|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
Journal of Clinical Oncology, Vol 24, No 6 (February 20), 2006: pp. 856-862 © 2006 American Society of Clinical Oncology. DOI: 10.1200/JCO.2005.02.1790 Identification in Administrative Databases of Women Dying of Breast CancerFrom the Divisions of Stroke, Rehabilitation, Health Care Services, Breast Cancer, and Outcomes; Palliative Care, Ethics at the End of Life, and Communication; and Statistical Models, Methodology, and Clinical Epidemiology; and Oncology, Bioethics, Anorexia-Cachexia, and Palliative Care, McGill University, Montréal, Quebéc, Canada. Address reprint requests to Bruno Gagnon, MD, MSC, 687 Pine Ave W, Ross 4.29, Montréal, Quebéc H3A 1A1, Canada; e-mail: bruno.gagnon{at}clinepi.mcgill.ca
PURPOSE: Palliative care is an essential component of cancer care, and population-based research is needed to monitor its impact. Administrative databases are the cornerstone of health services research. Their limitation is that cause of death is not sufficient to readily classify decedents as terminally ill for the study of the health services they received at the end of life. The study purpose is to develop and test the validity of an algorithm allowing the classification of the decedents as dying of breast cancer (BC), using administrative data. METHODS: Validation was carried out through a chart review of 119 BC decedents extracted from hospital-based databases. This algorithm was applied to 3,384 deceased women with BC representative of the whole population. The effect of the classification by the algorithm was illustrated by the shift in the distributions of age and place of death. RESULTS: The validation showed a sensitivity of 95%, a specificity of 89%, a positive predictive value of 98%, and negative predictive value of 77% for the classification of women dying of BC. Of the 3,384 decedents, 2,293 were classified as dying of, and 1,091 as not dying of BC. Women dying of BC were younger, died less often at home (6.9% v 17.9%), and in chronic care institutions (4.1% v 14.8%), and more often in acute-care beds (69.9% v 57.1%). CONCLUSION: This novel way to classify decedents is conceptually based and empirically validated through chart review and impact on distribution of age and place of death.
Hospice and palliative care is now recognized as an essential component of cancer care.1 Various initiatives in the delivery of health care services for people who are terminally ill with cancer are being developed around the world, and evidence of their clinical value2-7 and cost effectiveness is emerging.8 Population-based studies will be mandatory to monitor the effects of the implementation of these new health services. The cornerstone of health services research is administrative databases because they capture the entire insured population, and they are available at low cost. The main limitation of these databases for health service research in the area of palliative care is that they are diagnostically oriented and lack specificity about disease processes and functional limitations before death. It is these indicators that vary across diagnoses and from patient to patient, and drive the provision of care for the terminally ill.9 Given the current limitations of these administrative databases, functional status at the end of life is not directly discernible.10 Relying solely on diagnostic codes indicating the cause of death to identify people who could be provided with palliative care services is insufficient because the diagnosis listed may not reflect their condition or experience before death. This problem was raised recently with regard to appropriately classifying all decedents with a previous diagnosis of cancer, and those who were dying of cancer, meaning those who were terminally ill.11 Here are some possible scenarios of clinical course before death of persons previously diagnosed with cancer. Scenario 1: A 60-year-old man with disseminated prostate carcinoma with multiple bone and lung metastases. This patient experienced progressive deterioration, and death is associated with a terminal clinical condition usually encountered as a consequence of the previously diagnosed cancer. Scenario 2: A 40-year-old woman with breast carcinoma after mastectomy and lymph node dissection. The treatment with adjuvant chemotherapy was complicated by septic shock leading to death. Scenario 3: An 80-year-old man diagnosed with a stage II, non–small-cell adenocarcinoma of the lung treated by radiotherapy alone. Weeks later, he suffered a massive stroke and died. These three premortality clinical scenarios represent distinct terminal processes encountered by cancer patients who died and who are typical of patients being cared for by different disciplines at the end of life. We propose a model (Fig 1) representing how these clinical scenarios can be situated within a new integrated cancer care model.
In this model, certainty of dying of cancer is depicted as a function of time divided into two distinct periods, one when the cancer is still curable, and a second when it has progressed beyond cure. The roles of curative and supportive modalities have been shown to be mandated during a period of time at which the cancer is curable. However, provision of curative and supportive modalities is not at the exclusion of palliative modalities, which are targeted at the prevention of physical, emotional, and existential suffering. Thus, Figure 1 illustrates a place for provision of palliative modalities even when the cancer is still curable. However, once the cancer has passed into an incurable stage, all modalities provided are within a palliative paradigm, including hope in prolonging life, and reaching out to the grieving family after death. This model provides a conceptual framework for describing disease experience before death that would allow decedents to be classified according to premortality clinical scenarios. This is important because, as illustrated by the model, it is these scenarios that determine modalities of treatment and provision of health services at the end of life. These premortality clinical scenarios are categorized as follows: The 60-year-old man with prostate cancer was dying of cancer as he experienced a progressive general deterioration, and his death occurred as an "expected" event of his terminal cancer. The 40-year-old woman with breast cancer died of complications of cancer treatment provided with a curative intent. The 80-year-old man with lung cancer died with cancer as he died of a stroke, and not of cancer progression. It is obvious that the appropriateness of the provision of health services at the end of life cannot be studied without accurate classification of decedents into these three distinct premortality clinical scenarios. Decedents who were dying of cancer would be candidates for cancer-specific palliative modalities. Decedents who died of complications of curative treatment for the cancer needed aggressive acute care to reverse the life-threatening complications resulting from the curative modalities. Decedents who died with cancer would be candidates for palliative modalities adapted to the disease processes responsible for the terminal state. Persons who died of complications of a curative treatment for the cancer, or who died with cancer would not be primary candidates for programs involving cancer-specific palliative care services and should not be included in population-based studies on patterns of end-of-life care received by people with terminal cancer. In this study, these decedents are classified as "not dying of cancer." Such premortality clinical scenarios are easily assignable to patients in clinical studies. Population-based studies of palliative care services would be enhanced if decedents could be classified appropriately as terminally ill11 using information presently available in administrative databases. Therefore, the overall aim of this study was to develop and test an algorithm derived from diagnostic codes on administrative databases to classify persons as dying of or not dying of cancer. The specific objectives were: (1) to develop the algorithm; (2) to test the sensitivity and specificity of the algorithm against clinical information available in the patients’ hospital records in classifying decedents as dying of breast cancer; and (3) to estimate the effect, using the classification obtained through the algorithm, on the distribution of decedents by age and place of death in a general population.
Population We elected to study a group of decedents previously diagnosed with breast cancer, as this population has a mix of women who were dying of cancer and others who were not dying of cancer. The following two groups of decedents with breast cancer will be analyzed for this study: (A) a hospital-based group of deceased women and (B) a population-based group of deceased women from the Province of Québec, Canada.
Data Sources Two sources of data were used for this study: (1) administrative data maintained by hospitals and by provincial agencies and (2) the medical record. For the purposes of reporting to provincial agencies that maintain health databases, each hospital transmits a record for each person discharged. These records form the provincial hospital discharge database. A copy of this record is kept at the hospital and can be retrieved from the local hospital discharge database. The discharge records provide information on the hospital (acute or chronic), medical record number, final diagnosis and 14 secondary diagnoses, dates of admission and discharge, type of bed (acute care, chronic care, or palliative care) during hospitalization, status at discharge (alive or dead), discharge destination (home, other institution, and so on), age, sex, regional indicators, and specialized tests and procedures. For people with cancer, there are three additional fields for cancer-specific codes using the International Classification of Diseases, Ninth Edition (ICD-9). These codes provide information on the site of the primary tumor and any documented metastases to lymph nodes or distant organs. Through the patient’s medical record number listed on the local hospital discharge database, the medical record was retrieved and reviewed to identify clinical status at the time of death. Access to hospital charts for the purposes of research using only denominalized information is governed by the Directors of Professional Services of the two hospitals that reviewed the protocol and granted permission. The province of Québec maintains several centralized administrative health databases, including the hospital discharge database called MedEcho (Maintenance et Exploitation des Données pour l’Étude de la Clientèle Hospitaliére) and others for fee-for-service health care billing called RAMQ (Régie de l’Assurance Maladie du Québec). The provincial hospital discharge database receives abstracts of the discharge record from all acute care hospitals and most chronic care institutions. The RAMQ database provides records of all fee-for-service billings, and indicates, among other information, the type of act and place of act (at home, in chronic care institution, emergency room, out patient clinic, and acute care institution). The linkage of the provincial hospital discharge and billing databases is done at the ministry level through a unique and common patient identifier. Once the two files are merged, the identification number is scrambled to protect confidentiality. Permission to obtain denominalized linked data was obtained from the Commission d’accès à l’information,12 the provincial agency that oversees access to public administrative databases.
Procedures
Step 2 consisted of classifying deceased women with disseminated breast cancer into one of two possible premortality clinical scenarios: "dying of" and "not dying of" cancer. In the hospital discharge database, the final diagnosis recorded for the admission that ended in the death is identical to the cause of death on the death certificate. To be classified as dying of breast cancer, the electronic record of a deceased woman should include, in addition to evidence of metastastic disease, a final diagnosis compatible with terminal illness (Appendix 1). Testing the algorithm. To constitute the hospital-based group of deceased women, the administrative records from two local hospital discharge databases of university teaching hospitals were searched to identify women with a previous diagnosis of breast cancer who died during the years 1996 to 2002. To test the accuracy of the algorithm in classifying decedents into the two premortality clinical scenarios, we compared the classification obtained through the algorithm (Fig 2) using the information available from the local hospital administrative database (cancer-specific codes and final diagnosis) to the "true" classification determined from the review of the medical records. The review of medical records was done by a physician who specialized in palliative care (C.L.) who was blinded as to the algorithm classification. This physician reviewed the whole medical chart to look for evidence of metastatic disease and to determine if patient medical evolution was compatible with terminal cancer. A woman was classified as dying of cancer when her medical record included evidence of metastatic breast cancer and of a terminal illness related to breast cancer. Other women were classified as not dying of cancer. We considered his opinion as the gold standard. Estimating the effect of classification by premortality clinical scenario on the distribution of age and place of death in a general population. A representative cohort of women diagnosed with breast cancer during the years 1992 to 1999, previously created from administrative databases for the purpose of studying waiting time for breast cancer surgery in Québec,13 was used to create a population of decedents with breast cancer. From this cohort, we identified women who died and associated date of death. The sources of information used to identify deaths were: hospital discharge status (MedEcho), fact of death from the beneficiary file of the RAMQ, and physician’s billing for completion of a death certificate, also from RAMQ. Before applying to this group of decedents the algorithm previously developed and tested, decedents with a concomitant cancer of other origin (ICD-9 codes: 140.0 to 173.9, 175.0 to 195.9, 200.0 to 208.9, and 235.0 to 239.9) documented in the provincial hospital discharge database were excluded in order to assemble a population of decedents with only breast cancer. The algorithm was used to classify these decedents into the two clinical scenarios. Figure 3 illustrates the steps to assign place of death. The hospitalization record (MedEcho) was used to identify people who died in the hospital. This record also provides an indicator for the unit where the women received care at the time of death. For the purposes of this study, we grouped the types of units into acute care, chronic care, and palliative care units.
For women discharged alive, the destination code at the last discharge before death was used to identify women who were transferred to chronic care institutions, and they were considered to have died there. A small proportion of women were transferred to another province, and they were considered to have a place of death unknown. For the women discharged home, we used the physician billing database (RAMQ) to determine place of death. First, if there were physician billings from a chronic care institution any time between the date of the last hospital discharge and death, these women were classified as dying in chronic care. This assumes that once admitted into chronic care, a woman would not return home. Second, if in the 48 hours before death, there were more than two physician billings from the emergency department, over and above the one billing for the death certificate, women were considered to have died in the emergency department. One single billing from the emergency department was not considered sufficient to classify a woman as dying in the emergency department, and in this situation, women were considered to have been dead on arrival at the hospital, and the death was classified as a home death. Third, the remaining group of women were classified as home deaths.
Statistical Analysis The distribution of age at breast cancer death using data from vital statistics14 was plotted, as were the age-at-death distributions for women classified as dying of or not dying of breast cancer. Descriptive statistics for the distribution of places of death and the effect of the classification into premortality clinical scenarios are presented.
Testing the Algorithm Using Hospital-Based Decedents From two teaching hospitals with oncology services, 119 records of decedents with a previous diagnosis of breast cancer in 1996 to 2002 were retrieved and reviewed. In Tables 1 and 2, the results of the classification using chart data and using the algorithm are presented. For disease extension (Table 1), the algorithm had a sensitivity of 0.98, meaning that 98% of women (101 of 104 women) with disseminated breast cancer were correctly classified as subjects with disseminated disease, and a specificity of 0.87, meaning that 87% of women (13 of 15 women) with locoregional breast cancer were classified as subjects with locoregional disease.
Table 2 shows how the premortality clinical scenario classification using the algorithm agreed with that from the medical record review. The sensitivity of the algorithm was 95% (95 of 100 women), and the specificity was 89% (17 of 19 women). The positive predictive value of the algorithm for dying of was 0.98, meaning that 98% (95 of 97) of women dying of cancer were appropriately classified. The negative predictive value of the algorithm was 0.77, meaning that 77% (17 of 22) of women not dying of cancer were correctly classified, and were therefore excluded from the group of women dying of breast cancer.
Estimating the Impact of Ignoring Premortality Clinical Scenarios Figure 4 shows the classification of the 3,384 decedents with breast cancer using the algorithm. Of these women, 2,380 women had a disseminated breast cancer documented at least once in the whole study period, and 1,004 women never had more than locoregional breast cancer documented. The final classification resulted in 2,293 women said to be dying of breast cancer and in 1,091 women, not dying of cancer. Most of the latter group of women died with breast cancer.
Figure 5 presents the age distribution of death using the algorithm for the premortality clinical scenarios and vital statistics. Women who were dying of breast cancer are compared with women who were not dying of breast cancer. This graph clearly shows that women who were not dying of breast cancer are older, as their age distribution does not correspond to the age distribution provided by vital statistics for breast cancer death.
Impact on Place of Death The distributions of places of death are reported in Table 3 for the entire cohort and are stratified according to each of the two premortality clinical scenarios. Overall, only 10.5% of women died at home; 13.4%, in palliative care beds; and 65.5%, in acute-care beds. However, the classification of women by clinical scenario modified these proportions. In comparison to women who were not dying of breast cancer, the women who were dying of breast cancer died more rarely at home (6.9% v 17.9%), had more access to a palliative care bed (18.0% v 3.5%), remained preferentially in acute-care beds (69.9% v 57.1%), died less often in the emergency department (1.3% v 6.5%), and were rarely transferred to a chronic-care institution (4.1% v 14.8%).
The testing of the algorithm provided supporting evidence of the adequacy of the algorithm for the purposes of classifying persons dying of breast cancer. The algorithm was quite accurate for disease extension, which was determined longitudinally through an examination of diagnostic information during all subsequent hospitalizations following the first mention of breast cancer. The distribution of age at death was affected by applying the algorithm to the base cohort. Women who were not dying of breast cancer were older than women who were dying of breast cancer, as is strikingly demonstrated by the marked, right shift of the age distribution of the former group. This is expected, as this group includes mostly women who died with breast cancer and were likely to be older with other comorbidities likely to explain death. By separating these two groups of women, it was possible to demonstrate a better correspondence between the age distribution of women who were dying of and the age distribution as reported by vital statistics; this is in contrast to the difference in the age distribution for women classified as not dying of breast cancer. This finding shows the advantage of refining the classification of decedents. Furthermore, for the study of end-of-life care, the inclusion of all women with breast cancer who died rather than specifying whether they were dying of breast cancer or not, would have potentially overestimated the population in need of end-of-life cancer care by almost 33% (Table 3). As presented in Table 3, the classification of women into the two premortality clinical scenarios had important influence on the proportion of women who died in each specific place of death. The proportion of women who died at home would have been overestimated (10.5%) by including all women, as only 6.9% of women who were dying of breast cancer died at home. This would have important ramifications for providing home-based palliative care services. More importantly, the number of women with terminal breast cancer who died at home is much lower than would be reported without proper classification. The application of the algorithm to classify women also had corresponding effects of respective proportions of death in acute care, chronic care, and palliative care. The shift in distribution of place of death suggests that a better identification of women really requiring palliative care services is essential in order to remove the effect introduced by a group of older women requiring other types of terminal care. In a recent article, McCarthy et al15 found that access to hospice services seemed to differ by type of cancer, with 27% of people with lung cancer having access, compared with 16.4% of women with breast cancer. This apparent discrepancy could well be explained by misclassification of premortality clinical scenarios.16 We found that only two thirds of decedents with breast cancer were dying of breast cancer, and one third died of or died with their disease. Applying our estimate of the size of the population in need of such care would increase their proportion to 24.2% (essentially be reducing their denominator by a third). This rate is quite similar to the population with lung cancer, removing any appearance of unequal access. In their reply, McCarthy et al17 recognized that by restricting the analysis to participants who died from cancer, as ascertained by cause of death, this difference decreased. This shift in proportion of patients receiving hospice care depending on the way cohort of patients is defined, from tumor registry or death certificate, has been pointed out as a major source of bias.11 Our study is an attempt to decrease this bias by excluding decedents who were not dying of cancer from a group of terminally ill cancer patients. Our study could not evaluate whether our algorithm improves classification of decedents above the information provided by death certificates, as these death certificates were not accessible on the provincial health database. We think that it does because it included only women with stage IV breast cancer and also removed a proportion of women with advanced breast cancer who died from other causes and who therefore were not dying of breast cancer, and women who died of complications of curative treatment of cancer. In this article, Bach et al11 suggested that such studies should not be carried out because of the bias they introduce. Our study suggests that new research methodologies could be developed to reduce this bias to an acceptable level. Our methodology to identify cases of women dying of breast cancer is applicable in the other provinces in Canada as they maintain similar administrative databases. It could also be used in other countries where hospital discharge databases are maintained as long as they contain cancer diagnosis codes, and causes of death could be documented. Our study has some limitations. The testing of the algorithm did not include women dying outside the hospital. To include decedents outside the hospital would have necessitated constructing a cohort of advanced breast cancer patients and follow them prospectively until death, this was beyond the scope of this research. We do not believe that the validity of the algorithm is affected by this limitation as the main step of the algorithm is to classify appropriately the extension of cancer and this step should not be influenced by place of death as the information is obtained from hospital discharge extracts available over many years before death. Place of death was determined in some cases (home and chronic care institutions) based solely on physicians’ billing. However, this represents a very small proportion of cases and error in this classification should have a very minimal effect overall. Conceptually, classifying all people with advanced cancer as dying of represents a paradigm shift for care providers, as dying of implies need for services that are not curative, but are palliative in intent. We have represented this paradigm shift graphically in Figure 1, redrawn from a model of continuity of care proposed by the American Society of Clinical Oncology18 and the European Society of Medical Oncology.19-21 We feel that this representation provides a framework for defining populations suitable for research into end-of-life health services research. This model may also serve to promote better integration of care for patients and families. We have tested an algorithm to classify women as "dying of" or "not dying of" breast cancer, using administrative data. The validation process suggested a high degree of accuracy. Appropriate classification according to premortality clinical scenario resulted in a shift of the age and place of death distribution more adequately defining the population in need of end-of-life cancer care. This algorithm was developed for database studies of health services at end of life for women with breast cancer, but it could be readily adapted to other cancer types or other diseases.
The appendix is included in the full-text version of this article, available online at www.jco.org. It is not included in the PDF (via Adobe® Acrobat Reader®) version.
The authors indicated no potential conflicts of interest.
Authors' disclosures of potential conflicts of interest and author contributions are found at the end of this article.
1. Pain Relief and Palliative Care: National Cancer Control Programmes: Policies and managerial guidelines (ed 2). WHO Library Cataloguing-in-Publication Data, 2002, pp 83-91 2. Bruera E, Neumann CM, Gagnon B, et al: Edmonton Regional Paliative Care Program: Impact on patterns of terminal cancer care. Can Med Assoc J 161:290-293, 1999 3. Rinck GC, van den Bos GA, Kleijnen J, et al: Methodologic issues in effectiveness research on palliative cancer care: A systematic review. J Clin Oncol 15:1697-1707, 1997[Abstract] 4. Grande GE, Todd CJ, Barclay SIG, et al: Does hospital at home for palliative care facilitate death at home? Randomised controlled trial. BMJ 319:1472-1475, 1999 5. Hanks GW, Robbins M, Sharp D, et al: The imPaCT study: A randomised controlled trial to evaluate a hospital palliative care team. Br J Cancer 87:733-739, 2002[CrossRef][Medline] 6. Rabow MW, Dibble SL, Pantilat SZ, et al: The comprehensive care team: A controlled trial of outpatient palliative medicine consultation. Arch Intern Med 164:83-91, 2004 7. Jordhoy MS, Fayers P, Loge JH, et al: Quality of life in palliative cancer care: Results from a cluster randomized trial. J Clin Oncol 19:3884-3894, 2001 8. Bruera E, Neumann CM, Gagnon B, et al: The Impact of a Regional Palliative Care program on the Cost of Palliative Care Delivery. J Palliat Med 3:181-186, 2000[CrossRef][Medline] 9. Lunney JR, Lynn J, Foley DJ, et al: Patterns of functional decline at the end of life. JAMA 289:2387-2392, 2003 10. Mayo NE, Poissant L, Ahmed S, et al: Incorporating the International Classification of Functioning, Disability, and Health (ICF) into an electronic health record to create indicators of function: Proof of concept using the SF-12. J Am Med Inform Assoc 11:514-522, 2004 11. Bach PB, Schrag D, Begg CB: Resurrecting treatment histories of dead patients: A study design that should be laid to rest. JAMA 292:2765-2770, 2004 12. Commission d'accès à l'information: Commission d'accès à l'information. http://www.cai.gouv.qc.ca/index-en.html 13. Mayo NE, Scott SC, Shen N, et al: Waiting time for breast cancer surgery in Quebec. Can Med Assoc J 164:1133-1138, 2001 14. National Cancer Institute of Canada: Canadian Cancer Statistics. Toronto, Ontario, Canadian Cancer Society, Health Canada, 2000 15. McCarthy EP, Burns RB, Ngo-Metzger Q, et al: Hospice use among Medicare managed care and fee-for-service patients dying with cancer. JAMA 289:2238-2245, 2003 16. Gagnon B: Hospice use in Medicare beneficiaries with cancer. JAMA 290:1578-1579, 2003 17. McCarthy EP, Phillips RS: Hospice use in medicare beneficiaries with cancer. JAMA 290:1578-1579, 2003 18. Cancer care during the last phase of life: J Clin Oncol 16:1986-1996, 1998[Abstract] 19. Cherny NI: European Society of Medical Oncology (ESMO) joins the palliative care community. Palliat Med 17:475-476, 2003 20. Cherny NI, Catane R, Kosmidis P: ESMO takes a stand on supportive and palliative care. Ann Oncol 14:1335-1337, 2003 21. Cherny NI, Catane R: Attitudes of medical oncologists toward palliative care for patients with advanced and incurable cancer: Report on a survery by the European Society of Medical Oncology Taskforce on Palliative and Supportive Care. Cancer 98:2502-2510, 2003[CrossRef][Medline] Submitted March 29, 2005; accepted August 17, 2005. Related Editorial
This article has been cited by other articles:
|
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2006 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
TOP
ABSTRACT
INTRODUCTION
