Handling uncertainty in health care management using the

cardinality-constrained approach: Advantages and remarks

Bernardetta Addis a , Giuliana Carello b , Andrea Grosso c , Ettore Lanzarone d,∗ ,

Sara Mattia e , Elena Tànfani f
a
LORIA (UMR 7503 CNRS), Université de Lorraine, INRIA Nancy-Grand Est, France
b
DEIB, Politecnico di Milano, Milan, Italy
c
DEI, Università di Torino, Turin, Italy
d
CNR-IMATI, Milan, Italy
e
CNR-IASI, Rome, Italy
f
DIEC, Unversità di Genova, Genoa, Italy

Received 13 June 2014

Accepted 10 October 2014
Available online 8 November 2014

1. Aim and scope 2. Uncertainty in health care management

The purpose of this short communication is to highlight the The demand for mathematically based decision support tools
potential of a specific robust optimisation approach, namely the to optimise the delivery of health care services has been growing
cardinality-constrained model, in handling health care manage- significantly in recent years. In Western countries, the need for
ment problems. In fact, although robust optimisation approaches efficient resource management is constantly increasing due to the
have been widely applied in health care, the cardinality-constrain- ageing population on the one hand, which is increasing the demand
ed model has seldom been considered. We briefly report our ex- for health care services, and the reduction in public health funding
perience with this approach, which in our view can be fruitfully on the other. Several studies have been conducted with the aim
applied to health care management problems, pointing out some of improving the ratio of service quality to cost in health care
advantages and considerations that can be useful for operations re- systems. Among the adopted quantitative techniques, operations
searchers while applying this approach to the same field. research methods have proved to be effective in many sectors
where decision making support is necessary, such as industrial
planning and logistics, and are currently widely applied to health
care-related problems.
∗
Corresponding author. Typical features of health care systems include a high level of
E-mail addresses: [email protected] (B. Addis), complexity and a large amount of available data, which have only
[email protected] (G. Carello), [email protected] (A. Grosso),
recently begun to be stored digitally. Moreover, one of the major
[email protected] (E. Lanzarone), [email protected] (S. Mattia),
[email protected] (E. Tànfani). problems faced by people working in health care management is
data uncertainty [1]. Available data need to be properly analysed tractable than the stochastic programming approaches. The pro-
and processed to obtain reliable input parameters. In fact, data re- duced solutions may be over-conservative, but the possibility of
liability is a key factor in guaranteeing the feasibility and efficiency managing the convex sets and the parameters’ movement within
of the obtained solution when it is applied to the real system. In them allow the level of robustness to be adjusted.
health care, uncertainty may arise in different contexts and due In the simplest definition of the uncertainty set, the uncertain
to different causes. For example, in the emergency vehicles loca- parameters are assumed to belong to an interval. Hence, when
tion problems, uncertainty arises from ambulance availability [2]; addressing a (possibly mixed-integer) linear program of the form
in planning and scheduling operating rooms, uncertain data are the 
minimise cj xj
duration of surgery, length of stay and rate of patient arrivals [3];
j
in home care services, the demand for care is uncertain in terms
s.t.
of the number, duration and types of visits [4]. In all cases, uncer- 
tainty cannot be neglected, as it may have a significant impact on aij xj ≤ bi ∀i
j
the solution and on the quality of service provided to patients. For
xj ≥ 0 ∀j
this reason, robust methods that explicitly take data uncertainty
into account are drawing the attention of the research community. a nominal value āij and a maximum deviation δij are given for each
uncertain parameter aij , i.e., aij ∈ [āij − δij , āij + δij ].
3. Methods and approaches for robustness The cardinality-constrained robust optimisation method [14]
considers such uncertainty sets. As it is unlikely that all parameters
The concept of solution robustness is crucial when handling will deviate from their nominal value at the same time, the method
uncertain data or parameters. Generally speaking, a solution is assumes also that only a subset of at most Γi parameters is allowed
considered robust if it is able to protect against uncertainty; i.e., to change in the ith constraint. In this sense, the cardinality of
the solution remains feasible even when parameters change with the parameters permitted to change is constrained. The generated
respect to their nominal value, and the deterioration of the solution solution must be feasible for the worst possible scenario in which
quality is controlled or limited when it is applied to the real Γi parameters assume their maximum values while the others
uncertain system instead of the nominal case. Moreover, a trade- assume their nominal ones.
off must be pursued between these two aspects, i.e., the level We remark that very little knowledge of the parameters’ vari-
of robustness (the feasibility in different scenarios defined by ability is required; furthermore, the decision maker can tune the
possible realisations of parameters), and the efficiency and cost of level of robustness by setting the parameters Γi . It is interesting to
the solution (how much the solution deteriorates when applied to note that, as the constraints added by the cardinality-constrained
approach are linear [14], the robust counterpart of an integer lin-
the uncertain system). Indeed, a very conservative solution tailored
ear programming problem can be solved by any commercial solver,
to unlikely scenarios may turn out to be highly expensive for more
except in very large instances. Additionally, the robust counterpart
likely scenarios or, alternatively, a solution that is not conservative
of a linear programming model is still a linear programming model,
enough may become infeasible even for small variations with
and can therefore be solved in polynomial time.
respect to the nominal problem.
The area of robust optimisation is vast and continuously ex-
Different approaches have been proposed in the literature to ad-
panding. For instance, scenario-based approaches are proposed
dress uncertain parameters in optimisation problems. They belong
in [15], whose aim is to produce solutions that have the best
to three main groups: stochastic programming, distributionally ro-
worst-case objective value considering all possible scenarios. In
bust optimisation, and robust optimisation.
adjustable robust optimisation methods [16,17], a multistage de-
In stochastic programming [5,6], uncertain parameters are
cision policy is used. In [18,19] robust optimisation approaches
modelled as random variables and the probability distribution
have been proposed in which the robustness is enforced by cutting
of each parameter is assumed to be known. Hence, to approach
planes. A computational study on cutting planes versus reformu-
real life problems, stochastic programming requires both a strong
lations is also provided in [20]. For other approaches and in-depth
statistical background to manage the mathematical models and a analyses of robust optimisation, we address the reader to Ben-Tal
thorough knowledge of the real problem to derive the probability et al. [21], Bertsimas et al. [22], Gorissen et al. [23], and the refer-
distributions, which are not always easy to derive. The main ences within.
advantage of stochastic programming is that it produces solutions
that are usually not over-conservative as they protect against the 4. Applying the cardinality-constrained approach to health
most likely realisations. The resulting optimisation problems can care management problems
be difficult to solve, however, often involving a very wide number
of scenarios. In addition, if the given distributions are not reliable Several approaches are largely applied when handling uncer-
(i.e., they are different from the real ones), the solutions produced tainty in health care problems, such as probabilistic models [2,24]
by a stochastic method may not prove to be robust. or stochastic optimisation approaches [25,26]. On the contrary, to
On the contrary, distributionally robust optimisation [7,8] and the best of our knowledge, the cardinality-constrained approach
ambiguous chance-constrained approaches [9] assume that the has thus far been rarely applied to health care problems. In May
probability distribution is not known, but lies within a known fam- 2014, a search on Scopus for papers citing Bertsimas and Sim [14]
ily of distributions. The solution must protect against the worst- in relation to health care yielded only five results [27–31] other
case realisation given by the admissible probability distributions. than our contributions in a conference book [32,33], the latter also
The problem is difficult, but computationally tractable approxima- extended in a journal [34].
tions exist. However, the cardinality-constrained approach seems to be
Robust optimisation approaches [10–14] are a good compro- suitable for handling several health care management optimisa-
mise between the above mentioned methods. Robust optimisa- tion problems. We recently tested this hypothesis by applying
tion approaches assume that each uncertain parameter belongs to the cardinality-constrained approach to two relevant health care
a given convex set, and no detailed knowledge of its probability problems: the operating room planning problem and the nurse-to-
distribution is required. These approaches guarantee that the so- patient assignment problem in home care services. In the follow-
lution is feasible for all of the values of the parameters within the ing, we briefly describe our experience in applying this approach
considered uncertainty set. They are usually computationally more to these problems, focusing on general considerations, advantages,
and remarks that can help operations researchers in evaluating the generated savings of up to 30% on penalty costs, considering his-
approach for other applications in the health care field. torical data of a real home care provider over six months. Both
palliative and non-palliative patients are treated by the consid-
5. Our experience in two applications ered provider. Advantages are mainly ascribed to non-palliative
patients rather than to palliative ones because the latter have a
The surgical case assignment problem is considered in [35] more defined care pathway associated with lower uncertainty.
(extended work of Addis et al. [32]). For each specialty, a given Computational times needed to solve the model in practice with a
number of surgical blocks is available, together with their allocable weekly basis are reasonable. Compared to the other methodologies
timespan. Patients must be assigned to these blocks taking into applied to this problem [24,26,36], the cardinality-constrained ap-
account their surgery durations, with the goal of minimising the proach is able to produce competitive solutions. It requires reduced
total delay for patients. The surgery duration is uncertain, and computational time and fewer assumptions on the variability of
the cardinality-constrained approach is applied to compute an patients’ demands. On the contrary, the stochastic programming
assignment of surgery cases to the blocks, which shall not exceed approach [26] requires a high computational time due to the in-
each block’s capacity, even if a subset of surgery durations per clusion of stochasticity with the scenario generation, and the an-
block assumes the maximum value instead of the expected one. alytical approach based on stochastic ordering [24,36] requires
In this way, a certain degree of robustness is obtained without the introducing several assumptions on the shape of the density func-
need to protect against all possible realisations of surgery duration. tions.
The nurse-to-patient assignment problem arises in home care
service management when the continuity of care is pursued [34]. 6. Analysis and general conclusions
Continuity of care means that only one nurse (i.e., the reference
nurse) is assigned to each patient, and the assignment is kept The cardinality-constrained approach proves to behave well
over a long period. The problem consists of assigning each new in the two considered applications, based on which, some of the
patient entering the service to his/her reference nurse. Nurses have advantages of this approach that can be useful for research into
a certain amount of working hours over a defined time interval, other health care management problems are highlighted below.
usually a week, and must be extra paid for working overtime. As already mentioned, this approach uses a very simple ge-
Moreover, a maximum overtime is allowed for each nurse. The aim ometry for the convex sets: parameters are assumed to lie on
is to provide a suitable service quality, minimising overtime costs an interval, and a limited number of them are assumed to take
and/or balancing nurses’ workloads. The amount of treatment time the maximum possible value instead of the nominal one. There-
required for each patient in the time slot may vary with respect fore, the approach does not require the knowledge of the en-
to its expected planned value, e.g., due to sudden changes in the tire probability density functions of the uncertain parameters,
patient’s health conditions. Therefore, the time amount for visits but only some knowledge of them. In health care applications,
required by each patient in each time interval is the uncertain enormous amounts of historical data are usually available, from
parameter of the planning process. We tackled the problem by which the little information required can be easily derived. In
applying the cardinality-constrained approach, with the aim of contrast, it can be difficult to find the more detailed information
guaranteeing that the obtained assignments are feasible even if a required for other approaches. Moreover, unlike other methods,
subset of patients assigned to each nurse requires the maximum the cardinality-constrained approach does not rely on a detailed
treatment time. description of the uncertainty, and for the same reason, it is less
The results of both applications are promising: computational sensitive to estimation errors in the data and their probability dis-
times are reasonable, thus guaranteeing the applicability of the tributions. Finally, the approach provides a robust solution with a
cardinality-constrained approach in real life cases. Moreover, reasonable computational effort.
when tested on a set of randomly generated scenarios and/or on Another relevant advantage is that the basic idea of the ap-
historical data, the solutions display good behaviour in terms of proach can be easily understood by clinicians and health care ser-
quality performance metrics and improvement with respect to vice managers without any background in statistics or operations
their non-robust counterparts. research. The robustness of the solution can be easily tuned by the
Indeed, concerning the surgical case assignment problem, the service managers themselves according to the desired level of risk
robust model has been tested on a set of realistic instances and protection, by imposing the cardinality of the subsets of parame-
has been compared with the non-robust counterpart. The obtained ters assuming their maximum values. Such cardinality has a prac-
solutions, both robust and non-robust, have also been tested tical meaning that can easily be interpreted, and its impact can be
on a set of scenarios. The number of patients operated on may tested to properly select the trade-off between the level of robust-
decrease slightly when robustness is guaranteed; however, when ness and the cost of the solution.
the solutions are applied to scenarios, the robustness dramatically It is worth noting several factors with respect to the param-
reduces – up to one third – the number of patients who cannot be eter setting. Other parameters must be set besides cardinalities
operated on in the scheduled day due to surgery time variations Γi , and their values must be carefully evaluated. Indeed, the ap-
with respect to the expected value. In this way, from the patient’s proach assumes that the majority of uncertain parameters take
point of view, the quality of the solution is highly improved. their expected values, whereas a subset of them takes their max-
Increasing the level of robustness, the number of cancelled patients imum ones. Thus, each nominal value āij must have a real mean-
is almost zero, but on the other hand, the utilisation rate is reduced. ing in the considered application. In the health care applications
Even when the highest level of robustness is required, however, the described above, the lengths of treatment (e.g., surgery and visit
utilisation rate never drops below 50%. durations) are usually continuous parameters, and their nominal
Concerning the home care application, the proposed model has values can be taken as the expected value of a suitable statistical
been tested on both historical data and a set of generated sample distribution, or from standard values given by national or regional
paths. The robust model guarantees a better performance in terms health systems. If an integer parameter is considered, however, a
of overtime costs, as well as a more evenly distributed workload fractional expected value does not occur in any realisation. Regard-
for the nurses, with respect to the nominal model. Briefly, apply- ing the maximum values āij + δij , if a distribution is derived for aij
ing the robust model instead of the non-robust counterpart has based on historical data, attention must be paid when setting the
value of the maximum deviation δij . If we are not willing to pro- [13] D. Bertsimas, M. Sim, Robust discrete optimization and network flows, Math.
tect against some values as they are considered to be very unlikely, Program. B 98 (2003) 49–71.
[14] D. Bertsimas, M. Sim, The price of robustness, Oper. Res. 52 (2004) 35–53.
these values must be outside the given interval. To do so, one pos- [15] P. Kouvelis, G. Yu, Robust Discrete Optimization and its Applications, Kluwer
sibility is to take āij + δij in correspondence of a given quantile of Academic Publisher, 1997.
the distribution rather than the maximum value. [16] A. Ben-Tal, A. Goryashko, E. Guslitzer, A. Nemirovski, Adjusting robust
solutions of uncertain linear programs, Math. Program. 99 (2004) 351–376.
In addition to the benefits, we have experienced two main [17] X. Chen, Y. Zhang, Uncertain linear programs: extended affinely adjustable
drawbacks in conducting our work [34,35] that we consider note- robust counterparts, European J. Oper. Res. 57 (2009) 1469–1482.
worthy. First, the approach may become computationally expen- [18] D. Bienstock, Histogram models for robust portfolio optimization, J. Comput.
Finance 11 (2007) 1–64.
sive for large instances, and computational time may also depend [19] M. Fischetti, M. Monaci, Cutting plane versus compact formulations for
on the value of parameters Γi . Second, characterising the uncer- uncertain (integer) linear programs, Math. Program. Comput. 4 (2012)
tain parameters with only two values, āij and āij + δij , might not be 239–273.
[20] D. Bertsimas, I. Dunning, M. Lubin, Reformulations versus cutting planes
sufficient to capture the behaviour of peculiar parameters. Quite for robust optimization. a computational and machine learning perspective,
straightforward improvements in the method can overcome both Optimization (2014) Online.
drawbacks. Concerning the computational effort, cutting plane- [21] A. Ben-Tal, L.E. Ghaoui, A. Nemirovski, Robust Optimization, in: Princeton
Series in Applied Mathematics, 2009.
based approaches [19] can be applied to reduce computational
[22] D. Bertsimas, D. Brown, C. Caramanis, Theory and applications of robust
time. On the other hand, to provide a more detailed representation optimization, SIAM Rev. 53 (2011) 464–501.
of uncertain parameters, additional side-constraints may be added [23] B. Gorissen, I. Yanikoglu, D.D. Hertog, Hints for practical robust optimization,
to the model. in: Center Discussion Paper 065, 2013.
[24] E. Lanzarone, A. Matta, Robust nurse-to-patient assignment in home care
In general, we think that great advantages can be obtained by services to minimize overtimes under continuity of care, Oper. Res. Health Care
the standard formulation of the cardinality-constrained approach. 3 (2014) 48–58.
Should any of these two drawbacks occur in the solution of a [25] P. Beraldi, M. Bruni, D. Conforti, Designing robust emergency medical service
via stochastic programming, European J. Oper. Res. 158 (2004) 183–193.
specific application, the alternatives we discussed are worthy of [26] E. Lanzarone, A. Matta, E. Sahin, Operations management applied to home care
implementation. services: the problem of assigning human resources to patients, IEEE Trans.
Syst. Man Cybern. A 42 (2012) 1346–1363.
[27] T. Chan, T. Bortfeld, J. Tsitsiklis, A robust approach to imrt optimization, Phys.
References Med. Biol. 51 (2006) 2567–2583.
[28] B. Denton, A. Miller, H. Balasubramanian, T. Huschka, Optimal allocation of
surgery blocks to operating rooms under uncertainty, Oper. Res. 58 (2010)
[1] P. Han, W. Klein, N. Arora, Varieties of uncertainty in health care: a conceptual
802–816.
taxonomy, Med. Decis. Making 31 (2011) 828–838.
[29] M. Holte, C. Mannino, The implementor/adversary algorithm for the cyclic and
[2] L. Brotcorne, G. Laporte, F. Semet, Ambulance location and relocation models,
robust scheduling problem in health-care, European J. Oper. Res. 226 (2013)
European J. Oper. Res. 147 (2003) 451–463.
551–559.
[3] B. Cardoen, E. Demeulemeester, J. Beliën, Operating room planning and [30] C. Mannino, E. Nilssen, T. Nordlander, A pattern based, robust approach to
scheduling: a literature review, European J. Oper. Res. 201 (2010) 921–932. cyclic master surgery scheduling, J. Sched. 15 (2012) 553–563.
[4] E. Lanzarone, A. Matta, G. Scaccabarozzi, A patient stochastic model to support [31] C. Banditori, P. Cappanera, F. Visintin, A combined optimization-simulation
human resource panning in home care, Prod. Plan. Control 21 (2010) 3–25. approach to the master surgical scheduling problem, IMA J. Manag. Math. 24
[5] J. Birge, F. Louveaux, Introduction to Stochastic Programming, Springer-Verlag, (2013) 155–186.
1997.
[32] B. Addis, G. Carello, E. Tànfani, A robust optimization approach for the
[6] A. Shapiro, D. Dentcheva, A. Ruszczyński, Lectures on Stochastic
operating room planning problem with uncertain surgery durations, in:
Programming: Modeling and Theory, in: MPS/SIAM Series on Optimization,
Springer Proceedings in Mathematics & Statistics (Proceedings of HCSE 2013),
2009.
2014, pp. 185–189.
[7] A. Ben-Tal, D. Bertsimas, D. Brown, A soft robust model for optimization under
[33] E. Lanzarone, G. Carello, Applying the cardinality-constrained approach in
ambiguity, Oper. Res. 58 (2010) 1220–1234.
health care systems: the home care example, in: Springer Proceedings in
[8] J. Goh, M. Sim, Distributionally robust optimization and its tractable
Mathematics & Statistics, vol. 61 (Proceedings of HCSE 2013), 2014, pp. 61–72.
approximations, Oper. Res. 58 (2010) 902–917.
[34] G. Carello, E. Lanzarone, A cardinality-constrained robust model for the
[9] E. Erdogan, G. Iyengar, Ambiguous chance constrained problems and robust
assignment problem in home care services, European J. Oper. Res. 236 (2014)
optimization, Math. Program. 107 (2006) 37–61.
748–762.
[10] A. Soyster, Convex programming with set-inclusive constraints and applica-
[35] B. Addis, G. Carello, E. Tànfani, A robust optimization approach for the
tions to inexact linear programming, Oper. Res. 21 (1973) 1154–1157.
operating room planning problem with uncertain surgery duration—an
[11] A. Ben-Tal, A. Nemirowski, Robust convex optimization, Math. Oper. Res. 23
extended analysis. HAL, Hyper Articles en ligne, 2014.
(1998) 769–805.
[36] E. Lanzarone, A. Matta, A cost assignment policy for home care patients,
[12] L. El-Ghaoui, F. Oustry, H. Lebret, Robust solutions to uncertain semidefinite
Flexible Serv. Manuf. J. 24 (2012) 465–495.
programs, SIAM J. Optim. 9 (1998) 33–52.