Stochastics and StatisticsClassification trees: A possible method for maternity risk grouping
Introduction
Childbirth is a natural, fundamental process for human survival. Like any other natural process, it is governed by social, biological and environmental factors that can have adverse effects on the health of mothers and their babies. Over recent years, there has been considerable debate concerning the provision of maternity care in the UK. The government document Changing Childbirth [1] outlines new principles for good maternity care, suggesting that women should be the focus of the process and should be fully informed about all the possible options of care during pregnancy and at delivery.
Pregnancy still remains subject to numerous complications and risks. Although many procedures, such as caesarean section, are increasingly becoming more routine, there still remains an element of risk to both mother and child. It would be impossible to totally eliminate all risks and so different women receive different types of care based on the clinician’s assessment of the pregnancy. There is an obvious benefit to both midwife/obstetrician and expectant mother if an early warning system could flag up potential risks to individual women. Such risks might include a preterm (premature) delivery, the need for a caesarean section and a non-spontaneous delivery.1 Such an early warning system would allow for more suitable types of care to be offered given the likely risks involved with a particular pregnancy.
Classification and Regression Tree (CART) analysis [2] may offer help in producing easy to use statistically and clinically meaningful maternity risk groups. CART investigates the effects of selected independent variables on a dependent variable. Starting with all observations in a single group and a set of independent variables, the observations are split into two groups on the basis of the independent variable that is judged to be the most important in reducing the total variation or deviance in the continuous or categorical (respectfully) dependent variable. The two resulting groups are split further by the same or an alternative independent variable that continues to reduce the total variation within the dependent variable. The partitioning is repeated until further splitting will not reduce the variation of the dependent variable.
The partitioning may be unrestricted, i.e. the total variation within the dependent variable is minimised by mathematical means alone, or fixed by the investigator, i.e. the effect of a selected independent variable on splitting is explored. As a result, either the statistical method or clinical judgement, or a combination of both, can be used. A statistical approach helps to minimise subjective clinical bias about which patients ought to be grouped together and the classification rules used to group the patients.
CART has been successfully used in other areas of healthcare such as case mix groups [3], hospital bed capacities [4], cancer survival groups [5] and intensive care [6], but has yet to be used to create risk groups for maternity care.
This paper describes a study of 3902 data records from an obstetric database. Consultant obstetricians and midwives from the Princess Anne Hospital, part of the Southampton University Hospitals NHS Trust, provided clinical guidance during the study. The study evaluated the statistical effectiveness and user-interpretability of the CART algorithm for the creation of risk groups for a number of childbirth complications. We found that in many cases, the resulting trees reinforce clinical judgement, highlight the heterogeneous nature of the observations, and help to quantify the associated risks through the construction of homogeneous patient groupings.
Section snippets
Childbirth risks
Medical professionals routinely assess the risks of their patients, often by relying on their clinical judgement and experience to decide on the best course of care to set for each expectant mother. Risk scores have evolved as an aid to this process, but by no means act as a replacement. More recently, and with the advent of improved obstetric databases and data collection, many different risk-scoring systems have been proposed, with varying levels of success [7], [8]. All proposed systems
Methods
The CART algorithm generates groups of individuals based on a selected criterion for splitting a group [2]. The method requires certain inputs by the user, which include the maximum number of final groups, the minimum number of individuals within each final group and an acceptable reduction in variance used to determine further splitting into new groups. The groups are called nodes and form a branching node tree. The investigator can determine the partitioning of the data into nodes so that the
Results
A preliminary analysis of the data was performed. A summary of interesting results is provided below:
- •
The average age of the women was 28.4 years, ranging from 14 to 44.
- •
Body mass index (ratio of weight to height) averaged 24.8 with 90% of the observations between 19.07 and 33.9.
- •
A mean gestation period of 39.3 weeks was observed. 11.2% of pregnancies have a gestation period of less than 37 weeks, classed as preterm deliveries.
- •
The birth weight of the baby is approximately normally distributed with
Defining risk groups
From the regression tree analyses, as discussed in the previous section, it is possible to define risk groups which will be of use to medical professionals. A woman is to be given a classification for each specific outcome; for example she will be given a classification defining her chances of a preterm delivery and a risk assigned to chance of a caesarean section. This will give the obstetricians and midwives quantitative information on what specific complications they should be aware of as
Development of a maternity database tool with an early warning system
In a move to update the maternity database at the Princess Anne Hospital, a new multi-functional database has been developed to store the necessary information and to produce required reports. Its primary use is for the storage and maintenance of the personal, antenatal and medical/obstetric history information of each pregnant woman. In addition to this, the database incorporates an early warning system for possible complications to each pregnant woman. The database, on entering data for a new
Discussion
A classification tree algorithm (CART) has been used to analyse a dataset of 3902 births consisting of 34 recorded variables. The results of the presented study suggest that CART analysis is a useful tool in helping to produce statistically and clinically meaningful maternity risk groups for a number of outcomes of interest. Risk groupings have been constructed for a non-spontaneous (complicated) birth, a caesarean section and a preterm birth.
Use of the CART algorithm with staff from the
Acknowledgements
The authors express gratitude to staff at the Princess Anne Hospital for their medical guidance and enthusiasm during this study. In particular we would like to thank Dr. Nigel Saunders for his time and medical assistance. We would also like to take this opportunity to thank Dr. Arjan Shahani and Professor Valter de Senna for their statistical assistance during the work.
References (13)
- et al.
Prenatal and intrapartum high risk screening: II. Risk factors reassessed
American Journal of Obstetrics and Gynecology
(1979) On the quest for easy-to-understand splitting rules
Data and Knowledge Engineering
(2003)- Department of Health, Changing Childbirth, Part I. Report of the Expert Maternity Group, London, HMSO,...
Classification and Regression Trees
(1993)Case-mix groups for hospital-based home care
Medical Care
(1992)- et al.
Modelling for the planning and management of bed capacities in hospitals
Journal of the Operational Research Society
(2002)