Extraction of Useful Knowledge for Making Roster by Analyzing

Nurse Scheduling Data and Incident Data

Koichiro Okada

, Masanori Akiyoshi

, Yukie Majima

, Hiroe Takahashi

, Sayuri Tanaka

Misae Tanioka

and Miwako Hori

Department of Computing, Osaka Prefecture University, 1-1 Gakuencho, Naka-ku, 599-8531, Sakai-shi, Osaka, Japan

Department of Computing, Kanagawa University, 3-27-1, Rokkakubashi, Kanagawa-ku, 221-8686,

Yokohama-shi, Kanagawa, Japan

Department of Planning Management, Japan Community Health care Organization (JCHO), Tokyo, Japan

JCHO Osaka Hospital, Osaka, Japan

Keywords: Incident Data, Ladder Level, Nurse Rostering, Nurse Scheduling, Pattern Mining.

Abstract: As described herein, we sought knowledge necessary to make a roster for nurses by analyzing nurse

scheduling data and incident reports on the night shift. Even today, it is difficult to say that computers are

used effectively producing nurse rosters. One reason is that algorithms suggested by researchers are not

practical for nurses working at various sites because they are built without consideration of medical accidents

known as “incidents”. Another reason is that the study of incidents from a team's perspective, which is the

original mode of working as a nurse, is not available. Therefore, this study was conducted for discovery of

knowledge to help produce a nursing roster by analyzing nurse scheduling data and incident data for night

shifts from the viewpoint of teams, which is the original mode of working for nurses.

1 INTRODUCTION

When it is necessary to arrange staff during a day at

institutions such as hospitals, managers usually adopt

a shift-work system to fill a schedule, with shift staff

working days and hours. This system demands a duty

roster to show which duties are covered by which

employee on which time shift. Therefore, this roster

must be produced with consideration of the number

of people needed for duties and various constraints

such as workers’ requests and their respective

capabilities. This scheduling invariably takes much

time. Generally, an experienced expert who is

familiar with the duties and personnel produces this

roster over a long period of time. The problem of

making such a roster entails "the necessity of

assigning work as staff have requested as effectively

as possible." This so-called shift scheduling problem

can be resolved through mathematical programming.

Producing a roster at a medical institution such as a

hospital is a nurse scheduling problem.

Nurse scheduling problems are regarded as a

difficult optimization problem. That is true because

no matter how one might try to resolve the problem,

it is difficult to define an optimum solution for any

reason. If we define such a solution, then it is merely

difficult to obtain the solution. A survey paper

(Smith-Daniels et al., 1988) has presented that

research of nurse scheduling problems began in the

United States in 1976 (Miller et al., 1976). It declined

thereafter, but became popular again from about 1998

(Dowsland, 1998; Miller, Kiragu, 1998). In recent

years, many researchers engaged in the study of this

problem (Burke et al., 2004; Ikegami, 2005; Ikegami,

Tanaka, 2014; Tassopoulos et al., 2015). However, in

spite of the fact that many researchers have developed

the study and have suggested many algorithms for

solving nurse scheduling problems, it remains

difficult to say that computers are used for making

rosters effectively. This fact is apparent from a survey

(Ikegami et al., 1995) that found "the mean time of

making roster is 6.8 hours." Moreover, nurses of a

certain General Hospital, which provided data for our

study, reported long preparation times. Apparently,

hospitals have insufficient cost and technical

knowledge to adopt scheduling software.

Furthermore, the suggested algorithms are not

practical for nurses working at a site because they are

Okada, K., Akiyoshi, M., Majima, Y., Takahashi, H., Tanaka, S., Tanioka, M. and Hori, M.

Extraction of Useful Knowledge for Making Roster by Analyzing Nurse Scheduling Data and Incident Data.

DOI: 10.5220/0005776303830388

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 5: HEALTHINF, pages 383-388

ISBN: 978-989-758-170-0

383

suggested without consideration of medical accidents

known as “incidents.”

Table 1: Number of nurses per level.

Ladder level Number of nurses

1 26

2 16

3 19

4 5

5 6

Temporary (0) 18

Studies of incidents (Bates et al., 1995; Onozawa

et al., 2000) have been conducted by many

researchers along with the nurse scheduling problem.

Most depend on examinations of individual nurses to

examine "What kind of nurses and circumstances

cause incidents and what should we do for preventing

those incidents." Research about incidents from a

team perspective, the original mode of working for

nurses, has not been conducted to date.

Therefore, this study was undertaken for

discovery of knowledge to help produce a nursing

roster by analyzing nurse scheduling data of 2009–

2013 (5 years) and incident reports related to night-

work provided by A General Hospital from the

viewpoint of teams.

2 EXPERIMENTAL DATA

DETAILS

Data we use are nurse scheduling data covering five

years (April 1, 2009 – March 31, 2013 ) and incident

date data on night shifts from the same period of time

provided by A General Hospital (number of hospital

beds, 68; mean number of hospital patients per day,

62; number of nurses, 89). The number of nurses

includes temporary employees. Changing of nurses

occurred during those five years.

Through research collaboration with A General

Hospital, we received consent to use these data.

2.1 Details of Nurse Scheduling Data

These nurse scheduling data (e.g. see Appendix A)

comprise seven items.

 Nurse ID

 Ladder level (Skill level)

 Date

 Day of the week

 Types of shift work

 Shift total of the month for each nurse (e.g. nurse

A [day shift, 10; night shift, 5])

 Total number of nurses with each shift (e.g. Day

1 [day shift, 12; night shift, 6])

Nurse ID consists of five digit numbers. The ladder

level has five stages of 1–5 (temporary nurse is 0,

number of nurses per level is shown in Table 1) and

Types of shift work consist of 11 items.

 Day shift

 Late shift

 Night shift

 Training

 Business trip

 Recreation

 Public holiday

 Holiday

 National holiday

 Substitute holiday

 Absence

2.2 Details of Incident Date Data

These incident date data on the night shift (e.g. see

Appendix B) consist of two items as follows.

 Types of incidents

 Time and date of occurrence of each incident

Types of incidents consist of three items.

 Care: About fall occurrence

 Pharmacy: About wrong medication

 Tube: About removal of tube

This study was conducted for discovery of the

knowledge to support the production of a nurse roster

by analyzing these data from a team viewpoint.

3 EXPERIMENT

For this study, we conducted three experiments.

 Extraction of a shift pattern that all nurses meet

working throughout the year

 Correlation of incidents and nurse group in the

night shift

 Extraction of nurse pairs with no mutual work on

a day or night shift during five years

Table 2: Time series database (Cited by Zaki, 2001).

SID EID Items

1 10 CD

1 15 ABC

1 20 ABF

1 25 ACDF

HEALTHINF 2016 - 9th International Conference on Health Informatics

384

Table 2: Time series database (Cited by Zaki, 2001) (cont.).

2 15 ABF

2 20 E

3 10 ABF

4 10 DGH

4 20 BF

4 25 AGH

Table 3: Part of list coupling SID and TID per item.

A B

…

SID EID SID EID

…

1 15 1 15

…

1 20 1 20

…

1 25 2 15

…

2 15 3 10

…

2 15 4 20

…

3 10

…

4 25

…

Table 4: Lis obtained by combination of item A and B in

Table 3.

A-B

SID EID

1 20

3.1 Extraction of Shift Patterns with

All Nurses Working throughout the

Year

3.1.1 Overview

Nurse replacement occurred in each year, but we

specifically examined nurses who worked throughout

the year and extracted shift patterns such that all

nurses meet using a sequential pattern mining

algorithm called SPADE (Sequential Pattern

Discovery using Equivalence classes) algorithm

(Zaki, 2001). Constraint conditions for making shifts

are likely to be hidden in such shift patterns.

3.1.2 Spade Algorithm

Sequential pattern mining is extraction of pattern

which have ordering. SPADE algorithm is the way of

this mining. SPADE receives a time series database

such as Table 2 as input, and this database consists of

time series data. Time series data has Sequential ID

(SID), Time ID (TID) and Items. SPADE combine

SID and TID into list per item like Table 3 and find a

new time series pattern by coupling these lists like

Table 4.

3.2 Correlation of Incident and Nurse

Groups on the Night Shift

3.2.1 Overview

On the night shift each day, we sorted nurse groups

on shifts according to whether that group is likely to

cause an incident or not based on the following

definition of the group. Then we assess the correlation

of the incident and that group.

Definition: when group work at night shift as

hellows, we regard its group as “Risky group” that is

likely to cause incident and other as “Non-risky

group”.

Group have over 2 nurses whose ladder level is

under 3 and have no nurse whose level is over 3

3.2.2 Procedure

We make an experiment by following the steps

described below.

1. We pick out all nurses as each night shift from

scheduling data. Every nurse work at night shift

are one group.

2. We distinguish these groups “Risky group” form

“Non-risky group” using above definition.

3. We calculate the rate of each incident occurrence

in the case of “Risky group” and “Non-risky

group”.

3.3 Extraction of Nurse Pairs with No

Work Together on Day or Night

Shifts during Five Years

3.3.1 Overview

Nurse work is team work related directly with the life

and death of patients. Therefore, human relations

must be considered, particularly excluding pairs that

are more likely to cause incidents. In addition, human

relations can cause resignation of freshly recruited

nurses (Mizuta et al., 2004). Therefore, it is important

to formalize such pairs without leaving it as tacit

knowledge.

Table 5: Extracted shift pattern that has maximum length in

each year.

Year Extracted shift pattern that has maximum length

2009 <N,N,PH,D,D,N,N,PH,D>

〃

<N,N,PH,PH,D,D,N,N,PH>

〃

<N,N,PH,D,D,D,N,N,PH>

〃

<N,N,PH,D,D,N,N,PH,PH>

Extraction of Useful Knowledge for Making Roster by Analyzing Nurse Scheduling Data and Incident Data

385

Table 5: Extracted shift pattern that has maximum length in

each year (cont.).

Year Extracted shift pattern that has maximum length

2010 <D,D,N,N,PH,D,D>

〃

<N,N,PH,PH,D,D,D>

2011 <D,N,N,PH,D,D,N,N>

2012 <D,D,N,N,PH,PH,D,D>

2013 <D,D,N,N,PH,PH,D,D>

〃

<D,N,N,PH,PH,D,D,D>

〃

<PH,PH,D,D,D,N,N,PH>

〃

<PH,D,D,D,N,N,PH,PH>

*D: Day Shift, N: Night Shift, PH: Public Holiday

Table 6: Combination of shift patterns of length 3.

<D,D,D> <D,D,N> <D,N,N>

<N,N,PH> <N,PH,D> <N,PH,PH>

<PH,D,D> <PH,PH,D>

*D: Day Shift, N: Night Shift, PH: Public Holiday

3.3.2 Procedure

We make an experiment by following the steps

described below.

1. We acquire all nurse IDs from scheduling data and

make pairs of IDs.

2. We make pairs of IDs at day shift and delete them

form pairs created procedure 1. Similarly, we do

at night shift.

3. After procedure 2, leftover pairs are regarded as

nurse pairs with no work together on day or night

shifts during five years.

4 RESULTS AND DISCUSSION

4.1 Extraction of Shift Patterns with

All Nurses Working throughout a

Year

Table 5 shows the extracted shift pattern that has

maximum length in each year. The maximum lengths

of shift patterns were high values such as 7–9, but a

nurse actually making a roster considered these

longest shift patterns only slightly when producing

the roster. In other words, roster makers consider

shorter shift patterns. Longer shift patterns result

from combined constraints such as "Two consecutive

night shifts are forbidden." Actually, these longer

patterns are explainable by the combination of shift

patterns of length 3, as shown in Table 6. In both

Table 5 and 6, “D”, “N” and “PH” indicate “Day

shift”, “Night shift” and “PH”, respectively. However,

“Night shift” is a set of two “N” like <N,N>. Since,

first “N” indicate “begin working” and second “N”

indicate “finish working”.

4.1 Correlation of Incidents and Nurse

Groups on the Night Shift

Tables 7 and 8 respectively show the rates of incident

occurrence when nurses work in a "Risky group" or

"Non-risky group". Figures 1, 2 and 3 show the rates

of incident occurrence per incident type. According

to Tables 7, 8 and Figures 1, 2 and 3, the rate as a

"Risky group" was slightly higher than that of a "Non-

risky group". Therefore, we investigated them using

a t-test to clarify whether these results had significant

differences. Results show no significant difference in

any of the three kinds of incident. However, these

results are dubious because we only use data for last

5 years. Accordingly, we are set to prepare data for

more than 10 years and do significance test using that

data.

Table 7: Rates of incident occurrence when nurses work in

a "Risky group" (%).

Year Care Pharmacy Tube

2009 21.2 13.6 3.4

2010 18.5 9.2 4.6

2011 19.7 10.6 2.3

2012 23.6 10.8 8.1

2013 13.6 12.6 5.5

Five years 22.2 13.1 5.8

Table 8: Rates of incident occurrence when nurses work in

a "Non-risky group" (%).

Year Care Pharmacy Tube

2009 19.4 10.1 1.6

2010 18.3 7.7 2.6

2011 20.6 9.9 3.0

2012 24.7 6.4 7.8

2013 18.1 18.1 7.8

Five years 20.3 10.0 4.3

Figure 1: Rate of “Care” incident occurrence.

HEALTHINF 2016 - 9th International Conference on Health Informatics

386

Figure 2: Rate of “Pharmacy” incident occurrence.

Figure 3: Rate of “Tube” incident occurrence.

Table 9: Nurse pairs who did not work together on day and

night shifts during five years.

Type of

shift

Number

of pairs

Number of pairs between the nurses

with ladder level lower than 2

day 117 58

night 34 26

both 24 18

4.2 Extraction of Nurse Pairs Who Did

Not Work Together on Day and

Night Shifts during Five Years

Table 9 shows the results. We gave a breakdown of

the pair between the nurses with ladder level lower

than 2 because the extracted pairs had high

probabilities of causing incidents. We made a roster

devoting attention to these pairs.

5 CONCLUSIONS

As described herein, we sought knowledge to support

production of a nurse roster by analyzing nurse

scheduling data and incident reports from night-work

provided by A General Hospital from a team

viewpoint, reflecting a nurse working style.

Accordingly, we conducted three experiments as

follows from the viewpoint of a team.

 Extraction of shift patterns that all nurses meet

working throughout a year

 Correlation of incidents and nurse groups on the

night shift

 Extraction of nurse pairs with no mutual work on

day and night shifts during five years

In shift pattern extraction, a limitation condition in

roster making and a possible shift pattern were

extracted. The maximum length in each year was

explained using a combination of shift patterns of

length 3. The real incident rate was judged from

viewing 2 in a case of working in the dangerous nurse

group, which was more likely to produce incidents.

Although the rate of the dangerous group was slightly

high, no significant difference was found. Pairs who

did not take day and night duty together during five

years were extracted.

Future studies should examine techniques to

identify shift patterns that are most appropriate as

constraints to roster making by analyzing extracted

shift patterns and correlations with incidents.

REFERENCES

Bates, D. W., Cullen, D. J., Laird, N., Petersen, L. A.,

Small, S. D., Servi, D., Laffel, G., Sweitzer, B. J., Shea,

B. F., Hallisey, R., Vliet, M. V., Nemeskal, R. & Leape,

L. L. 1995. Incidence of adverse drug events and

potential adverse drug events: Implications for

prevention. The Journal of the American Medical

Association, 274, 29-34.

Burke, E. K., Causmaecker, P. D., Berghe, G. V. &

Landeghem, H. V. 2004. The state of the art of nurse

rostering. Journal of Scheduling, 7, 441-499.

Dowsland, K. A. 1998. Nurse scheduling with tabu search

and strategic oscillation. European Journal of

Operational Research, 106, 393-407.

Ikegami, A. 2005 Nurse Scheduling - Site research,

Modeling and Algorithms -. Proceedings of the

Institute of Statistical Mathematics, 53, 231-259.

Ikegami, A., Aizawa, M., Ohkura, M., Wakasa, K.,

Matsudaira, N. & Kosugo, R. 1995. A preliminary study

of the development of a scheduling system for hospital

nurses. Journal of Science of Labour, 71, 413-423.

Ikegami, A., Tanaka, Y. 2014. Re-challenge to the nurse

scheduling problem (in Japanese). The operations

research as a management science, 59, 26-33.

Millar, H. H., Kiragu, M. 1998. Cyclic and noncyclic

scheduling of 12 h shift nurses by network

programming. European Journal of Operational

Extraction of Useful Knowledge for Making Roster by Analyzing Nurse Scheduling Data and Incident Data

387

Research, 104, 582-592.

Miller, H. E., Pierskalla, W. P. & Rath, G. J. 1976. Nurse

scheduling using mathematical programming.

Operations Research, 24, 857-870.

Mizuta, M., Kousaka R., Tsuji, S., Nakano, M. & Inoue, J.

2004. The degree of psychological distress and turnover

wish among new graduate nurses. The bulletin of

Nursing College, Wakayama Medical University, 7, 21-

27.

Onozawa, Y., Yoshioka, N., Kanbayashi, M., Tosaka, K.,

Tamura, T. & Kawaguti H. 2000. A study of incident

on clinical nursing: incident mode and nurses

perceptions. The bulletin of Niigata College of Nursing,

6, 71-90.

Smith-Daniels, V. L., Schweikhart, S. B. & Smith-Daniels,

D. E. 1988. Capacity management in health care

services: Review and future research directions.

Decision Sciences, 19, 889-919.

Tassopoulos, I. X., Solos, I. P. & Beligiannis, G. N. 2015.

A two-phase adaptive variable neighborhood approach

for nurse rostering. Computers and Operations

Research, 60, September 2015, 150-169.

Warner, D. M. 1976. Scheduling nursing personnel

according to nursing preference: A mathematical

programming approach. Operations Research, 24, 842-

856.

Zaki, M. J. 2001. SPADE: an efficient algorithm for

mining frequent sequences. Machine Learning, 42, 31-

60.

APPENDIX

Appendix A: Sample of nurse scheduling data.

2009/04 1 2 3 4 … 30

Shift

total

NurseID -

Ladder

level

Wed. Thurs. Fri. Sat. … Thurs. Day Night … Absence

11111 - 2 PH N N PH … N 10 7 … 0

11122 - 1 N PH D D … PH 12 4 … 0

…

44111 - 0 D D PH N N PH 7 7 … 0

Shift

total

Day 17 12 15 16 15

Night 7 7 8 6 7

…

Absence 0 0 0 0 0

※ We define Temporary nurse’s ladder level is 0

Appendix B: Sample of incident date data.

2009 Types of incident Date

Care 04/01 04/03 04/10 04/21 05/09 05/13 … 03/12

Pharmacy 05/12 05/27 06/22 06/28 07/05 07/18 … 02/22

Tube 04/21 07/22 08/14 09/04 09/25 10/11 … 02/02

HEALTHINF 2016 - 9th International Conference on Health Informatics

388