A lack of recognition of the proximity of death is often blamed for inappropriate admission to hospital at the end of an older person's life. An accurate prognosis for older adults living in a residential or nursing home can facilitate end of life decision making and planning for preferred place of care at the end of life. In the UK more than 80% of all deaths occur in the over 65 year old age group and 68% are over 75 years. Furthermore, more than 80% of UK deaths occur in institutions, including more than 20% in the care home setting [1
]. Higginson [2
] estimates that about three quarters of all deaths are predictable given they occur following a period of chronic illness. Even though chronic illnesses and death are expected in this population, there is still a reluctance to instigate end of life care planning [3
]. Planning, choice and communication are essential for supporting a good death in which care is individualised and symptoms well managed. Recognising that the patient is at the end of life is the essential first step in accessing and providing excellent palliative care.
General guidance has been provided in the Gold Standards Framework for Community Palliative Care [5
] prognosis but this has only limited application. However, as Mckillop [7
] states, there needs not only to be a general prognosis based on tissue diagnosis and prognostic markers but also an individual prognosis specific to the person sitting before the physician. What is needed now, to complement and operationalise the GSF, is an individualised prognostic "marker" to help initiate these vital conversations about prognosis and EOL care.
The Minimum Data Set (MDS) is federally mandated in the USA for monitoring the quality of long term care in nursing homes certified by Medicare or Medicaid (Health Care Financing Administration, 1995) [8
]. The MDS has been used previously to develop predictive models for mortality of residents in relation to specific conditions [9
] as well as in general [12
] including our own previous work [16
In our article in 2005 we reported the development and validation of a predictive model, the MDS Mortality Risk Index (MMRI), of death in 6-months for older adults in nursing homes was published [16
]. The model was based on a regression analysis of data from over 43,000 residents taken from Missouri's state MDS data. Two components of the original scoring system made it difficult to implement in practice without the aid of a computer program. Several readers of the article contacted us to enquire if a simplified system could be devised as they wished to trial the scoring system in their own practice and research and in settings where the MDS was not available. Furthermore, it was found that when relying on the usual MDS processing to get the score, too much time had passed and residents had already passed away. Thus a method of scoring that was not reliant on the MDS data collection was needed [17
]. The purpose of this paper is to briefly summarise our original work; describe the decision making process associated with simplifying the model; and then report on the performance of the revised MMRI; the MMRI-R.
Development of the original MDS Mortality Risk Index
The MMRI was developed as a result of a series of studies exploring items relating to mortality using the Minimum Data Set (MDS) for long term care in the USA. Two forms of the MDS are used. The full MDS is a questionnaire with around 400 items which is administered annually or on admission to the long term care facility. A shortened MDS is used quarterly and following any adverse events. The data for the MDS is collected by designated nurses within each facility and are kept centrally in each state and federally. The MDS is primarily used to monitor care quality in long term care facilities and is linked to funding through the Medicare and Medicaid systems in the USA. The MDS includes a broad range of items associated with the health and social wellbeing of residents and although it was not originally designed for research in long term care, the MDS provides a wealth of data and research has developed alongside its use [17
As palliative care researchers our interest began with a single item in the MDS - J5c - that simply stated "The resident has six or fewer months to live" to which a tick in the box indicated the affirmative response and clearly identifying residents known to be at the end of life (EOL). Our first study used admission MDS data to describe residents who were identified as EOL and compare their six month survival with new admissions not so designated [19
]. We found only 4% of admissions were designated EOL but that this item was a very good predictor with 50% of these residents dying in the first month of admission and only 17% still being alive 6-months later. Interestingly, 5% of the non-EOL admissions also died in the first month and 15% of them had also died by 6 months. Although J5c was a reliable predictor it was not used frequently enough to make it a useful identifier. Furthermore, our second study revealed that despite the sampling only including residents from facilities that had an active contract with a hospice service only half of the EOL designated residents were receiving input from specialist services [20
]. Our conclusion from these studies was that there might be a more reliable way of predicting death using the MDS given the breadth of data available.
On reviewing all items in the MDS we identified 50 that could be related theoretically to the likelihood of death within 6 months. The items were justified by existing empirical research and clinical experience and fell into four categories: demographics (e.g. age and sex); disease (e.g. cancer, chronic heart failure); clinical signs and symptoms (e.g. shortness of breath, poor appetite); and adverse events (e.g. falls, hospitalisation, loss of a spouse). With approval of the University of Missouri institutional review board (IRB) and the appropriate data use agreement, a dataset from the Missouri data was created of residents using the first full MDS (annual or new admission) completed in 1999. The outcome, death of a resident within 6 months of the full MDS assessment, was determined by linking the MDS data to Missouri death certificate data using social security number, sex, and date of birth. This produced a dataset with a full MDS assessment and date of death information for a sample of 43,510 nursing home residents. The death rate in this dataset over the 6 months follow-up was 26%. As described in Porock et al. [16
] stepwise logistic regression methods along with a data-splitting strategy was employed to develop a predictive model. The resulting regression model included the following 14 independent variables, listed in priority order for entry into the model: dependency with activities of daily living, shortness of breath, diagnosis of cancer, being an admission assessment, having a poor appetite, being male, general physical deterioration, unintended weight loss over the past 90 days, chronic heart failure, increasing age, renal failure, poor cognitive performance score, diagnosis of Alzheimer's disease or dementia and dehydration. The model also included interaction terms between age and cancer and deterioration and admission, giving a final model with 16 terms. A point-value for each risk factor was derived by transforming the associated logistic regression coefficients to integer values, which are then summed to give the MMRI score. Depending on the selected "cut-point" the summated scale scores can provide a highly sensitive or highly specific instrument.
In developing the original MMRI, the goal was to provide simple yet accurate instrument that could predict short term mortality. In practice the calculation of a hierarchical Activities of Daily Living (ADL) score [21
] and the Cognitive Performance Scale (CPS) [22
] were too complex to be easily implemented by hand. As the authors and others attempted to use the MMRI we found using the algorithms too unwieldy and time consuming [23
]; the computations were confusing and the number of variables excessive.
The modification of the MMRI was undertaken with two research questions in mind. First, can a simpler ADL and Cognitive performance measure produce an equally strong model? Secondly, can we eliminate some of the variables while maintaining predictive validity?