X)» (Sj # Sr) that causes constraint infeasibility to P, infeasibility problem is transferred from P, to Py-*/ bet Spe Go t0 Step 2; ee Solving th | Undo the temporally transfers, keeping the init ‘1 U{S;] does not increase the value of Q then schedule and go to Step 1; Algor 3. YNDS Proce fr rpairing solitons ti based on afer sf contributing tthe nfl of one _hysicin’s schethle to anther pian These ters of shifts sucesively nee several decors amd are repeated eral ines Solving the network low problem. The problem is solved by us ing an algorithm to find the minimum-cost feasible flow. The resulting networks are small in size and can be solved quickly by efficient algorithms such as Network Simplex, Out of Kilter, Cycle Canceling, or Suecessive Shortest Path (see Thulasiaman ct al [55]), Our implementation uses a successive shortest path algorithm, as described in Ahuja, Magnanti, and Orlin [56]. After network optimization, each physician can transfer and receive, at ‘most, one shift. For this reason, this optimization step is repeated ‘maxje-ye0 times. In each iteration, the limits that define the partition of physicians into sets Prs (X), Pys (X). and Phy (X) are ‘modified, starting with small values, which are gradually in- creased. Any fluctuation of the zone of indifference between two values contributes to the variability ofthe created networks and the stabilization of the shift transfers as the algorithm progresses Consecutive iterations of this procedure lead to different net= works, which gradually improve the balancing of shifts and work- ing hours. When this NFO phase is iterated with the VNDS algo- rithm because the solution is still infeasible, the NFO helps the \VNDS algorithm by providing new starting solutions from which {o search for good shift transfer chains (as in a shaking procedure) and also helping to redress any imbalance inthe shift distribution that may be introduced due to the application of VNDS. 5. Computational analysis This section reports the results of the empirical assessment of the algorithm presented above, which was implemented in Java Its practical effectiveness is tested in Section 5.1 by solving the problem of scheduling all the ER shifts for the year 2018 among 42 physicians in the Hospital Compound of Navarre (HCN) in Spain, In addition, in Section 5.2, a set of synthetic scheduling problems of varying degrees of difficulty is used to assess the per- formance of the algorithm under different conditions. The results are compared with those obtained by CPLEX. Finally, Section 5.3 investigates the influence of the different phases ofthe algorithm on the solutions to the physician scheduling problem as well as the value of its parameters for obtaining goad solutions 5.1. The physician scheduling problem at the Hospital Compound of Navarre (HON) The ED of the HCN, which is located in Pamplona (Spain), serves a population of half a million people, and attends to aver 140,000 patients per year. This ED is staffed 24 h per day by 42 board-certified emergency physicians. Currently, each year’s shift schedule is planned manually by one of the physicians, who dedicates three weeks’ work to this task. Although, this person is an experienced physician and has been in charge of schedille planning for many years, the task becomes more com plicated every year, because new labor laws create new con- straints and new categories of workers with different working conditions. This physician creates the schedule without any tech- ‘nological/computational support using only large spread sheets, similar to the one shown in Fig. 5, where there is a row for each physician and a column for each day. Starting with simple rotational rules, the scheduler uses hisfher own heuristics to consecutively balance holiday shifts, weekend shift, nights, an, finally, regular shifts, while also trying. to satisfy a large set of constraints (ergonomic, workload, etc. as described in Sec- tion 3). The resulting schedule violates many conditions as well as provoking numerous complaints from other physicians, who consider the schedule unbalanced and conditioned by subjective preferences. ‘Staff characteristics. The staff comprises 42 physicians who can be grouped into two types: (1) a first group G; of 3 physicians who are exempt from night shifts (denoted by 0. AS. G1, G2 and G3 in