|
|
RAMS
Rapid Analysis of Mental health Services
Carinci F, Damodaran S, Yonchev A,
Cesnik B, Anderson D, Villanueva E and Talman P, 2003
Monash Institute of Health Services Research
Background
This project built upon the foundation development of the SPHIS model.
The SPHIS project has set
the principles and tools needed to directly using routine data
for population-based statistical analysis.
Following its conclusion, three preliminary meetings have been held
with the key staff from Mental Health, Southern Health and
Monash Institute of Health Service Research in Melbourne, Australia.
It was determined by consensus that a project for analysis of mental health data
in order to inform clinical practice was both feasible and desirable.
Objectives
In 2003 it became apparent that there were differences in the frequency of admission
and length of stay of patients attending the mental health centres across
the Southern Metropolitan Health Service.
In order to further the understanding of these differences
an analysis of available mental health data has been undertaken.
The main objective of the project was to explore correlates of high demand
and utilisation of health services both at the individual and organisational level.
It has been agreed that the methodological reference had to be the SPHIS approach.
Secondary objective was the realisation of a user interface that would have allowed
to simplify and automate the procedure of rapid analysis of mental services,
developing a specific prototype tool, and guidelines for refinement, update,
and further extension of the model, towards the establishment of a clinical
management system.
As a whole, the project has been planned to deliver a specialised approach
for Rapid Analysis of Mental Services (RAMS), including a first report and
a usable tool.
Tasks
To fulfil the objectives, the following steps were required:
analysis of the literature,
epidemiological design for the structure of the reference dataset,
- data extraction
- statistical analysis
- design
- development and testing of a user-friendly tool for rapid analysis.
Therefore, the project required multidisciplinary teamwork
and interaction between clinical leaders, epidemiologists,
statisticians and database managers, from both SH and MIHSR.
The Project has consisted of three sequential tasks with predefined deliverables.
At the end of each task the group members reviewed progress and determined whether
subsequent tasks needed modification in order to reach the project objectives.
Task 1: Data preparation - Coordinator: Paul Talman
1a) Definition of the characteristics of interest
An initial list of data fields from the mental health data set thought
to be most relevant for analysis was selected.
The choice of fields was assisted by the published literature and direct local knowledge,
looking at patients admitted to psychiatric services in relation to length of stay
and recurrent admissions.
1b) Epidemiologic design
Definition of the epidemiological design and statistical criteria for the
preparation of the minimum dataset.
To fulfil the objectives, it was essential that the data,
though deidentified, would still allow subject-level longitudinal analysis.
As for the SPHIS model, this required the construction of a cohort from
the multilevel database, hence proper guidelines to execute sequential query
through the data and further manipulation.
1c) Data extraction
Review of the available dataset and extraction of the selected fields
as determined by previous tasks. Determine the integrity of the extracted data
to inform the feasibility of proceeding further.
In this, ensure that patient privacy can be maintained without significant
erosion of data specificity and that additional primary data collection
is not required.
Task 2: Statistical and epidemiologic analysis - Coordinator: Fabrizio Carinci
2a) Statistical analysis
This task has used statistical and data mining tools to provide
a plausible and unbiased epidemiological interpretation to service problems,
at the best of the knowledge acquirable through the available routine datasets.
This phase included the analysis of the congruence and quality of the data,
the transformation of available variables, descriptive analysis, predictive model
design and building, graphing and mapping.
The main output was a complete statistical report based on the SPHIS model.
The deliverable allowed to set the strategies for the exploration of correlates
of high services utilisation, as indicated by abnormal lengths of stay, high impact
readmission rates, and excessive use of mental health services.
Inputs and outputs formed the basis of a user-friendly tool
to be developed as a subsequent task
2b) Database management
The statistical analysis required consultancy and support from database managers,
for refinement, further extraction and update of information from the mental health database.
2c) Clinical interpretation
Statistical analysis also required high level clinical consultancy for the
interpretation and further refinement of the statistical outputs.
Task 3: Development of a user-friendly prototype tool - Coordinator: Branko Cesnik
3a) Design of a user-friendly prototype tool
This task identified a simple way to automate the process of rapid
analysis of mental services. Specialised software allowed to link data extraction
and statistical analysis to rapid reporting and output delivery.
The tool included the possibility to browse statistical reports as documents online,
as well as downloading data in various forms for further insertion and use
into spreadsheets and end-user applications.
3b) Development of the tool
This task was devoted to the development and implementation of the
analysis tool as previously designed.
It included mechanisms to link data extraction to statistical analysis
and back to the user.
It made also possible to design a two-steps approach
to statistical analysis, where a basic level will be provided locally
using simple data processing, and a more advanced application had to be
accessible through connection to SPHIS.
The task included testing of the design within the network infrastructure
and services available within SHCN, including early training of nominated staff
who started to use the system.
Methods
A multidisciplinary team including psychiatrists, health services researchers
(biostatistician, epidemiologist and clinical scientist), a data manager and an
information officer has been formed to progress a project for a Rapid Analysis
of Mental health Services (RAMS project).
Between October 2002 and February 2003, the team met regularly to set the criteria
to produce a research-oriented, de-identified clinical database, as a scaled-down version
of the Victorian Database for Mental Health Services.
The register, also known as the Clinical Management Interface (CMI), holds a large amount
of socio-demographic, diagnosis/intervention-related activities.
The framework for diagnosis coding is ICD10 classification, Australian Modification .
To refine the criteria for the construction of a range of outcomes and risk factors,
a review of the literature was carried out, leading to the final structure of a minimum dataset,
or the "RAMS database".
Observations included in the RAMS study population refer to a cohort of patients admitted
to Southern Health acute inpatient units between 1.10.1999 and 30.9.2002.
We defined an excessive number of admissions equal to 4 admissions or more over three years.
Clinical, socio-economic and demographic risk factors were selected as potential correlates
of excessive readmissions. All characteristics with relative categories are shown in Table 1.
Fundamental population-based indicators chosen were geographical areas by postcode.
Only areas presenting at least 50 subjects in the sample were considered as potential risk
factors.
Univariate analysis was carried out using frequency tables, in combination with chi-square tests,
to describe the association between each factor alone and the outcome of interest.
Multivariate Cox regression analysis was used to evaluate the independent impact of
potential risk factors on an excessive number of admissions, adjusting for all other
characteristics.
The final model included age and gender, forced into the model as fixed effects,
along with factors retained by a backward elimination procedure, at a level of a= .05.
All results are expressed in terms of odds and hazard ratios, where values above 1
indicate increased risk, and those below 1 an equivalent percentage of decreased risk.
Although multiple observations per patient contributed to the construction of outcome
indicators and risk factors, only one observation per patient contributed to univariate
results and the multivariate risk model.
This corresponds to either the time when the outcome was recorded, or the last
observation available.
All analyses have been performed using the SAS language.
We have adopted a specialised package conceived and developed by F.Carinci
to coordinate the complex operations of cohort selection, data-management
and predictive modelling involved in the
Southern Population Health Information System (SPHIS) .
Statistical Results
The RAMS database was accessed repeatedly to match each subject with own
medical records over the study period.
An initial sample of 523,948 records was drawn from the CMI,
relative to all services provided by Southern Health during the timeframe.
A total of 10,835 subjects were assisted. Among these, N=3,243 subjects were hospitalised
at least once (52% males, average age at last contact equal to 41,
48% with at least a schizophrenic diagnosis, median LOS per admission equal to 7 days,
median cumulative LOS equal to 22), the total number of admissions and contacts for
these patients being equal to 302,664.
A total of N=347 (10.7%) experienced at least four admissions within the three years
study period. This category has been elsewhere described as "revolving door patients" .
The general characteristics of the population, by outcome group, and correspondent odds ratios,
are showed in Table 1.
Univariate results highlighted different areas of interest to be significantly
associated with excessive readmissions. Among these, several were discarded by
the multivariate model: diagnoses of substance abuse, schizophrenia and risk factors,
employment, an intermediate average value of post-discharge delay, and residence in
two geographical areas.
The multivariate model summarized the effect of variables by several
dimensions significantly associated to the outcome (results presented as hazard ratios,
HR, and 95% CIs).
Among socio-economic characteristics, the following were significant: occupied in a
subordinate or non-managerial position (HR=1.70,1.14-2.54), being assisted by a carer
(HR=1.46,1.081.96), and being born in Asia (0.60;0.37-0.96).
Compared to the overall sample, subjects with multiple readmissions were over represented
with schizophrenia and substance abuse, but these clinical diagnoses were found to be
associated to a high rate of readmissions.
Among other clinical characteristics, the following factors were significantly associated
to an increased number of admissions: mood disorders (HR=1.32,1.06-1.65), neurological
(HR=1.37, 1.05-1.80), personality disorders (HR=1.95,1.52-2.50), mental retardation
(HR=2.10;1.18-3.73).
Emotional disorders were associated to a much lower risk of multiple readmissions
(HR=0.39;0.17-0.88).
Among service-related indicators, the following were found significant: a high average LOS:
(1-4 weeks: HR=2.91,1.86-4.54;>4 weeks: HR=4.58,2.88-7.27), a high average number of contacts
per month (1-3: HR=2.37,1.68-3.33; >3: HR=6.30;4.69-8.45), and a long delay in contacting
a patient after discharge (more than a week: HR=1.62;1.30-2.03).
Four geographical areas and two clinical centres have been found also significantly
associated with the outcome. All these areas are socioeconomically disadvantaged with high
unemployment, supported residential accommodation services and low home ownership .
Discussion
This study was based on the extraction of valuable information from a very large
and well maintained administrative database, originally establihed for operational reasons.
A key feature of the study was the establishment of a multidisciplinary team that has
defined the procedures needed to transform a collection of medical records into
a system capable of modelling health services utilization.
Specific disease categories were found to be relevant.
Among these, mental retardation, and neurotic, mood, and personality disorders
were associated to an increase in risk, showing that the impact of the diagnostic
profile on health services utilization is still high.
On the other hand, emotional disorders appear to be strongly "protective"
against the risk of many readmissions.
From the socio-economic point of view, occupation, presence of a carer,
and being born in Asia are all associated to an increased number of readmissions.
The impact of cultural differences and social networks on more hospital admissions
is well documented.
The adoption of the community model was meant to target the social context
of the individual within the specific pathological diagnosis.
Nevertheless, still few basic characteristics show a marked influence
on the organization of health services.
This finding suggests that our ability to enhance understanding and acceptance
of the community model still needs to be improved.
The impact of social factors is confirmed by the result on postcode areas,
showing that the most disadvantaged areas are also associated to a more problematic
pattern of readmissions.
Some of the limitations of the present study are worth to be outlined.
First, CMI lacks important information on potential predictors (medicines, compliance)
that may play a role in readmissions.
The medication component should be linked to the database to address integrated care.
Second, readmissions are underestimated since admissions for non-mental related services
are not currently included.
Third, the role of comorbidity on readmissions is underestimated,
since CMI currently is tuned to include mainly diagnosis that are related to mental health.
The above factors, if included in the RAMS database, can shed light on particular
profiles that are worth attention, e.g. patients with chronic diseases,
poor compliance to pharmacological therapy, or complications.
However, since this information is not currently available in routine practice,
it would not be possible to apply the resullts of our predictive model
if they included unobservable factors.
Hopefully the Victorian Database could be extended to include other sources,
currently automated but not directly linked to it, so that a more accurate model
would be carried out and made directly usable.
Implication of this research
This application has shown the relevance of an operational database even for service planning.
This model helped us to focus on the problem of an excessive number of multiple admissions,
which may be caused by different factors related to clinical, demographical,
and socio-economic characteristics.
Rapid identification of the specific role of each dimension involved in care delivery
is key to support policy guidelines aimed at improving population health and systems
efficiency.
Each case need careful interpretation of results in terms of the level of association
of the different risk factors to the particular outcome.
Here, several clinical, socio-economic and service-related factors show
to be associated to increased rates of an excessive number of multiple readmissions.
Two important findings highlight the role of health services organization
on multiple readmissions.
First, a length of stay higher on average is associated to an increased rate of many readmissions.
In other terms, those who stay longer are also more likely to come back again and again.
The result is consistent with previous findings .
Second, and most important, patients showing a high rate of community contacts come
back to the hospital more often.
In particular, an average of more than 3 contacts per month is associated to a
more than six-fold risk of excessive readmissions, as opposed to an average of one contact per month.
This may reflect the high service need of this client group.
Along with the seven geographical areas of high socio-economic disadvantage,
this factor accounts for the highest risk of readmissions.
Summing up the two results, patients showing an average LOS of more than
four weeks and more than three contacts per month are likely to represent the
hardcore "frequent flyers" in terms of readmissions that should be carefully monitored
by the service providers.
Finally, the result relative to some clinical centres (increased risks for acute
child and acute care) is generally useful only as an adjustment factor for better model fit.
Conclusions
This study has exploited the potential of a very large and well maintained
administrative database for the management of mental health services.
A summary multivariate model was fit to analyse the problem of multiple readmissions.
Findings suggest that socio-economic variables more than clinical
characteristics should urge us to look at the patient context in greater and
perhaps different detail.
In particular, solutions addressing the issue of increased community
contacts and disadvantaged geographical areas should be taken under serious
consideration by health care providers to respond to current population needs.
The method applied in the RAMS project represents a model that could be easily
updated and replicated in other areas to identify trigging levels and flag patients
who are at high risk of multiple readmissions.
This model, if routinely used, can provide essential insight in the mechanisms
involved in the provision of acute care and promptly inform us on how to balance
community and inpatient services.
|
FabCarinci.net
Copyright © Fabrizio Carinci 2004
|
|
