Article Text

PDF
Accuracy of prognosis prediction by PPI in hospice inpatients with cancer: a multi-centre prospective study
  1. Sivakumar Subramaniam1,
  2. Andrew Thorns2,
  3. Martin Ridout3,
  4. Thiru Thirukkumaran2 and
  5. Thomas Richard Osborne4
  1. 1EllenorLions Hospice, Gravesend, UK
  2. 2Department of Palliative Medicine, Pilgrims Hospices/University of Kent, Canterbury, East Kent, UK
  3. 3Department of Statistics, University of Kent, Canterbury, Kent, UK
  4. 4Departments of Palliative Care, Policy and Rehabilitation, King's College London, London, UK
  1. Correspondence to Dr Sivakumar Subramaniam, Pilgrims Hospice, Canterbury, Kent CT2 8JA, UK; drsivakumar{at}doctors.org.uk

Abstract

The Palliative Prognostic Index (PPI) is a prognostication tool for palliative care patients based on clinical indices developed in Japan and further validated by one study in the UK. The aim of this study was to test its prediction accuracy in a large inpatient hospice sample. The admitting doctor in three inpatient hospices calculated the PPI score on admission. Two hundred and sixty-two patients were included in this study. Based on the PPI score, three subgroups were identified. Group 1 corresponded to patients with PPI ≤4 and the median survival of 53 days (95% CI 40 to 80 days). Group 2 corresponded to those with PPI >4 and ≤6 and the median survival 15 days (95% CI 12 to 26 days) and Group 3 corresponded to patients with PPI >6 and the median survival of 5 days (95% CI 3 to 7 days). In this study, PPI was able to identify patients’ likelihood of dying within 3 weeks with a sensitivity of 64% and specificity of 83%. It was able to identify a 6-week survival chance with a sensitivity of 62% and specificity of 86%. A one-unit increase in PPI score was estimated to increase the hazard for death by a factor of 1.33 (95% CI 1.26 to 1.40), based on fitting a stratified Cox proportional hazards model. The authors conclude that PPI can be used to predict prognosis for patients with advanced cancer.

Statistics from Altmetric.com

Introduction

Prediction of prognosis is an important skill for palliative care professionals. It is helpful for patients and clinicians when making informed decisions about treatments, preparing for likely future events and enabling timely referrals to appropriate services such as inpatient hospice care.

Many studies have confirmed that the majority of patients with cancer want to know their prognosis.1–4 A recent systematic review in 2009 found that many patients wanted some type of broad indication of their prognosis.5 A retrospective cohort study with 3445 patients concluded that diagnosing and communicating a terminal diagnosis formally to patients, could reduce hospital admissions and late referrals to palliative care services and increase the possibilities of dying at home.6 However, one of the factors identified as contributing to the healthcare professionals’ reluctance to talk about prognosis was ‘uncertainty of prognostication’.7

Research confirms that healthcare professionals are unable to prognosticate accurately.8 ,9 Many studies have attempted to identify clinical and biological indices to improve prognosis prediction.10 A retrospective study with 1081 patients with cancer concluded that performance status was a key prognostic factor in terminally ill patients with cancer.11 A prospective study in 2005, conducted with 466 hospice patients, concluded that the Palliative Performance Scale (PPS) was a strong predictor of prognosis.12 However, rather than identification of just the prognostic factors, organising those factors into clinically usable tools might be of more practical value for professionals. Many studies have attempted this.13 ,14

We were interested in introducing a ‘user friendly’ tool to assess prognosis without invasive assessments and investigations. Our literature search suggested that the best tool to achieve these aims was Palliative Prognostic Index (PPI). PPI was developed in 1999 by Morita et al14 in Japan for inpatient hospice patients. At the start of our study, the only validation was in Japan.

Aim of the study

To examine the accuracy and to further validate PPI as a prognostication tool for UK hospice inpatients with cancer.

Methods

The study took place at three UK inpatient hospices. These hospices were specifically selected for the study as they were located in Kent and were easy to access. All patients above 18 years with cancer, who were admitted from January to June 2009, were included in the study. The study was approved by the East Kent Local Research Ethics Committee, who advised against seeking patient consent, as the data collection involved no change to clinical practice and the data were anonymised.

The five clinical indices used to calculate PPI, such as PPS (a modified Karnofsky index), oral intake, oedema, dyspnoea at rest and delirium, were recorded by the admitting doctor, and the PPI score was calculated (table 1). The PPS15 measures performance status in 10% decrements from 100% (healthy) to 0% (dead). Oral intake was classified as normal, moderately reduced (reduced, but more than mouthfuls) or severely reduced (mouthfuls or less). Patients with parenteral feeding were recorded as normal intake. The other three indices, oedema, dyspnoea at rest and delirium were recorded as present or absent. Delirium was recorded according to the criteria outlined in the Diagnostic and Statistical Manual of Mental Disorders, 4th edition, but was recorded as absent if it was considered to be caused by a single medication, following Morita et al.14

Table 1

Palliative Prognostic Index (PPI)

All the patients were followed up to their death or for 6 weeks after the recruitment of the last patient at each hospice (both inpatients and discharged), whichever was earlier. The status (alive or dead) of the discharged patients was ascertained from the patient’ records. All the centres had electronic patients’ record and their respective community palliative care teams recorded the date of death by checking with the patients’ general practitioner. This was helpful for confirming the date of death for those patients who were discharged from the hospice.

Data analysis

Patient ages at the three hospices were compared using one-way analysis of variance. Contingency table χ2 tests were used to compare the frequency data between hospices; for example, to compare prevalence of symptoms.

Following Morita et al,14 the PPI scores were used to classify patients as PPI ≤4 (predicted survival more than 6 weeks), 4>PPI (predicted survival 3–6 weeks) and PPI >6 (predicted survival less than 3 weeks). Kaplan–Meier curves were constructed to compare the survival between groups, and the log-rank test was conducted to give a formal test for differences in survival between groups. A Cox proportional hazards regression model was used to compare the effect of PPI on survival at different hospices. Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were calculated for predicting survival for less than 3 weeks and less than 6 weeks based on patients’ PPI values. Statistical analyses were completed using R, V.2.10.0.16

Results

Data were obtained for 265 patients, but 3 patients were omitted from the analysis because their PPI score was not collected in strict accordance with the protocol. Table 2 summarises the characteristics of the remaining 262 patients at the three hospices. Approximately half of the patients at each hospice were men, and there was no significant difference between hospices in the average age of the patients.

Table 2

Summary of patient characteristics’ and other factors of PPI

Table 2 also shows the prevalence of symptoms among the patients and the average PPI score at each hospice. There were some significant differences in prevalence of symptoms between patients at different hospices. In particular, Hospice 1 had a higher proportion of patients with delirium (17% vs 9% and 5%) and a lower proportion of patients with normal oral intake than the other two hospices (14% vs 21% and 32%), and Hospice 2 had a higher prevalence of oedema (68% vs 50% and 39%) and a higher proportion of patients in the worst PPS category (34% vs 22% and 20%). Mean PPI score was lowest for patients at Hospice 3 (5.4 vs 6.1 and 6.3), although the differences were not statistically significant.

Figure 1 shows Kaplan–Meier survival curves for the three groups of patients classified by PPI score. Curves are shown for each hospice separately and for the combined data from all three hospices. Actual survival against predicted survival is shown in table 3, and table 4 gives the corresponding median survival times. There are clear differences in the survival of groups with different PPI scores, as confirmed by log-rank test (p<0.001). However, except for the patients with PPI score greater than 6, survival times tended to be shorter at Hospice 2 than at the other hospices, and at this hospice, there was little difference in survival between the patients with medium and high PPI scores.

Figure 1

Kaplan–Meier survival curves for patients grouped by Palliative Prognostic Index (PPI) score.

Table 3

Actual survival time in each three hospices against the Palliative Prognostic Index (PPI) scores

Table 4

Median survival times for patients grouped by Palliative Prognostic Index (PPI) score, with 95% CIs*

The Cox proportional hazards model assumes that a one-unit increase in PPI score increases the hazard of death by a constant factor (the HR). Fitting a stratified Cox model, allowing the baseline hazard to differ between hospices, gave an estimated HR of 1.33 with 95% CI (1.26 to 1.40). However, a check of the proportionality assumption revealed evidence of a significant departure from proportionality for Hospice 2, even though the model appeared satisfactory for Hospices 1 and 3. There was no evidence that the HR differed between these two hospices (p=0.15), and the estimated HR from pooling the data from Hospices 1 and 3 was 1.40, with 95% CI (1.32 to 1.49). Thus, each one-unit increase in PPI is estimated to increase the hazard of death by 40% at these two hospices.

The sensitivity, specificity, PPV and NPV for predicting survival as less than 3 weeks (based on PPI >6) or more than 6 weeks (based on PPI ≤4) are shown in table 5. Values for individual hospices are estimated less precisely than the overall values due to smaller sample sizes; SEs of estimates are typically in the range 5%–10%. Nonetheless, sensitivity and specificity were higher at Hospice 1 than at the other two hospices.

Table 5

Sensitivity, specificity, positive predictive value and negative predictive value for predicting survival as less than 3 weeks or more than 6 weeks based on Palliative Prognostic Index (PPI)

Table 6

Comparison of this study results with those of other studies (sensitivity, specificity for predicting survival as less than 3 weeks or less than 6 weeks based on Palliative Prognostic Index (PPI))

Discussion

In this study, the PPI score was able to differentiate between the groups of patients based on their PPI scores. The median survival for patients with scores ≤4 was 53 days (95% CI 40 to 80), PPI score >4–6 was 15 days (95% CI 12 to 26) and the group of patients with the PPI scores >6 was 5 days (95% CI 3 to 7).

We considered two predictions based on the PPI score, that is, an individual with PPI >6 would survive less than 3 weeks and an individual with PPI ≤4 would survive for 6 weeks or more. The sensitivity of these predictions was only moderate, 62% and 64%, respectively. Thus, more than one-third of the patients who survived for less than 3 weeks had PPI ≤6, and more than one-third who survived for 6 weeks or more had PPI >4. On the other hand, the specificity of the predictions was much higher. Among patients who survived 3 weeks or more, 86% had PPI ≤6 and patients who survived for less than 6 weeks, 83% had PPI >4. This was comparable to the results from Morita et al and Stone et al. However, the overall accuracy was 60% of the calculations for the PPI in this study.

The PPV of the PPI score of >6 was high, with 87% of the individuals with a PPI score of >6 survived less than 3 weeks. Conversely, the lower PPI scores (<4) had high NPV, with 86% of the individuals with PPI <4 surviving less than 6 weeks. These point to a similar conclusion, that is, a high PPI is generally a good indicator of a short survival time. Conversely, a low PPI is not a strong predictor of a long survival time. This is most probably because the PPI cannot anticipate the sudden deterioration that occurs in some patients. Clinicians who choose to use the PPI scores to prognosticate should always explain this risk when discussing prognosis with patients or their carers.

While the tool can be used to classify patients into three prognostic groups, this study neither compared the use of the PPI with clinicians’ unaided decision making nor did it use the tool as a clinical decision aid.

The limitations of the PPI are not a reason to discard it. One of the reasons that the prognostication tools might not be used routinely in palliative care is that the tools are not user friendly and not validated enough to gain the clinicians’ trust. This study has generated further evidence and the experience gained from the study could be used to improve this tool by further studies. The fact that a different population produces new data should allow modification to improve its performance.17 It is possible that different weighting to the clinical indices could help to improve its accuracy. In this study, the calculation of the PPI on admission to the hospice before simple symptom control measures have been put in place to improve some of the clinical indices, might have affected its accuracy. Hence, delaying calculation of the score by a day or two may help in this situation.

Another important factor noticed in this study was the variability in results among the three hospices. There were no statistically significant differences in the characteristics of the study population of the three hospices. However, the authors did find that in Hospice 2 patients had more acute events (3 out of 8 patients in the PPI score group of <4) compared with others and this could have affected its results. Even though this could account for some variance, this could also be an indication that the calculation of the PPI scores by different clinicians (inter-rater variability) might have affected its accuracy. This could be a significant factor for the results, as the number of clinicians calculating the scores in this hospice was more than the other two hospices. We could not analyse the data for individual accuracy or its relationship with the experience of the clinicians, as this was not included in the data collected. However, it would be impossible to expect the calculation to be 100% the same from different clinicians. Training the clinicians to use the tool could rectify this aspect. However, the variance of the results between the Hospices was unexplainable due to the small size of the data in each centre. This also requires further testing with a larger sample across more centres.

Conclusion

The inaccurate nature of prognostication in palliative care is well documented.8–10 ,18 Even though a number of prognostic tools are available, they are not widely used in clinical practice.19 This prospective, multi-centre study examining the use of the PPI in hospice inpatients with cancer demonstrated comparable results with other studies from Japan14 and Ireland20 as shown in table 6, and provided further external validation of the PPI as a clinical tool for estimating prognosis.

The PPI was shown to differentiate between the prognosis of groups of patients based on their scores with moderate sensitivity and high specificity. This means, although the PPI might not identify all the patients at risk of dying, the patients identified by the PPI score are highly likely to die within the predicted time. This tool could be useful for the clinicians to facilitate to achieve preferred place of care, make informed decisions about interventions for these group of patients. Further studies are needed to identify its usability in routine clinical practice and whether it can be extended to other diagnoses and places of care.

Acknowledgments

This study was sponsored by University of Kent. The authors would like to thank all the healthcare professionals from EllenorLions Hospice and Pilgrims Hospices who generously gave their time to support the study.

  • Received 18 March 2012.
  • Revision received 12 January 2013.
  • Accepted 13 March 2013.

References

View Abstract

Footnotes

  • Contributors SS was responsible for the conception and design, data interpretation, drafting article, revised critically and final approval of the version submitted; AT made substantial contributions to design, drafting the article, revising it critically for important intellectual content and final approval of the version to be published; MR made substantial contribution to the design, data interpretation and approval of the final draft for submission; TT and TRO made substantial contribution to the design, data collection, critical revising of the final draft for submission.

  • Competing interests None.

  • Ethics approval The East Kent Local Research Ethics Committee.

  • Provenance and peer review Not commissioned; externally peer reviewed.

Request permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.