Talk to the Veterans Crisis Line now
U.S. flag
An official website of the United States government

VA Health Systems Research

Go to the VA ORD website
Go to the QUERI website

IIR 02-103 – HSR Study

IIR 02-103
Development of Survival Prediction Models for Advanced Cancer Patients
Victor Tsu-Shih Chang, MD
East Orange Campus of the VA New Jersey Health Care System, East Orange, NJ
East Orange, NJ
Funding Period: July 2003 - June 2009
Survival estimates for advanced cancer patients are currently based upon the Karnofsky Performance Status (KPS) and clinician estimates. However, their ability to predict survival is poor. Although KPS is reasonably reliable in predicting imminent death if low ( KPS<50), a high KPS does not automatically mean a long survival. Survival is usually overestimated by both clinicians and cancer patients. Advanced cancer patients tend to choose life-extending therapy if they think their survival is more than 6 months. Survival prediction models with higher accuracy and better predictive performance are needed. A better survival prediction model will facilitate decision making processes related to treatment plans for advanced cancer patients. In our previous study, recursive partitioning analysis (RPA) results suggest that KPS and patient- rated QOL (FACT physical wellbeing PWB) and physical symptom distress scale (PHYS) can be combined to refine estimates of prognosis. From an analysis of a retrospective cohort of veterans, we identified four distinct prognostic subgroups with clear cutoff points for each significant variable. The median survival time for each survival group was 29 days, 146 days, 292 days and 1.7 years. A Cox model found different survival predictors in addition to KPS and physical wellbeing; these were psychological symptom distress, global distress index and age. These two models (RPA and Cox models) need to be further validated. We proposed to develop new multidimensional survival prediction models by combining additional important variables from a multilevel QOL model.

(1)To validate newly developed survival models with the KPS, symptom distress and QOL variables from a prospective cohort of patients. We would like to describe patterns of hazards and survival in the prospective cohort and in such stratified patient demographic groups as age, race/ethnicity and spouse status; compare the patterns with those in the existing sample. We will examine the extent to which KPS, physical wellbeing and physical symptom distress are associated with survival in the prospective cohort; compare the patterns of associations with those in the existing sample. (2) To develop a complete multidimensional model by adding the physiological variables and individual characteristics to models developed for objective 1. The survival model developed from our previous study did not include all the important variables suggested by the literature, especially in the physiological dimension and individual characteristics. We will test the potential links between survival and the additional dimensions in the prospective cohort and examine the mechanisms by which the multidimensional measures are associated with survival through independent paths, mediation or moderation.

This was a prospective longitudinal study. Patients with metastatic cancer whose cancer had already been treated with standard or experimental chemotherapy with KPS < 80, or who did not wish to receive systemic chemotherapy, were recruited and stratified by KPS, PWB and PHYS based on our pilot study results. Accrual to each subgroup was specified with a goal of 40 for Group 1 (KPS<50), 135 for Group 2 (KPS>50, PWB <25), 130 for Group 3 (KPS >50, PWB >25, and PHYS > 0.6), and 55 for Group 4 (KPS > 50, PWB > 25, PHYS < 0.6). All patients were followed for survival. Each participant had demographic data recorded and a Karnofsky Performance Status (KPS) evaluation, and were interviewed once with the following validated instruments: Memorial Symptom Assessment Scale-Short Form (MSAS-SF), Functional Assessment of Cancer Therapy (FACT-G) and European Quality of Life Scale (EuroQOL), the Herth Hope Index, and the Hospital and Anxiety Depression Scale, and observer rated instruments - Comorbidity scales (Charlson Comorbidity Scale, Cumulative Illness Rating Scale), Memorial Delirium Assessment Scale and Barthel Index. Weight and laboratory tests (CBC, chemistry profile, LDH) were obtained.

The study was initiated in July, 2003 at VANJHCS and we recruited a total of 241 patients through October 2009. The goal was 360 patients; accrual was full for Groups 1 and 2 and incomplete for Groups 3 and 4, with 43 in Group I, 136 in Group 2, 25 in Group 3, and 38 in Group 4. Analyses were also done with a combined cohort of 453 patients with advanced cancer who were prospectively assessed (Prospective Cohort) similarly for other studies with 67 in Group 1, 58 in Group 2, 241 in Group 3, and 87 in Group 4, and with the original retrospective cohort.
Patients with advanced cancer who decide to receive palliative care are a diverse group of patients, ranging from patients with stable metastatic disease to those who are about to go on hospice care, with different median survivals.
Objective 1:
A combination of performance status ratings and patient reported physical symptom ratings improves survival prediction. We confirmed the relevance of the original RPA tree in this propsective validation study. The original criteria again defined 4 distinct groups (p<0.0001), but survival times were shorter in this cohort compared to their counterparts in the retrospective group. The MST for the PC was 33, 46.5, 124, and 209.5 days for the 4 prognostic groups (p<0.0001, all pairwise comparisons <0.02).
New improved RPA trees have been developed for metastatic patients that use the Karnofsky performance status, patient reported outcomes (the PHYS subscale of the Memorial Symptom Assessment Scale Short Form or the Physical Well Being scale of the Functional Assessment Therapy), and additional items. We developed 4 RPA trees for veterans with advanced cancer. These RPA trees are able to define distinct groups of patients, including patients with short median survivals of 30 days.
Cox analyses showed that a Cox survival model which includes KPS, the PHYS subscale, and 2 items from the MSAS SF was optimal.
For the domain of patient characteristics, age and weight loss were predictors of survival. Black patients had an 18% risk reduction for death compared to white patients. The presence of metastatic cancer increases the risk of death threefold compared to patients with incurable locally advanced cancer. Other features, such as gender, employment status, marital status, education, and caregiver status were not predictors.
For the domain of patient reported outcomes, other PRO scales, such as the Herth Hope Index and Bartels scale, and components of the EuroQOL 5D and the Hospital Anxiety Depression scale did correlate with survival. This result suggests that specific items can be incorporated into future predictive models.
For the domain of laboratory values, bilirubin, neutrophil count, platelet count, albumin, recent hypercalcemia, and potentially C Reactive Protein levels may be useful markers.
For disease domain, the underlying cancer was not a predictor of survival. Comorbidities had borderline significance in contributing to survival models, with items from the Charlson scale, and items from the Cumulative Illness Rating Scale for pulmonary disease, and liver disease.
Objective 2
For the multidimensional QOL model, mediation analyses showed that KPS directly mediates survival. In addition, KPS affects hope which mediates survival. In turn, KPS is mediated by symptoms and albumin.

This study took the perspective of a hematology oncology service to prospectively characterize patients at the transition from disease-oriented to palliative goals of care in patients with advanced cancer. This study represents the first comprehensive study of potential predictors of survival in veterans with advanced cancer, and the first validation of a recursive partitioning tree in modelling the survival of veterans with advanced cancer .

Objective 1 - The new RPA trees consistently identify a group of patients with very short median survival times of 30 days. Until now, the only known predictive rule was that patients with a KPS of 50% had a median survival of 2 months. The new trees have the potential to decrease costs and improve end of life care by improving risk stratification, patient counselling and decision making, and quality of end of life care by limiting unneeded interventions at the end of life. These trees will be relevant to National VA Oncology goals to improve quality of care for patients with advanced cancer, and to VA Palliative Care Program. They are patient-centered trees, and do not rely on physician judgment other than a Karnofsky Performance Status rating. They can be used by other health care professionals such as nurses and social workers. Further research should be performed with these trees.

The data on psychometric and prognostic aspects of the additional PRO instruments, comorbidity instruments, and laboratory dimensions will strengthen the evidence base for epidemiologic research in oncology and palliative care in the VA, and point to a newer generation of disease specific RPA models that may be supplemented with specific items from quality of life and symptom instruments.

Objective 2 - The new multidimensional quality of life and survival model is one of the first of its kind, and is novel for its inclusion of hope as a mediator of survival. This model may stimulate further efforts in the field.

External Links for this Project

Dimensions for VA

Dimensions for VA is a web-based tool available to VA staff that enables detailed searches of published research and research projects.

Learn more about Dimensions for VA.

VA staff not currently on the VA network can access Dimensions by registering for an account using their VA email address.
    Search Dimensions for this project


Journal Articles

  1. Buffum MD, Hutt E, Chang VT, Craine MH, Snow AL. Cognitive impairment and pain management: review of issues and challenges. Journal of rehabilitation research and development. 2007 Jan 1; 44(2):315-30. [view]
  2. Chang VT, Sorger B, Rosenfeld KE, Lorenz KA, Bailey AF, Bui T, Weinberger L, Montagnini M. Pain and palliative medicine. Journal of rehabilitation research and development. 2007 Jan 1; 44(2):279-94. [view]
  3. Chang VT, Xia Q, Kasimis B. The Functional Assessment of Anorexia/Cachexia Therapy (FAACT) Appetite Scale in veteran cancer patients. The journal of supportive oncology. 2005 Sep 1; 3(5):377-82. [view]
Conference Presentations

  1. Cogswell J, Badin S, Oliphant TL, Hoover D, Chang VT. Comorbidity and Survival of Prostate Cancer D3 patients. Paper presented at: American Society of Clinical Oncology Annual Meeting; 2008 May 30; Chicago, IL. [view]
  2. Chang VT, Bhatty MU, Hoover D, Sikder M, Gounder S, Gonzalez MK, McPherson M, Zhong F, Nazha NT, Kasimis BS. End-of-life quality-of-care indicators for medical oncology patients at a VA medical center. Paper presented at: American Society of Clinical Oncology Annual Meeting; 2008 May 30; Chicago, IL. [view]
  3. Chang VT, Scott CB, Yan H, Gonzalez ML, Einhorn J, Zhou B, Cogswell J, Crump B, McPherson M, Kasimis B. Patient-reported outcomes for determining prognostic groups in Veterans with advanced cancer. Poster session presented at: American Society of Clinical Oncology Annual Meeting; 2010 Jun 4; Chicago, IL. [view]
  4. Gonzalez ML, Chang VT, Hoover D, Bhatty M, Gounder SK, Sikder M, Wu H, Cogswell J, Crump B, Kasimis BS. Processes as predictors of outcomes of EOL care at a VA medical center. Paper presented at: American Society of Clinical Oncology Annual Meeting; 2008 May 30; Chicago, IL. [view]

DRA: Health Systems
DRE: Diagnosis, Technology Development and Assessment
Keywords: Cancer, Decision support, End-of-life
MeSH Terms: none

Questions about the HSR website? Email the Web Team

Any health information on this website is strictly for informational purposes and is not intended as medical advice. It should not be used to diagnose or treat any condition.