Health Care Manag Sci (2012) 15:79–90

DOI 10.1007/s10729-011-9183-6

Use of Self-Organizing Maps for Balanced Scorecard

analysis to monitor the performance of dialysis clinic chains
Isabella Cattinelli · Elena Bolzoni · Carlo Barbieri · Flavio Mari ·
José David Martin-Guerrero · Emilio Soria-Olivas · José Maria Martinez-Martinez ·
Juan Gomez-Sanchis · Claudia Amato · Andrea Stopper · Emanuele Gatti

Received: 12 August 2011 / Accepted: 25 October 2011 / Published online: 15 November 2011
© Springer Science+Business Media, LLC 2011

Abstract The Balanced Scorecard (BSC) is a vali- space (map), thus providing a compressed representa-
dated tool to monitor enterprise performances against tion. The SOM unsupervised (self-organizing) training
specific objectives. Through the choice and the evalu- procedure results in a map that preserves similarity
ation of strategic Key Performance Indicators (KPIs), relations existing in the original dataset; in this way, the
it provides a measure of the past company’s outcome information contained in the high-dimensional space
and allows planning future managerial strategies. The can be more easily visualized and understood. The
Fresenius Medical Care (FME) BSC makes use of 30 present work demonstrates the effectiveness of the
KPIs for a continuous quality improvement strategy SOM approach in extracting useful information from
within its dialysis clinics. Each KPI is monthly asso- the 30-dimensional BSC dataset: indeed, SOMs en-
ciated to a score that summarizes the clinic efficiency abled both to highlight expected relationships between
for that month. Standard statistical methods are cur- the KPIs and to uncover results not predictable with
rently used to analyze the BSC data and to give a traditional analyses. Hence we suggest SOMs as a reli-
comprehensive view of the corporate improvements able complementary approach to the standard methods
to the top management. We herein propose the Self- for BSC interpretation.
Organizing Maps (SOMs) as an innovative approach
to extrapolate information from the FME BSC data Keywords Artificial Neural Networks ·
and to present it in an easy-readable informative form. Balanced Scorecard · Dialysis · Healthcare ·
A SOM is a computational technique that allows pro- Quality control system · Self-Organizing Maps
jecting high-dimensional datasets to a two-dimensional

1 Introduction

Cases of End Stage Renal Disease (ESRD) requiring

dialysis treatment are becoming more frequent world-
I. Cattinelli · E. Bolzoni · C. Barbieri (B) · F. Mari ·
C. Amato · A. Stopper · E. Gatti wide [27, 33]. In the particular case of Europe, preva-
Fresenius Medical Care, Bad Homburg, Germany lence and incidence of ESRD are increasing making
e-mail: [email protected] the kidney disease emerge as a crucial health, social,
and economic concern [15]. Therefore, governments
E. Gatti
Centre for Biomedical Technology at the Danube, and healthcare companies must aim to apply adequate
University of Krems, Krems, Austria strategies that allow the highest quality at the best cost,
and that ensure the patient well-being as a manda-
J. D. Martin-Guerrero · E. Soria-Olivas ·
J. M. Martinez-Martinez · J. Gomez-Sanchis
tory duty. Currently, Fresenius Medical Care (FME)
Department of Electronic Engineering, European activities involve more than 400 dialysis cen-
University of Valencia, Valencia, Spain ters located in 19 countries [31]. This wide distribution
80 I. Cattinelli et al.

implies a high variability concerning both the different external standards (i.e. European Best Practice Guide-
governments’ healthcare systems and the heterogeneity lines). According to this framework, each clinic is (on
of the dialytic population [6]. To maximize the results, a monthly basis) associated with 30 values, each repre-
FME has chosen a continuous quality improvement senting the score of that clinic with respect to one single
strategy which combines clinical enhancements with KPI (see Section 2 and [31]).
management benchmarks [31]. To address this purpose, The use of the BSC within FME is now fully estab-
a robust warehouse of clinical, operational and financial lished as the main instrument for performance moni-
data was designed in the past years and continuously toring: monthly reports are generated that detail the
implemented [19, 29]. This framework represents the scores of single KPIs, focusing on groups of clinics
fundamental decision support for the use of a continu- based on their geographical area, on single clinics, and
ous performance monitoring system. even on single patients. The management, based on
Over the years, several methodologies have been such reports, can ask for more information on the
presented for business performance measurement, in- clinics that show defective performance under some
cluding techniques based both on reliable mathematical KPI, and take corrective actions if needed; on the other
algorithms and on more empirical approaches [2, 3, 13, hand, clinics whose KPI scores are in the excellence
20, 28]. area can receive incentives. Currently, BSC reports
Given the complexity of reconciling clinical and make it possible to look at the monthly trend of single
financial requirements, FME has been adopting since KPIs (or perspectives) separately, or alternatively at
2007 the Balanced Scorecard (BSC) approach in order the whole set of KPIs within the same month. However,
to successfully align the business strategy of the organi- there is no simple way for the management to compare
zation at any level (as extensively described in [31]). data for a group of selected KPIs over a large time
The BSC was first introduced in 1992 as a powerful window. In particular, one cannot, based on the infor-
tool to monitor and align the performance of all the mation provided by standard reports, identify groups
branches of a single enterprise [12]. This methodology of clinics based on their performance on combinations
differentiates from the traditional accounting measures of KPIs—that is, clinics with a well-defined behavior,
in that it combines financial measures (which give infor- characterized by correlated scores for a set of KPIs.
mation about actions taken in the past) and operational In other words, the relations existing between KPIs
measures which will drive the future performance. The cannot be easily extracted with the currently employed
operational measures include aspects like customer analytical tools. Moreover, even if such information
satisfaction, internal processes, and innovations. This was to be processed with ad-hoc analyses, it still would
gives a balanced view of the general efficiency of a remain difficult to effectively convey the results to the
company—not only of its productivity—providing cues management, as the nude numbers do not offer an
for the future plans of business. In general, a BSC intuitive depiction of relations across KPIs, especially
identifies main areas of business in which the executives when dealing with high-dimensional data: in this case,
should address the improvement effort of the company “a picture is worth a thousand numbers”. Thus, there is
and that are usually named perspectives. The perspec- a need for analytical techniques that can easily extract
tives have the aim to guide the BSC implementers in the interesting correlation patterns on groups of KPIs, and
selection of those Key Performance Indicators (KPIs) at the same time offer an effective visualization of such
that are crucial for tracking the whole company growth complex information. To this end, we propose here the
and that, therefore, embrace both financial and non- use of Self-Organizing Maps (SOMs).
financial topics. Basically, the use of a BSC implies A SOM is a type of Artif icial Neural Network (ANN)
the definition of strategic KPIs describing the company for data clustering and visualization [14]. An ANN is
requirements of efficiency and the evaluation of each a computational model inspired by the basics of hu-
of them with reference to a real or hypothetical best man brain functioning. In general, an ANN employs
standard model. The final score for every KPI draws connections between its processing units (called neu-
a picture of the whole company achievements. Over rons) for storing the knowledge required to perform
the years, the BSC has been modified to be adopted some specified task. The fundamental characteristic of
as a conceptual framework in totally different organi- an ANN is the ability to learn from the environment
zations, including healthcare organizations [11, 36]. and to improve its performance in accordance with a
In the particular case of FME, the use of this tech- prescribed model that constitutes the learning para-
nique implies the selection of long-term objectives, digm [10]. In particular, SOMs operate to produce a
clear perspectives, and specific KPIs in conjunction low-dimensional (typically 2D) representation of high-
with the respect of both internal requirements and dimensional data (such as records containing several
Use of SOMs for Balanced Scorecard analysis in dialysis clinics 81

variables, which in our case correspond to the 30 KPIs) demonstrate that SOMs, providing a compact and unbi-
by identifying data that are similar in the input space, ased representation of complex datasets, are a valuable
and grouping them on a grid [14]. method to complete the analysis performed on BSC
Beyond its consolidated reliability, the choice of with traditional statistics. Indeed, SOMs gave more
SOM was driven by two main advantages of this insights on the role of the different KPIs in driving the
technique. First, SOMs enable to summarize large clinic performance and highlighted unpredicted rela-
collections of complex data in a compact and easily tions and dynamics existing among the KPIs.
interpretable graphical representation. Moreover, the
modelling approach of SOMs is unsupervised, meaning
that no a priori hypotheses need to be injected by the 2 Balanced Scorecard and Key Performance
user. Results are, therefore, data-driven and unbiased. Indicators definition
This allows unanticipated relationships between vari-
ables to freely emerge. It must be noted that the SOM The FME Balanced Scorecard framework is based on
is by no means the only technique proposed in the four main perspectives (i.e. relevant topics of business
literature for performing dimensionality reduction and or area of improvement): i) patients, ii) employees,
data visualization [16]: other popular methods include iii) shareholders, and iv) the community. For each
Principal Component Analysis [23], Sammon’s non- perspective specific quality goals (KPIs) have been
linear mapping [26], Isomap [32], and Locally Linear defined. For each KPI a target of excellence has been
Embedding [25], just to name a few of them. However, selected, according to healthcare, financial, and man-
the simplicity and intuitiveness of the SOM makes it agerial guidelines, so that each KPI value can be scored
preferable for our aims, as the ability to produce an with reference to its excellence target. More precisely,
easily-readable map that immediately conveys crucial the numeric raw data concerning every single KPI are
information content in the data is a key factor for sup- collected every month (extracted from the FME clin-
porting the use of SOMs in the everyday management ical/financial data warehouse) and elaborated, so that
practice, which was in fact one of the main goals of a score can be associated to the final value of each
our work. indicator. The closest the KPI value is to the target of
SOMs have already been applied in the healthcare excellence, the highest its score will be for that month.
field, for population studies [1], clinical diagnosis [18, On the whole, the four perspectives embrace 30 KPIs
22], and for organization [17] or economic consider- that selectively describe i) patients’ outcomes (i.e. satis-
ations [21, 24]. Herein, we propose the application faction, compliance and prolonged life expectancy); ii)
of SOMs to the FME BSC for cost-benefit analyses personnel qualification and its continuous professional
and for company efficiency evaluations. We mean to growth; iii) financial control and company develop-

Table 1 KPIs, and corresponding perspectives, as defined in the FME BSC

Patient perspective Employee perspective Shareholder perspective Community perspective
High flux dialysis Turnover of personnel Treatment growth Accidents to employees
(per 1.000)
HDF online dialysis Absenteeism Patient growth Patient education and
support program
eKt/V ≥ 1.0/1.2/ 1.7 Overtime New patient inflow ISO 9001 and equivalent
Hgb ≥ 10 and ≤13 g/dL Employee satisfaction survey Scheduling efficiency ISO 14001 certification
Vascular access (native fistula) Training hours Personnel costs Compliance program
Treatment adequacy Other costs Contaminated waste:
kg per treatment
Peritonitis rate Electricity consumption:
kWh per treatment
Reporting compliance Water consumption:
liter per treatment
Patient satisfaction survey
Patients at risk for
HepB infection
Seroconversion HepB-C
82 I. Cattinelli et al.

ment; iv) enterprise’s social responsibilities (e.g. energy were discarded: High flux dialysis, HDF online dialysis,
savings and preservation of the environment). Table Peritonitis rate, Reporting compliance, Absenteeism,
1 lists the 30 KPIs grouped based on the perspective Overtime, Patient education and support program,
they belong to. An exhaustive description of the imple- Electricity consumption, and Water consumption. The
mentation process of the FME BSC has been reported other KPI data were then normalized to mean = 0 and
in [31]. standard deviation = 1.
BSC data are collected for 19 European countries.
The present work is focused on data from the Italian
clinics (30 clinics monitored from January, 2008 to 3.2 SOM theory
April, 2010), which we have chosen as our case study
to demonstrate the potentialities of the SOM approach Each neuron in the SOM is represented by a vector
on this kind of data. of D real values, where D is the number of dimen-
sions of the input data: w j = (w j1 . . . , w jD ) ∈ R D for
j = 1, . . . , q (q being the total number of neurons in
the grid that is one parameter to be tuned). Such vector
3 Self-Organizing Maps for European clinic is called the weight vector: its components are called
performance evaluation weights, each corresponding to one component of the
input vectors. Notice that in our case D = N, as we
As mentioned above, a SOM is organized as a grid of have as many weights as KPIs. Weight vectors are to
computing units, called neurons. Neurons are trained to be interpreted as prototypes of the original data: a
effectively map the input space: in other words, neurons competitive learning process takes place whereby the
learn how to position themselves in the space so that prototype that is closest to a given input vector “wins”
regions of similar input data are modeled by neurons the vector itself. Therefore, in our case each neuron will
that are close in the grid (that is, neighboring neurons). stand for a set of similar KPI records, thus representing
a specific region in the 30-dimensional input space.
3.1 Clinic data After the neuronal grid has been initialized, the
learning process takes place. For each input vector,
The SOM algorithm takes, as input data, N- the Best Matching Unit (BMU—that is, the neuron
dimensional real-valued vectors xi : whose weight vector is closest to that input vector, and
therefore the winning prototype) is identified:
X = {x1 , . . . , xm },
 
jt∗ = argmin d x(t), w j (1)
where j=1,...,q

xi = (v1 , . . . , v N ) ∈ R N for i = 1, . . . , m
where x(t) is the input vector presented at current step
t, and d is a suitable distance function (typically, the
In our case, each vector xi represents a KPI record, that
squared Euclidean distance is used). The network is
is, it collects KPI scores for one clinic in one month.
trained by moving the weight vector of this BMU, and
As our dataset consisted of 30 clinics, each monitored
those of its neighboring units, even closer to the cur-
over 28 consecutive months, we have a total of m =
rently presented input pattern (Fig. 1). More precisely,
840 KPI records (that is, input vectors). Each input
let x(t) be the currently presented input vector, and
vector contains one score, vl , for each KPI, with N
w j(t) be the current weight vector for a generic neuron
denoting here the number of considered KPIs; thus,
j. The new set of weights after input presentation is
N ≤ 30 (N = 30 when all the KPIs are taken into ac-
computed according to:
count). Notice that KPI scores can take values between
0 (worst performance) and 100 (best performance).  
Prior to submitting them to the SOM algorithm, data w j(t + 1) = w j(t) + α(t)h j∗ j(t) x(t) − w j(t) (2)
were preprocessed so that those KPIs for which no data
were available, or that were almost constant (i. e., low Here, α(t) is the learning rate: this factor controls
standard deviation) were discarded; in fact, a constant how strongly the new weights are moved in the direc-
KPI means that all clinics in every month perform the tion of the input pattern. The learning rate is usually
same on that parameter, and therefore that KPI cannot decreased with training steps, to allow for the map
be used to distinguish among clinics having different to settle on its final configuration. h j∗ j(t) is the cur-
behaviors. Based on these criteria, the following KPIs rent neighborhood function, determining the extent to
Use of SOMs for Balanced Scorecard analysis in dialysis clinics 83

Once this final SOM has been selected, a number of

operations for both visualizing the SOM, and extracting
additional information from it, can be performed. First
of all, the component planes of the SOM can be plotted,
so that information regarding each single KPI can be
visualized. A component plane of a SOM is a map
where, for each neuron, only one component of its
(a) (b) weight vector (corresponding to a given input variable)
Fig. 1 Toy example of the basic functioning of a SOM. In panel is shown (see for instance Fig. 2); thus in our experi-
(a) the initial neuronal grid is shown. The first pattern, repre- ments we can display a total of N component planes. In
sented here with a cross sign, is presented to the network. Its the component plane i, each neuron in the SOM grid is
position with respect to the grid represents the similarity of this colored based on the value of the i-th component of its
input vector to the weight vectors of the SOM units. The weight
update process (panel (b)) drags the winner unit (the one closest weight vector; higher values are usually depicted in red
to the input, here shown in bold) closer to the presented pat- and lower ones in blue. In this way, similarly colored
tern; also neighboring units to the winner neuron (here marked neurons within a plane represent a set of KPI records
in gray) are affected by this weight update that are similar according to that specific KPI. The same
region within every map (say, the upper left corner)
identifies the same set of records but different maps
focus on a different KPI: for instance, one plane could
show values relative to the Treatment Growth KPI,
which units that are close to j∗ are dragged toward the
while another shows values for the Patient Growth KPI.
input pattern by the update step.
This kind of visualization allows comparing values over
By means of this competitive learning dynamics, the
different KPIs: for instance, one could easily identify
neurons of the SOM end up covering the input space
regions of the input space where a given KPI i1 takes
in such a way that the neighborhood relationships are
large values (a red region in plane i1 ) whereas KPI i2
mostly preserved (that is, close regions in the output
has small values instead (a blue region in plane i2 —
map correspond to close regions in the input space). As
negative correlation). Therefore, the comparison be-
typically the number of neurons in a SOM is smaller
tween different planes allows for an intuitive grasp of
than the size of the input dataset, at the end of the train-
the relationships existing between the variables under
ing process a given neuron will be the BMU for a group
study.
of input vectors—that is, that neuron weight vector is
Additional information is provided by the BMU
the prototype that sums up the common features of the
map, where one can see how many input vectors were
data falling into that region of the input space.
“won” by each neuron, and therefore which regions of
the SOM contain more data points. In a BMU map
3.3 Implementation (see for instance Fig. 2, bottom part) each neuron
of the SOM is represented with a size that is pro-
In this work, the SOM analysis was carried out in portional to the number of records that it represents
MATLAB using the SOM toolbox [34]. In order to (i.e., the records that it “won” as best matching unit).
ensure that a good quality map is produced, for each The largest neurons host the most records, while the
analysis we generated 4008 different maps. These smallest ones denote those regions of the SOM that are
differed in the way the weight vectors were initialized less interesting to the user as they are almost record-
(either randomly or linearly), in the type of algorithm less. Thus, the BMU map supports the interpretation
used for training (either sequential or batch), and in of the component planes as it highlights those regions
the neighborhood function employed in weight updates that are mostly worthy of attention when looking for
(e.g. gaussian). Among these maps, the best one was correlations across planes.
then selected. The quality of a SOM is typically eval- An additional clustering step on the SOM vectors
uated based on two error measures: the quantization was also performed, as this provides a higher-level
error, qe, (telling how well the prototypes represent the description of characteristic features of larger groups
input data) and the topographic error, te (measuring of KPI records, with respect to the more fine-grained
the degree of preservation of neighborhood relations). information encoded by single neurons. The weight
As both these measures are relevant to our aims, we vectors of the SOM were automatically grouped based
decided to choose as our final map the one minimizing on their similarity by employing the k-means algo-
the product of te × qe. rithm [5] (the number of clusters was set to k ≤ 5 for
84 I. Cattinelli et al.

Fig. 2 Component planes and BMU map for the SOM analysis to the smallest one; therefore, the maximum value for one KPI
over KPIs belonging to the Patient perspective, and to the Share- might not correspond to dark red for all planes. Also notice
holder one. The BMU map, reported for completeness, highlights that for KPIs marked with a *, red regions correspond to good
the regions of the map where most data reside. Here, a hexagonal scores which in turn are obtained for low raw values on those
lattice was chosen for the SOM, meaning that each neuron parameters: for instance, a red region in the HepB infection risk
(except those at the edges) has six adjacent neighbors. Notice component does not mean that the risk is high, but rather that
that color scales are normalized to the same interval, where 100 the performance regarding such aspect is a good one—that is, the
corresponds to the largest value over all weight vectors, and 0 infection risk is low

better readability of the results). This clustering step jectories superimposed on the SOM, so that temporal
defines macro-regions over the SOM, which can then trends could be intuitively inferred.
be characterized, for instance, as “positive” or “nega-
tive” clusters (that is, clusters collecting KPI records
having overall high scores vs those characterized by
lower scores) based on the weight vector values of 4 Case study
neurons falling inside them; in this way, one can get a
high-level categorization of larger sets of KPI records. The purpose of the present work is to demonstrate the
Finally, as our data is temporal in nature (records are power and the reliability of the SOM technique for the
taken on a monthly basis), we investigated how the KPI analysis of data concerning dialysis clinic performances
scores for individual clinics evolve in time. This was by discussing some representative results obtained for
done by identifying, for a given clinic, the sequence of FME BSC data. To this end, we chose to study the case
its BMUs, and then visualizing that evolution as tra- of Italian clinics.
Use of SOMs for Balanced Scorecard analysis in dialysis clinics 85

The choice of SOMs for our study was driven by contains KPI records characterized by high scores on
the well known potential of this technique [10, 14]. In KPI 1 and low scores on KPI 2.
fact, this method is not only able to elaborate complex First of all we verified the well known correla-
information summarizing it in an easy-readable two- tion between eKt/V and Treatment Adequacy, Treat-
dimensional map. Indeed, its main advantage for our ment Growth and Patient Growth, Vascular Access
purposes is the capability to extract important infor- and Hemoglobin. As expected, a positive correlation
mation from the data without any a priori assump- emerged between the scores for eKt/V and those for
tion. Hence, thanks to its unsupervised training process Treatment Adequacy, as shown by a strikingly similar
SOMs allow unpredicted outcomes to freely emerge distribution of values over the two maps. This direct
without the risk of artefacts induced by the user’s correlation was predictable since eKt/V, also defined
intervention. as dialysis dose [4, 8], is one of the main parameters
It is important to stress that the easy-readable rep- used for treatment efficacy evaluation; then, eKt/V and
resentation of multidimensional datasets provided by Treatment Adequacy KPIs measure different aspects
the SOMs preserves their main information content. of the same target. This result confirms the reliabil-
Hence SOMs offer a less complex but highly informa- ity of SOMs in extracting relations that actually exist
tive visualization of multivariable data that supports within the dataset. The effectiveness of the SOMs in
the BSC users and managers in the interpretation of highlighting real correlations is also stressed in the same
the performance measurements. Therefore, aim of our panel when the Treatment Growth and Patient Growth
study is not only to share the outcomes of the specific maps are compared. A predictable strong correlation
case study herein presented. On the contrary, the main between the two KPIs is, in fact, showed. Another inter-
goal of this work is to provide evidence that, given its esting outcome applies to the Hemoglobin and Vascu-
flexibility and reliability, the SOM methodology may lar Access KPIs. Indeed, a less sharp but still quite pre-
find a wider utilization in the interpretation of data dictable correlation can be seen when comparing these
coming from the BSCs of any healthcare enterprise, two maps. It has already been suggested that dialysis
with the potential of becoming a standard analytic tool performed through temporary subcutaneous catheters
for efficiency measurement. or grafts raises blood loss, inflammation and infection
As expected, SOMs enabled to highlight both un- events, all conditions which tend to deteriorate the
expected and more predictable correlations between patients’ anemic status and quality of life [30, 35].
KPIs. Interestingly, some other correlations that we Moreover, vascular accesses other than the permanent
were expecting to find were not confirmed by the arteriovenous fistula increase the risk of resistance to
SOM analysis, and this underlines that no preliminary the erythropoiesis stimulating agents’ therapy (such
assumption biased the outcomes. as recombinant human erythropoietin) as well [7, 9].
In a first analysis, we conjointly analyzed KPIs from Ultimately, this results in poor hemoglobin plasma
both the Patient and the Shareholder perspective, for a levels.
total of 13 KPIs. The component planes for the relative After this first assessment, we further examined the
SOM are shown in Fig. 2, together with the BMU map. maps in Fig. 2 by analyzing the overall information
In interpreting these and the following maps, one must therein contained. If we focus on the upper left corner
keep in mind that a SOM groups similar KPI records of the maps, it can be noticed that the highest Patient
(that is, vectors of KPI scores for one clinic in one Satisfaction can be found for the lowest values of eKt/V
month) by assigning them to the same neuron in the and Treatment Adequacy. This surprising outcome
grid; each neuron is represented by a weight vector could be explained by the fact that the goodness of the
that is the prototypical KPI record residing in that treatment implies conditions (such as longer or more
neuron. Each component plane of a SOM acts as a frequent sessions) that tend to decrease the patients’
separate filter on the SOM itself, showing only infor- compliance. This strongly suggests that the patients
mation pertaining to one KPI. In other words, the i-th do not always perceive an effective treatment as the
component plane shows only the i-th component of its key factor for a satisfactory quality of life. Rather, a
weight vector. Although we have multiple component more relevant factor seems to be a correct management
planes, these all correspond to the same SOM, and thus of the anemic condition (intended as low hemoglobin
the same hexagon (neuron) in all planes represents the values) that might affect the patient. In fact, comparing
same set of KPI records. Therefore, if a region is found the Hemoglobin and the Patient Satisfaction maps, it
that has a reddish coloring in component plane 1, and appears evident that low levels of hemoglobin often
bluish coloring in component plane 2, then that region correspond to a poor contentment in the patient, and
86 I. Cattinelli et al.

vice versa; this observation confirms the main role of to a clustering step. Figure 3b reports the trajectory
appropriate hemoglobin values in the dialysis patients’ of one clinic as an example; in this case, five clusters
welfare. were found on the SOM. The considered clinic had an
Comparing KPIs across the two different perspec- almost stable behavior over time, as its KPI records
tives (Patient and Shareholders), it is of interest under- remained mostly inside one cluster; in other words, its
lining that both the Patient Growth and the Treatment performance did not change much during the 28-month
Growth KPIs tend to be in general independent of window that we considered. Computing the average
medical performances: this suggests that a large frac- SOM (that is, a map where each neuron has a color
tion of the considered clinics is able to deal with an that is proportional to the average of its weight vector
increased work load while keeping high medical stan- as shown in Fig. 3a) helps characterize this cluster as
dards. On the other hand, a high rate of new patient containing KPI records corresponding to intermediate
income corresponds in some cases to a higher risk of clinic performance. As stated above, in Fig. 3 the result
hepatitis B infection (note that high risk is indicated by for a generic clinic of ours is described as an example.
low KPI scores, blue colors), and to suboptimal scores Additional insights into possible causes of an observed
in the Hemoglobin KPI. These partial inverse correla- behavior can be obtained by tracing back other rele-
tions find a possible explanation in that when new pa- vant features of the considered clinic (e.g. number of
tients are accepted in a FME clinic to start the therapy patients, clinic operating since).
they might display conditions that require time to be Similar steps were carried out for a second study
corrected. This comprehensibly delays the achievement where a new SOM was trained for studying the Em-
of optimal conditions and satisfying therapy outcomes. ployee and Community perspectives. The component
From another point of view, the inverse correlation planes for this SOM (Fig. 4) reveal remarkable and
between New Patient Inflow and viral infection risk unexpected relationships among the Employee KPIs,
does mean that a very good vaccination prophylaxis is confirming that the outcomes were not affected by any
implemented in the FME clinics. a priori assumptions.
The temporal evolution of single clinics over the Focusing on the Employee Satisfaction plane, two
considered variables can be analyzed by computing main regions can be identified. In the upper part of
their trajectories on the SOM. To facilitate an intuitive the map the clinics for which the personnel disclosed
understanding of the meaning of a trajectory, it is useful a poor gratification of its job are grouped (blue color);
to project it on the SOM after this has been submitted on the contrary, the lower region of the map clusters

KPI average Clinic trajectory on clustered SOM

83.9

75.2 3

66.6
(a) (b)

Fig. 3 In panel (a), the average SOM for the Patient-Shareholder panel (b) the trajectory of KPI records over time for one clinic is
KPI analysis is shown. The value of each unit is the average of reported, and superimposed on the clustered SOM. Clusters were
that unit’s weight vector. Notice that this is a weighted average, obtained by running the k-means algorithm on the weight vectors
since the FME Balanced Scorecard defines different weights for of the SOM
each KPI, based on their relevance for overall performance. In
Use of SOMs for Balanced Scorecard analysis in dialysis clinics 87

Fig. 4 Component planes

and BMU map for the
SOM analysis performed
on the Employee and
Community KPIs

clinics where the staff satisfaction is greater (yellow- Training Hours and the Accident to Employees KPIs.
red color). An intuitive expectation could have been Indeed, the emerging evidence is that there is no direct
that a higher turnover should correspond to a lower correlation between the level of the workforce prepa-
workers’ contentment. However, this assumption does ration and the capability to prevent injurious events in
not find confirmation in the SOM analysis, which shows the clinics. In this case, a deeper analysis of subjects
no direct correlation in the trend of these two KPIs. and contents of the educational sessions taken by the
In fact, observing the Turnover of Personnel’s map, it employees may be required to further investigate this
is evident that clinics with a high personnel turnover hypothesis.
(i.e. with a low score for this KPI, displayed in blue- The analysis on clinic trajectories introduced above
green color) are all clustered in the left corner of the can be applied to this case as well: the example shown
lower part of the map and that the trend ameliorates in Fig. 5 illustrates the temporal dynamics of one clinic
for clinics grouped in the upper and right section of the that started with mediocre performances, but tended to
map. This evidence might have different explanations; improve in later times.
for instance the procedure of investigation may not be Figures 3 and 5 do show that the potential of the
optimal. This could mean that the satisfaction survey SOM technique applied to the BSC analysis does
is not properly formulated and therefore fails to reveal not exhaust with the information provided by compo-
the real thought of the workers. Another explanation nent planes. Additional insights on the dataset can be
could lie on the timing of the survey; basically, whether gained by clustering the SOM itself, therefore iden-
the employees’ opinions were gathered early after the tifying macro-groups of KPI records that, on aver-
hiring or later in their career might affect the outcomes. age, share similar features. This allows detecting high-
Similar considerations could be inferred comparing the level trends in the data, such as groups of clinics that
88 I. Cattinelli et al.

Fig. 5 Average map (a) and KPI average Clinic trajectory on clustered SOM
clustered map together with 88.4
one clinic trajectory (b) for
the SOM analysis on KPIs 5
belonging to the Employee
perspective and to the
Community one 4

77.9 3

1
67.5
(a) (b)

consistently have top performances. This is especially ment these basic analyses with an innovative practical
useful when paired with an analysis of the temporal application of a well-known and validated computa-
evolution of records: we showed that it is possible tional technique, such as the Kohonen’s SOMs, to the
to reconstruct the trajectory of KPI records for one FME BSC.
selected clinic over time, so that relative improvements We have been able to provide evidence of the in-
can be tracked. Future work will focus on developing novation offered by the SOMs: indeed, SOMs allowed
a deeper analysis of cluster transitions based on such a comparative analysis among FME KPIs and per-
record trajectories: the aim is to infer from the ob- spectives over a large time lapse (not only month by
served, past records which behavior can be reasonably month, as the currently available reports on the BSC
expected from clinics in future months. Such prediction allow) highlighting relationships that could not easily
might then be used to take corrective actions to make be inferred before; moreover, it was possible to track
sure that clinics always maintain a high standard of clinic improvements (or declines) so suggesting future
performance. In general, in the light of a continuous interventions for business policy corrections. The extra
improvement policy, the SOM analysis proposed here value of SOMs in this context lies in that identified
naturally lends itself to effectively identify both areas of correlations among a set of KPIs can suggest a potential
excellence and aspects that might need improvements, causal link that was not apparent to the management
also suggesting possible strategies for intervention. at first (as standard reports do not support this kind
While ours can be considered a preliminary, pioneering of inference). For instance, a report showing an in-
study for this specific data domain, we believe that crease in accidents to the personnel would suggest the
the presented results provided evidences that support, need for corrective actions, but it would not provide
and indeed encourage, the adoption of SOM analy- any more clues for a focused intervention. On the
sis as a standard analytic tool for clinic performance other hand, when considering the component planes for
monitoring. the Training Hours and Accidents to Employees KPIs
(Fig. 4), no direct correspondence emerged between
the number of training hours and the capability of the
5 Conclusions personnel of avoiding injurious events. This result has
practical implications as it suggests that merely increas-
The BSC has been successfully adopted for several ing the number of training hours might not be enough
years by FME to monitor its clinic efficiency. At to improve safety. Rather, either the key factor for the
present, the BSC analysis monthly provides an evalua- high rate of accidents lies in the contents of the training
tion of financial and operational parameters with refer- program, or it lies in some totally different aspect: thus,
ence to specific perspectives and KPIs. This approach SOMs, narrowing the spectrum of possible interven-
gives a punctual depiction of every single care unit’s tions, provide more clues to guide the management
performance for the considered month. With our work strategy for corrective actions. Hence, what emerged
we have successfully explored the possibility to comple- from our particular case study is largely beyond a sim-
Use of SOMs for Balanced Scorecard analysis in dialysis clinics 89

ple theoretical speculation: indeed, the SOM concretely 11. Inamdar N, Kaplan RS, Bower M (2002) Applying the
demonstrated to be a powerful integrative method to balanced scorecard in healthcare provider organizations.
J Healthc Manag 47(3):195–196
the traditional interpretation of an active BSC.
12. Kaplan RS, Norton DP (1992) The balanced scorecard—
We would like to stress that the applicability of the measures that drive performance. Harvard Bus Rev 70(1):
SOMs does not exhaust in the specific case study above 71–79
presented; on the contrary, given their flexibility, we do 13. Kaplan RS, Norton DP (1996) The balanced scorecard: trans-
lating strategy into action. Harvard Business School Press,
not exclude, rather we encourage, a wider utilization of
Boston, USA
the SOMs also for the study of data coming from the 14. Kohonen T (2001) Self-organizing maps, 3rd edn. Springer
BSCs of totally unrelated areas of business. Hence, the 15. Lameire N, Jager K, Van Biesen W, de Bacquer D,
main goal of our work was not to provide unquestion- Vanholder R (2005) Chronic kidney disease: a European per-
spective. Kidney Inter, Suppl 68(99):S30–S38
able truths over a single case study; on the contrary,
16. Lee JA, Verleysen M (2007) Nonlinear dimensionality reduc-
sharing our results we aim to propose the SOMs as a tion. Springer
powerful tool to extrapolate cues for further considera- 17. Lloyd-Williams M, Williams T (1996) A neural network ap-
tions and analysis on general enterprise performances. proach to analyzing health care information. Top Health Inf
Manage 17(2):26–33
18. Makinen VP, Forsblom C, Thorn LM, Waden J, Gordin
Acknowledgements We would like to thank Rafael D, Heikkila O, Hietala K, Kyllonen L, Kyto J, Rosengard-
Magdalena-Benedito, Joan Vila-Francés, Marcelino Martínez- Barlund M, Saraheimo M, Tolonen N, Parkkonen M, Kaski
Sober, Antonio José Serrano-López, Pablo Escandell-Montero K, Ala-Korpela M, Groop PH (2008) Metabolic phenotypes,
and Ciro Tetta for their precious contribution and support. vascular complications, and premature deaths in a popula-
tion of 4,197 patients with type 1 diabetes. Diabetes 57:2480–
2487
19. Marcelli D, Kirchgessner J, Amato C, Steil H, Mitteregger
References A, Moscardò V, Carioni C, Orlandini G, Gatti E (2001)
EuCliD (European Clinical Database): a database comparing
1. Basara H, Yuan M (2008) Community health assessment us- different realities. J Nephrol 14(Suppl 4):S94–S100
ing self-organizing maps and geographic information systems. 20. Marr B, Schiuma G (2003) Business performance
Int J Health Geogr 7:67. doi:10.1186/1476-072X-7-67 measurement—past, present and future. Manage Decis
2. Bourne M, Neely A, Mills J, Platts K (2003) Implementing 41(8):680–687
performance measurement systems: a literature review. Int J 21. Montefiori M, Resta M (2008) A computational approach for
Bus Perform Manag 5(1):1–24 the health care market. Health Care Manag Sci 12(4):344–
3. Charnes A, Cooper WW, Rhodes E (1978) Measuring the 350
efficiency of decision making unit. Eur J Oper Res 2:429– 22. Nelson DW, Bellander BM, MacCallum RM, Axelsson J,
444 Alm M, Wallin M, Weitzberg E, Rudehill A (2004) Cere-
4. Daugirdas T (1993) Second generation logarithmic estimates bral microdialysis of patients with severe traumatic brain in-
of single-pool variable volume Kt/V: an analysis error. J Am jury exhibits highly individualistic patterns as visualized by
Soc Nephrol 4:1205–1213 cluster analysis with self-organizing maps. Crit Care Med
5. Forgy EW (1965) Cluster analysis of multivariate data: 32(12):2428–2436
efficiency vs interpretability of classifications. Biometrics 23. Pearson K (1901) On lines and planes of closest fit to systems
21:768–769 of points in space. Philos Mag 2(6):559–572
6. de Francisco ALM, Kim J, Anker SD, Belozeroff V, Canaud 24. Resta M (2011) Assessing the efficiency of health care
B, Chazot C, Drüeke TB, Eckardt KU, Floege J, Kronenberg providers: a SOM perspective. In: Laaksonen J, Honkela T
F, Macdougall IC, Marcelli D, Molemans B, Passlick-Deetjen (eds) Advances in self-organizing maps, lecture notes in com-
J, Schernthaner G, Stenvinkel P, Wheeler DC, Fouqueray puter science, vol 6731. Springer, pp 30–39
B, Aljama P (2010) An epidemiological study of hemodial- 25. Roweis ST, Saul LK (2000) Nonlinear dimensionality re-
ysis patients based on the European fresenius medical care duction by locally linear embedding. Science 290:2323–
hemodialysis network: results of the ARO study. Nephron 2326
Clin Pract 118(2):c143–c154 26. Sammon Jr JW (1969) A nonlinear mapping for data struc-
7. Goicoechea M, Caramelo C, Rodriguez P, Verde E, Gruss E, ture analysis. IEEE Trans Comput C-18(5):401–409
Albalate M, Ortiz A, Casado S, Valderrábano F (2001) Role 27. Schieppati A, Remuzzi G (2005) Chronic renal diseases as
of type of vascular access in erythropoietin and intravenous a public health problem: epidemiology, social, and economic
iron requirements in haemodialysis. Nephrol Dial Transplant implications. Kidney Inter Suppl 68(98):S7–S10
16(11):2188–93 28. Seiford LM (1996) Data envelopment analysis: the evolution
8. Gotch FA, Sargent JA (1985) A mechanistic analysis of the of the state of the art (1978–1995). Journal of Productivity
National Cooperative Dialysis Study (NCDS). Kidney Int Analysis 7(2–3):99–137
28(3):526–34 29. Steil H, Amato C, Carioni C, Kirchgessner J, Marcelli
9. Greenwood RN, Ronco C, Gastaldon F, Brendolan A, D, Mitteregger A, Moscardo V, Orlandini G, Gatti E
Homel P, Usvyat L, Bruno L, Carter M, Levin NW (2004) EuCliD—a medical registry. Methods Inf Med 43(1):
(2003) Erythropoeitin dose variation in different facilities in 83–88
different countries and its relationship to drug resistance. 30. Stevenson KB, Hannah EL, Lowder Ca, Adcox MJ,
Kidney Inter, Suppl 64(87):S78–86 Davidson RL, Mallea MC, Narasimhan N, Wagnild JP (2002)
10. Haykin S (1999) Neural networks. Prentice Hall Interna- Epidemiology of hemodialysis vascular access infections from
tional, Inc longitudinal infection surveillance data: predicting the impact
90 I. Cattinelli et al.

of NKF-DOQI clinical practice guidelines for vascular access. 34. Vesanto J, Himberg J, Alhoniemi E, Parhankangas J (2000)
Am J Kidney Dis 39(3):549–555 SOM toolbox for MATLAB 5. Tech. rep., Helsinki Univer-
31. Stopper A, Amato C, Gioberge S, Giordana G, Marcelli D, sity of Technology
Gatti E (2007) Managing complexity at dialysis service cen- 35. Wasse H, Kutner N, Zhang R, Huang Y (2007) Association
ters across Europe. Blood Purif 25(1):77–89 of initial hemodialysis vascular access with patient-reported
32. Tenenbaum JB, Silva VD, Langford JC (2000) A global geo- health status and quality of life. Clin J Am Soc Nephrol
metric framework for nonlinear dimensionality reduction. 2(4):708–714
Science 290:2319–2323 36. Zelman WN, Pink GH, Matthias CB (2003) Use of the
33. Udani S, Lazich I, Bakris GL (2011) Epidemiology of hyper- balanced scorecard in health care. J Health Care Finance
tensive kidney disease. Nat Rev Nephrol 7(1):11–21 29(4):1–16