Authors (including presenting author) :
KM Cheung, WM Leung, JCH Chow, KKH Bao, T Tsui, WWY Sung, KH Wong, HY Yiu and WL Leung
Affiliation :
Department of Clinical Oncology, Queen Elizabeth Hospital (all authors)
Introduction :
Prognostication by predicting the likelihood of surviving at certain time points is an important aspect in providing patient-centered palliative care service. The obvious benefit would be facilitating more realistic risk-benefit consideration in further medical treatments and discussions in advance care planning. Reliable survival estimates will boost clinicians' confidence in engaging patients in discussion regarding withholding futile medical treatments and do-not-resuscitate orders close to end-of-life. Providing patients and families with an expected disease trajectory has also been reported to improve patient satisfaction in medical encounters and reduce anxiety and depression due to unprepared clinical deterioration.
Currently, prognostic tools for ambulatory patients receiving palliative care are lacking. Most prognostic tools are validated against hospitalized patients in the hospice setting. Among current models, the Palliative care prognostic index (PPI) was most representative. The main problem of applying currently available tools is that the included prognostic factors were designed to look at the more-ill patients in ward settings and are usually not applicable to out-patients, which includes detection of confusion which tends to happen when death is imminent.
Objectives :
This is a retrospective cohort study aiming at developing a prognostic model for ambulatory cancer patients receiving palliative care, utilizing both clinical and laboratory parameters.
Methodology :
All patients attending our oncology palliative care new-case assessment clinic from 1st January 2013 to 31st December 2018 were eligible. As standard practice, patients were asked to fill in a baseline symptom questionnaire to facilitate medical care. Patients were stratified by the year of attendance and randomly sampled. A pilot sample of 240 patients was taken.
Information regarding underlying cancer, symptoms collected on baseline survey and laboratory results including blood counts, renal and liver function were collected.
Survival outcomes were dichotomized at 60 days. The answer to each question on symptoms is dichotomous. Laboratory abnormalities were dichotomized using cutoff provided on test reports when applicable.
For univariate analysis, significant prognostic factors were identified by Chi-square test and Fisher’s exact test (tests were chosen as appropriate). For multivariable analysis, logistic regression was used for variable selection and building a prognostic model. Variables with statistical significance of p< 0.1 were included. Backward selection by likelihood ratio was used for variable selection. Model fitness was assessed by Hosmer-Lemeshow test and R squared statistics. Predictive properties of the models were compared by their area-under curve (AUC) and calculation of sensitivity and specificity. The robustness of the prediction model was confirmed by bootstrapping.
This study is approved by KCC/KEC research ethics committee and supported by KCC research grant.
Result & Outcome :
240 patients were included. The median Karnofsky performance scale score is 70. The prevalence of self-reported symptoms and their prevalence are: tiredness (n=135, 55.6%), anorexia (n=106, 43.6%), shortness of breath on exertion (n=76, 31.3%), peripheral edema (n=58, 23.9%) and nausea (n=35, 14.4%). Laboratory abnormalities and their prevalence are elevated bilirubin (n=94, 38.7%), low albumin (n=94, 38.7%), leucocytosis (n=68, 28.0%), thrombocytosis (n=61, 25.1%) and lymphopenia (n=53, 21.8%). Patient health questionnaire, PHQ-9 score (cutoff = 8) was elevated in 81 patients (33.3%).
At 60 days, there were 70 death events. On univariate analyses, all the above factors were significant predictors. On multivariate analysis and prognostic model development, the most significant prognostic factors were self-reported symptoms of poor appetite and edema, laboratory test of elevated WCC above upper limit of normal (ULN), lymphocyte below normal limit and bilirubin above ULN. A prognostic model built upon these five factors showed high sensitivity of 73% and specificity of 55% in predicting survival at 60 days. The significance and prediction property was maintained after bootstrapping operation.
Compared to the most used model of palliative care prognostic index (PPI) which utilizes only clinical parameters of KPI, edema, poor oral intake, shortness of breath and confusion, the AUC of current predictive model is superior (0.782 vs 0.692).
The clear difference of clinical characteristics between ambulatory and hospitalized palliative care patients warrants development of tailored prognostic tools to bridge the current knowledge gap on providing accurate prognostication for our ambulatory palliative care patients. We hope this tool can be a powerful tool to improve the prediction of disease trajectory in terminal cancer patients, improve clinician-patient communication, inform medical decisions and facilitate advance care planning.
