Professional Documents
Culture Documents
Scheduling Systems
Hospital Scheduling
Hospital patients are usually classified as elective or emergency; emergency
patients, however, simply represent a limiting case of electives with no scheduling flexibility. Elective patients usually enter the scheduling system as the
result of a telephone call from the diagnosing physician to the hospital
admissions office. At this point the potential patient has four attributes of
importance to scheduling: (1) desired admission day; (2) flexibility in possible admission days; (3) potential length of stay in the hospital; and (4)
hospital service demands (medical or surgical patient, room type, operating
room needs, special care or service needs). Information on items (1), (2),
and (3) must be available if a scheduling system is to function; and schedSupported by U.S. Public Health Service Grant No. HM-00224, Division of Hospital
and Medical Facilities. This research was performed while the authors were with the Industrial Engineering Department, University of California, Berkeley.
130
PATIENT SCHEDULING
uling cannot be effected if item (2) does not exist, as is essentially the case
for the emergency patient. Demands under item (4) appear to add little to
the conceptual complexity of scheduling models, although they could detract
seriously from their effectiveness.
No information is available to the hospital on a potential patient prior to
the physician's request for admission. Before this point the hospital scheduling
system can obtain and use general predictive information describing such
phenomena as growth trends, seasonal fluctuations, and weekend differences.
These data can be used to perform appropriate parameter adjustments on the
various scheduling models. Our concern here is with the scheduling of individual patients.
OnIe important reason why quantitative scheduling methods are not employed in hospitals is a lack of data on the patient attributes defined above.
This lack applies both to the potential patient requesting admission and,
perhaps surprisingly, to the patient who is actually in the hospital. Prior to his
actual admission the patient must be described by all four attributes identified
above. The description of items under (4), service demands, is relatively
straightforward. Information on desired admission day is generated by the
physician in accordance with his and his patient's needs. Information on
possible flexibility of this day, however, is generated only implicitly by
negotiations between the admissions office and the physician. No normative
medical information is available indicating the explicit costs of deferring
hospital admission for specific patient disease states. A number of other factors
such as pain, discomfort, and inconvenience enter this determination, and a
quantitative consideration of these would appear to be beyond any presently
envisaged system. The quantification of this variable will therefore probably
remain a dual effort of the physician and the admissions office. It is perhaps
not unreasonable to assume, however, that this "game" will be played fairly
and well by both parties if it is clear to both that effective scheduling will
benefit the total hospital-medical system; that is, the game is not zero sum.
The simulation models discussed here use a highly simplified scheme for
representing the desired admission day and its flexibility. It is believed that
this scheme represents the quantitative nature of the results of the physicianadmissions office negotiation and, furthermore, may present a framework within
which the physician and the admissions office can carry out this negotiation.
Substantial progress has been made in uncovering ways to produce data on
item (3), the patient's estimated length of hospital stay. It is immediately
evident that whatever the status of all other information or whatever the
scheduling policy devised, information must be available on patient length of
stay, and it must be available at the time of scheduling. Recent studies [1-3]
have shown the feasibility of using the admitting physician as an estimator,
either directly or indirectly; his estimate is apparently at least as good as that
available from statistical techniques using diagnostic data, and the data
production cost is substantially less.
Descriptive data on the statistical nature of admission requests, time
Summer 1968
131
Robinson
et
al.
between request and actual admission, and length of stay were gathered at
two general hospitals in California: Alta Bates Community Hospital, Berkeley
(152 beds), and Mount Zion Hospital and Medical Center, San Francisco
Several simplifying assumptions were made in order to develop an operational simulation. These arose primarily as the result of incomplete data on
certain aspects of the hospitals or as programming expedients that were
judged not to have a large effect on the simulation results. They were:
1. Patients: Only elective patients are processed by the scheduler. Emergency
patients (including maternity) are admitted according to hospital policy
outside the present scheduling considerations. One reasonable policy is
simply to reserve a fixed block of beds for emergency admissions, the size
of the block being a function of the emergency history and the hospital's
policies toward emergencies.
2. Beds: The hospital has a fixed number of identical beds, available to any
patient regardless of his description. Problems of adjusting capacity to
meet varying demand and shifting patients from one bed to another are
therefore ignored. This assumption can be relaxed by classifying the beds
according to accommodation type (private, semiprivate, ward) and tagging
patients with desired accommodations and personal characteristics (sex,
medical isolation needs, and the like).
3. Hospital operations:
a. The number of available, or potentially available, beds is the only
scheduling criterion. This assumption can be interpreted as implying
either that all other facilities are adequate or that consideration of them
is incorporated in the patient demand description. (For example,
restrictions on the availability of the operating room could be handled
by modifying the patient's desired admission day and its flexibility).
b. Patients actually enter the hospital once they have been scheduled. No
data were available on the number of "no shows" in the hospital system.
If this number is significant, an additional routine can be added to
withdraw scheduled patients.
c. Discharge of any specific patient is certain on the morning of his actual
discharge. This is close to observed practice, in which the discharge
decision is usually made at the attending physician's morning visit, and
the patient actually leaves the hospital before a new patient arrives.
d. Requests for admission are processed after the day's discharges have
taken place and the state of the hospital has been subsequently updated.
e. Priorities arising from social or professional status are not considered.
This kind of priority can be considered as embedded in the dimensions
of desired admission day and scheduling flexibility.
132
Services Research
Health Services
Health
Research
PATIENT SCHEDULING
Summer 1968
133
Robinson et al.
tion with mean equal to ALOS and variance proportional to ALOS. Unbiasedness is not critical, since the proper operating point of the hospital can be
determined regardless of bias.
3. Earliest possible arrival date (EPA) and latest possible arrival date
(LPA), to establish bounds on admission dates for patients. No data exist on
these values, since they are generally only implicit in the scheduling process.
The present program selects one EPA-LPA pair for each patient as follows:
EPA is drawn from a geometric distribution approximating that portion of the
Alta Bates Hospital distribution of admission intervals (time between request
and actual admission) in excess of five days [4]. LPA is then selected with
mean and variance increasing with EPA. For an EPA of 1 or 2, LPA is approximately 1 or 2, with a probability of about 0.1 of being as high as 4 or 5. For
an EPA of 11-14, LPA is approximately 6, with a 0.1 probability of being as
high as 9-12. Although this is a fairly arbitrary procedure, the resulting
distribution of admission intervals gathered from the simulation compares
favorably with the descriptive data.
PATIENT SCHEULING
Summer 1968
135
Robinson et al.
of the different scheduling rules. A separate program was used for this function for economic reasons. The procedure consists in completing a series of
four or five runs at internal operating levels varying by five beds; interpolating
between these points to obtain a complete set of data; and then applying
relative costs to empty beds, to hospital overflow, and to the turning away of
prospective patients, in order to determine the "best" scheduling rule and its
associated "optimal" operating level. This procedure offers substantial savings
in computer time over the alternative of incorporating costs into the Scheduler
program, which would require a set of runs for each set of costs under consideration. Since the rules under investigation do not functionally depend on the
relative costs, there is no loss of accuracy with this approach.
As an approximation of the cost structure relevant to hospital scheduling,
the average daily cost of operation is computed as follows:
Mean daily cost of operation = cost of empty bed X mean number of empty beds
+ cost of overflow X mean number of overflows
+ cost of tumaway X mean number turned away
where the mean numbers of empty beds, overflows, and turnaways are outputs
of the Scheduler program. Turnaway costs have been included as indicative
of various costs and relationships that can, at this time, be discussed only in
subjective terms. Administrators who adopt quantitative methods in their
operations are continually called upon to define equally vague quantities as
their systems become more sophisticated. Attempts to grapple with these
problems often lead to a better understanding of the operation of the firm:
one of the hidden returns of quantitative methods. Extensions of the above
cost model rapidly lead to other subjective areas.
136
PATIENT SCHEDULING
Three provisions have been made to increase the validity of output data
and the statistical significance of comparisons between runs. Most important
of these is the large sample size. Each run of the simulation is 200 "days,"
processing approximately 3200 patients in a 100-bed hospital. A major source
of variation between runs is eliminated by using the same patient input data
for all runs, thus assuring that the scheduling rule is the only source of
variation. The hospital also is started at an initial condition with patients
in bed and on a scheduled admissions list and is run for 25 days before data
collection is started. Measurements indicate that any transients due to the
initial conditions have died out in 25 days. If the hospital is started empty,
as many as 50 days may be needed for stable operation.
Results
Six scheduling systems were investigated for one set of cost ratios: 1:10:1,
empty beds/overflows/turnaways. These scheduling systems were the three
discussed earlier (filled page, ELOS, and ELOS-PT), with four degrees of
estimate precision within the ELOS system ("poor," "normal," normal with
revision after hospitalization, and perfect).
With perfect estimation, ELOS = ALOS for all estimates. The results of
this system indicate the range of improvement possible with improvements in
estimating techniques. The costs measured in the system reflect those resulting
from patient-request variability.
The "poor" estimation system assigned ELOS values to patients according
to the following scheme:
ELOS, in days
ALOS, in days
2
1-3
6
4-8
10
>8
Length of stay is thus classified in three categories, corresponding roughly to
short, medium, and long stay in general hospitals. This represents possibly the
least accuracy that would be expected by any reasonable estimation technique.
The "normal" estimation system was designed to represent the results of
studies on the physician as an estimator of length of stay at the time an admission is requested [1,2]. The ELOS is drawn from a Poisson distribution with
mean equal to the ALOS and variance equal to 0.55 ALOS.
The normal estimation with revision selects a new value of ELOS after
three days of actual hospitalization. The variance of the new estimate is 0.55
times the remaining portion of stay (ALOS-3). Data on the third day of
hospitalization indicate this to be a conservative assumption. There is good
reason to believe that a carefully constructed revision system can substantially
reduce the ELOS variance [2,3].
The ELOS-PT scheduling system uses the normal estimate variance for
ELOS/ALOS and produces expected census values based on the probabilities
Summer 1968
137
Robinson et al.
of stay for each day, rather than simply assuming the precise termination of
stay at ELOS, as the ELOS systems do.
The results of this study are shown in the accompanying table. The optimal
operating level is simply the parameter value in the model yielding the lowest
cost figure. It is the size of the "page" for the filled-page technique, and it
represents the size of the hospital for the ELOS and ELOS-PT systems. The
actual size of the hospital is 100 beds.
Operating Characteristics for Several Scheduling Rules for a 100-bed "Hospital"
Basic
scheduling
rule
Filled page
ELOS "poor"
ELOS "nornal"
ELOS normal
with revisions
Variance
ELOS for
EachSALOr
-
0.55 x ALOS
0.55 X ALOS
and 0.55 X
(ALOS-3)
0.0
Census
Mean daily
Optimal
no. operating
operating mean standard Mean
level
mendeviation overflows
cost
16
82
95
95.0*
94.1
95.3
95
95.5
ELOS perfect
100
98.9
ELOS-PT
normal
0.55 x ALOS
105
93.4
*The probable error of these measures is somewhat less
All differences shown are, with relatively high probability,
4.8*
5.5
4.0
0.3*
0.3
0.3
12.5*
12.8
11.5
3.4
0.2
0.0
10.3
4.3
2.3
4.2
0.1
11.5
than the least significant figure.
indicative of true values.
PATIENT SCHEDULING
The Scheduler program used in the simulation could form the core of a
real-time automated patient-scheduling system. The economic feasibility of
such a design would depend, of course, on a number of factors specific to the
particular hospital, including the availability and cost of computer time. The
gain suggested by sophisticated scheduling techniques (see table) seems to
indicate feasibility, particularly if the data necessary to operate the system
are produced at essentially zero cost and if the present manual recordkeeping and scheduling function can be eliminated.
The data input to the system (EPA, LPA, ELOS, and special room and
service needs) would probably be generated by the admissions office and the
physician in a manner similar to present practice. With computer scheduling,
however, specific values for these quantities would have to be produced, as
contrasted with their implicit state in most current systems. Special EPA-LPA
needs could be handled by successive scheduling attempts; for example, such
a request as "Monday or Tuesday or the following Monday or Tuesday" could
be handled by two successive EPA-LPA inputs, if necessary. If the hospital
noted any patterns of requests not easily handled by the Scheduler program,
they could be incorporated in new subroutines.
It would probably be necessary to allow both the physician and the admis-
Summer 1968
139
Robinson et al.
REFERENCES
1. Robinson, G. H., L. E. Davis, and G. C. Johnson. The physician as an estimator of
hospital stay. Human Factors 8:201, 1966.
2. Robinson, G. H., L. E. Davis, and R. P. Leifer. Prediction of hospital length of stay.
Health Serv. Res. 1:287 Winter 1966.
3. Gustafson, D. H. Length of stay: Prediction and explanation. Health Serv. Res. 3:12
Spring 1968.
1For example, the system could be programmed to determine which wards should be
reserved for a particular sex or disease and to change these assignments if better utilization would result.
140
PATIENT SCHEDULING
4. Robinson, G. H. Statistics on Hospital Census, Admission and Discharge Rates, and Admission Intervals. Report No. HFT-64-4, Human Factors in Technology Research
Group, University of California, Berkeley, 1964.
5. Wing, P. and G. H. Robinson. Computer Programs for Simulating Hospital Scheduling
Systems. Report No. HFT-66-1, Human Factors in Technology Research Group, University of Califomia, Berkeley, 1966.
6. Markowitz, H. M., B. Hansner, and H. W. Carr. SIMSCRIPT, A Simulation Programming Language, Prentice-Hall, 1963.
7. Robinson, G. H. Hospital Admission Scheduling Control. Report No. HFT-65-4,
Human Factors in Technology Research Group, University of California, Berkeley, 1965.
8. Wing, P. Automated system for scheduling admissions, Hosp. Mgt., 104:53 October
1967.
Summer 1968
141