Japanese Journal of Clinical Oncology

ANALYSES

First, to identify the determinants for 3 and 6 weeks survival, the training set was reanalyzed by multiple logistic regression analyses using whether or not patients lived for 3 or 6 weeks as the dependent variable and PPS and the clinical symptoms investigated were entered into the equation in a backward stepwise fashion. PPS was categorized as 10-20, 30-50 and [ge]60 (reference category) using dummy codes; in addition, reduced oral intake and edema were represented as severe, moderate and absent (reference category). Dyspnea at rest and delirium were categorized as 0 (absence) and 1 (presence).

Second, on the training populations, the predictive values of the classifications by a combination of identified risk factors were examined. As it is impossible to consider all combinations of the symptoms because the sample size was not large enough and our intention was to formulate a judgment tool that would be as simple as possible, the classification was defined simply by whether or not a patient had at least one of the risk factors identified for each survival.

Finally, its validity was examined on the test data set.

All analyses were performed using the Statistical Package for the Social Sciences (ver. 6.1).

Table 1. Patients' backgrounds

	Training set	Test set	P
Number of patients	150	95
Males	85 (57%)	51 (54%)	0.65*
Mean age	67 ± 13	66 ± 13	0.39[dagger]
Death during study periods	140 (93%)	88 (93%)
Survival			0.90[Dagger]
Median survivals of deaths, days (range)	27 (1-181)	26 (1-217)
Median survivals of survivors, days (range)	284 (180-395)	277 (194-349)
Primary sites			0.55*
Lung	37 (25%)	18 (19%)
Stomach/esophagus	35 (23%)	16 (17%)
Colon/rectum	26 (17%)	14 (15%)
Breast	9 (6.0%)	6 (6.3%)
Prostate/bladder/kidney	7 (4.7%)	7 (7.4%)
Pancreas	7 (4.7%)	6 (6.3%)
Ovary/uterus	6 (4.0%)	4 (4.2%)
Biliary system	4 (2.7%)	6 (6.3%)
Blood	4 (2.7%)	1 (1.1%)
Liver	3 (2.0%)	8 (8.4%)
Soft tissue	3 (2.0%)	2 (2.2%)
Neck	3 (2.0%)	2 (2.2%)
Unknown	4 (2.7%)	4 (4.2%)
Others	2 (1.3%)	1 (1.1%)

*Chi-squared test. ^[dagger]Unpaired two-sided t-test, ^[Dagger]Log-rank test.

Table 2. Performance status and symptom prevalence

	Training set	Test set
Palliative Performance Status
PPS = 10-20	36 (24%)	22 (23%)
PPS = 30-50	83 (55%)	61 (64%)
PPS [ge] 60	31 (21%)	12 (13%)
Clinical symptoms
Moderately reduced oral intake*	50 (33%)	29 (31%)
Severely reduced oral intake*	44 (28%)	36 (38%)
Moderate edema	24 (16%)	8 (8.4%)
Severe edema	24 (16%)	26 (27%)
Dyspnea at rest	27 (16%)	17 (18%)
Delirium	29 (19%)	36 (38%)

*24 (training set) and 7 (test set) cases with total parental nutrition for bowel obstruction were rated as normal oral intake.

RESULTS

Table 3 shows the risk factors for 3 and 6 weeks survival identified by multiple logistic regression analyses on the training sample. The contributing factors to 6 weeks survival were PPS 10 or 20, edema, dyspnea at rest and delirium; whereas for 3 weeks survival, only PPS 10 to 20, dyspnea at rest and delirium were identified.

Second, two hypothetical classifications were defined by the above risk factors on the training set. In the classification to predict 3 weeks survival, the cases with PPS 10 or 20 or dyspnea at rest or delirium were classified in the group with a predicted survival of shorter than 3 weeks. In the other classification for the prediction of 6 weeks survival, the cases with PPS 10 or 20 or edema or dyspnea at rest or delirium were classified in the group whose prognoses were predicted to be shorter than 6 weeks.

Table 4 shows the associations between predicted and actual survival groups. These classifications predicted shorter than 3 weeks survival with sensitivity 75%, specificity 84%, positive predictive value 73%, negative predictive value 85% and overall accuracy 81%, while 6 weeks survival was predicted with sensitivity 76%, specificity 78%, positive predictive value 81%, negative predictive value 72% and overall accuracy 77%.

Finally, to validate these groupings, the test sample was examined. Again, these classifications were appropriate for 3 and 6 weeks survival prediction. Table 5 shows the associations between predicted and actual survival groups for the test sample. This classification predicted shorter than 3 weeks survival with sensitivity 85%, specificity 84%, positive predictive value 79%, negative predictive value 88% and overall accuracy 84%, while 6 weeks survival was predicted with sensitivity 79%, specificity 72%, positive predictive value 80%, negative predictive value 70% and overall accuracy 76%.

Table 3. Risk factors for shorter than 3 or 6 weeks survival

	<3 weeks survival	<6 weeks survival
Palliative Performance Status
PPS = 10-20	3.0 ± 0.83 [20]***	1.9 ± 0.79 [6.8]**
PPS = 30-50	-	-
PPS [ge] 60 (reference)	0.0 [1.0]	0.0 [1.0]
Clinical symptoms
Oral intake
Moderately reduced	-	-
Severely reduced	-	-
Normal (reference)	0.0 [1.0]	0.0 [1.0]
Edema
Moderate	-	1.1 ± 0.57 [3.1]**
Severe	-	1.9 ± 0.70 [6.8]***
Absent (reference)	0.0 [1.0]	0.0 [1.0]
Dyspnea at rest	1.0 ± 0.58 [2.8]*	1.2 ± 0.65 [3.2]*
Delirium	1.8 ± 0.61 [5.9]***	1.6 ± 0.71 [4.8]**

Data are shown as regression coefficient ± standard error [odds ratio].
*P < 0.1; **P < 0.05; ***P < 0.01; -, not significant.

Table 4. Associations of predicted and actual survival groups in training set

	Actual survival
	<3 weeks	>3weeks	Total
Predicted survival
<3 weeks	41	15	56
>3 weeks	14	80	94
Total	55	95	150
	Actual survival
	<6 weeks	>6 weeks	Total
Predicted survival
<6 weeks	63	15	78
>6 weeks	20	52	72
Total	83	67	150

Table 5. Associations of predicted and actual survival groups in test set

	Actual survival
	<3 weeks	>3weeks	Total
Predicted survival
<3 weeks	34	9	43
>3 weeks	6	46	52
Total	40	55	95
	Actual survival
	<6 weeks	>6 weeks	Total
Predicted survival
<6 weeks	44	11	55
>6 weeks	12	28	40
Total	56	39	95

DISCUSSION

Accurate prediction of survival is difficult but critically important not only for patients and their family members but also for health care providers. To determine whether medical intervention is appropriate for a patient or not, physicians are required to predicate patients' prognoses concretely. Especially in palliative care settings, risks and benefits should be estimated on every intervention. For example, although forced rehydration has been routinely utilized for terminally ill cancer patients, the importance of withholding or withdrawing rehydration is becoming stressed to avoid distress related to overhydration (12,13). Although corticosteroids and sedatives are often administered for symptom relief, inconsiderate long-term use can cause undesirable effects such as Cushing appearance, oral candidiasis and tolerance (14,15). Also, pharmacological treatment for depressed cancer patients whose death is estimated to be within a few weeks is not recommended because it requires at least 2 weeks for improvement and could cause severe anticholigenic side effects (16). Thus, whether a medical procedure should be performed or not essentially depends on survival evaluations.

While some clinical symptoms such as performance status, nutritional disturbance (e.g. anorexia, lowered oral intake, weight loss, edema, dysphagia), dyspnea especially at rest and delirium have been revealed as useful prognostic indicators in previous studies (1-8), no simple tools to predict concretely how long terminally ill cancer patients are likely to live have been established. In this study, using the clinical symptoms which had been identified as prognostic factors (5), we developed a simple and practical aid for survival prediction and examined its validity on two independent series of terminally ill cancer patients.

This study suggests that whether or not patients live shorter than 3 and 6 weeks can be satisfactorily predicted by a combination of performance status, edema, dyspnea at rest and delirium. The sensitivity, specificity and positive/negative predictive value of this classification were within an acceptable range: sensitivity is 75-85%, specificity is 72-84%, positive predictive value is 73-81%, negative predictive value is 70-88% and overall accuracy is 76-84%. The predictable accuracy of our methods is a match for or better than experts' estimations reported previously (1), where experienced physicians predicted 4 weeks survival with sensitivity 48-64% and specificity 73-91%. Moreover, this evaluation is fairly simple, does not require any invasive examinations and can be completed by a trained nurse. In conclusion, we believe that survival prediction for terminally ill cancer patients can be achieved by this simple technique with sufficient accuracy to choose adequate palliative interventions.

To develop more valid tools for survival prediction, the use of consistent values instead of categories, adoption of other items and more detailed classifications of each clinical symptom should be considered in further studies.