![](bg2.png)
Section 3 we introduce the domain, the data set, and
the data cleaning process. In Section 4, we present a
descriptive analysis of the data set using process min-
ing tools and standard statistical tools to identify in-
formative features of the data. This informs the pro-
cess of building predictive models which we describe
and evaluate in Section 5. In Section 6 we discuss our
results and in Section 7 we conclude.
2 RELATED WORK
Improving efficiency in surgical wards, specifically
improving utilization of operating rooms, has re-
ceived growing interested nationally and internation-
ally for a number of years now (Lalys and Jannin,
2014). The National Theatres Project in Scotland
states as its objective, “appropriately increasing pa-
tient throughput, thereby using resources more pro-
ductively and efficiently”(Scotland, 2006). The met-
rics for improvement include: reducing unutilized
(operating room) hours; reducing over/under-runs,
late-starts, cancellations and delayed discharges; and
avoiding unnecessary out-of-hours and nighttime pro-
cedures. Many of these objectives are strongly related
to appropriate scheduling, and would thereby benefit
from more accurate, data-informed, models of patient
flows.
A significant amount of research exists in model-
ing processes in the surgical domain. The modeling
scope of much existing work tends to fall on two ends
of a spectrum in terms of granularity: the level of sur-
gical procedures at one end and broader care flows
beyond the surgical ward at the other.
In (Lalys and Jannin, 2014), 46 publications on
surgical process modeling are categorized into a tax-
onomy ranging from the level of the surgical pro-
cedure at the lowest level of granularity, to low-
level physical movements at the highest. At the lat-
ter level, which is typically concerned with robot-
assisted surgery or training and assessment of sur-
geons, we see research on phase detection (Stauder,
2014) and detailed models of individual tool usage
patterns based on sensor data (Ahmadi, 2009). In-
dividual hand motions from video data are automat-
ically identified in (Lin, 2006) and (Haro, 2012). A
number of models based on sensor data collected dur-
ing Cholecystectomies (a highly standardized proce-
dure), were developed in (Blum, 2008), (Bouarfa and
Dankelman, 2012), (Bouarfa, 2011), and (Neumuth,
2011). All of these studies have the surgical proce-
dure at the highest level of abstraction. Our present
investigation lies above this level of granularity, with
only the procedure name and some other basic details
being present in the data.
Above the level of individual procedures, we see
work such as (Stahl, 2006) which describes the work-
flow within an operating room, including anesthesia,
surgery, and early recovery. Other studies also ad-
dress the process surrounding surgery, from admis-
sion to recovery (Funkner, 2017), which matches the
scope of our data set. Taking a view beyond the oper-
ating room is important, since activities downstream
from the actual surgical procedure can interrupt pa-
tient flows as shown in the case of ICU bottlenecks in
(Akkerman and Knip, 2004). Some studies have also
incorporated diagnosis and follow-up after surgery
such as (Mans, 2012) and (Huang, 2013).
Bayesian networks were used to model several as-
pects of stays in an emergency department in (Acid,
2004). While overall stay duration was one attribute
included in the model, the scope was at higher level
of abstraction, and not focused specifically on surgi-
cal patient flows. Furthermore, the main focus was
the comparison of structure learning algorithms.
Some work has looked specifically at model-
ing variance in surgery durations (Strum, 2000) and
incorporating this into sequencing and scheduling
strategies (Denton, 2007). In (Kayis, 2012), regres-
sion modeling is employed to predict surgery dura-
tion based on clinical, operational and temporal data.
Stochastic balancing of bed capacity based on fluctu-
ating demand patterns was explored in (Cochran and
Bharti, 2006) and length of stay patterns in (Akker-
man and Knip, 2004). Resource allocation and patient
admission was addressed in (Hulshof, 2013).
In summary, the scope of patient flows ranging
from admission, through surgery to recovery, is one
which has been less thoroughly addressed: most work
is positioned at a lower or higher level of abstrac-
tion. In regards to the distribution of surgery times,
our work has the corollary contribution of confirm-
ing previous findings. In terms of the more nuanced
conditional models we present of cycle times, specif-
ically the integration of patient clusters to Bayesian
networks, we believe our approach to be novel.
3 DOMAIN & DATA
PREPARATION
The Royal Infirmary of Edinburgh is the largest in
Scotland, housing 900 beds and with its 24-hour ac-
cident and emergency department, providing a full
range of acute medical and surgical services. The hos-
pital IT system is integrated with the Operating Room
Scheduling Office System (ORSOS), a surgery man-
agement and scheduling system.
HEALTHINF 2020 - 13th International Conference on Health Informatics
274