How accurate is the ‘Surprise Question’ at identifying patients at the end of life? A systematic review and meta-analysis

In both the UK and the USA, the Surprise Question (SQ; “Would you be surprised if this patient died within the next × months?”) has been suggested as a trigger for referral to specialist palliative care [1, 2].

It has been estimated that between 69% and 82% of dying patients in the UK would benefit from palliative care input (specialist or generalist) [3]. There were approximately 160,000–170,000 referrals made to specialist palliative care services in the UK between 2008 and 2009 [4]. In the USA, it has been estimated that between 1 million and 1.8 million (7.5% and 8.0%) of hospital admissions have a palliative care need [5].

It has been repeatedly shown that clinicians are inaccurate at prognostication [6‐8] and in recognising dying patients [9]. Therefore, there is an increased likelihood that patients who would benefit from palliative care are potentially being missed because validated prognostic tools (e.g. Palliative Prognostic Score [10] or Palliative Prognostic Indicator [11]) are not being used routinely, either due to their perceived complexity or inconvenience [12].

The SQ offers an alternative to a standard prognostic estimate. It does not require clinicians to collect clinical data or to use a scoring algorithm, nor does it require clinicians to make a specific estimate about length of survival; it simply asks whether the respondent would be surprised if the patient were to die within a specified time period (usually the next year). It was originally developed by Joanne Lynn as a method to identify patients who might benefit from palliative care services [13], asking the clinician: “Is this person sick enough that it would be no surprise for the person to die within the next 6 months, or a year?” Since its development, the SQ, or variants thereof, have been incorporated into clinical guidelines such as National Institute for Health and Care Excellence (NICE) for End of Life Care [14], and adopted into routine clinical practice in various settings, including hospitals, hospices, and General Practices. Although developed as a standalone item, it is now included as part of the Gold Standard Framework (GSF) proactive identification guidance tool in the UK [1], in which clinicians are encouraged to ask themselves “Would you be surprised if this patient were to die in the next 6–12 months”. A response of “No” to the SQ may trigger a referral to specialist palliative care services, or to the adoption of a palliative care approach to future care. This parsimonious approach could potentially identify more patients who need palliative care and could be incorporated into routine clinical practice with relative ease and at little or no extra cost.

Yet, how accurate or effective is the SQ at identifying people in the last year of life? Could it be used to identify people who are just days from dying? Is one clinical group more accurate at using the SQ than another? Is the SQ more accurate when used by one professional group rather than another? The present study aims to assess the accuracy of the SQ by time scale, clinician, and speciality.

In total, 357 studies were identified from the database searches (Fig. 1). No studies were identified through the grey literature search. Of those studies initially identified, 182 were subsequently found to be duplicates. There were 175 studies that were screened by title and abstract, of which 118 were excluded. Of the 57 full text articles retrieved, 34 were excluded for various reasons (Additional file 2: Table S1). Three additional studies were identified from a search of the references of the included studies. In total, 26 papers were included in this review [15, 18‐42].

We were able to obtain data from one unpublished paper with which we replaced one of the abstracts [33]. Another abstract [41] was excluded because the full paper had also been published [43].

Each paper included underwent a quality assessment with high agreement between the two raters (ICC 0.93) (Additional file 3: Table S2). No paper was excluded on the basis of their quality score.

A summary of the studies included can be seen in Table 2. Of the 26 papers, ten (38%) were from the UK and nine (35%) were from the USA. Ten (38%) of the studies were presented in abstract form, the rest were full papers. Eight of the included studies related to patients with end-stage renal disease, six related to cancer patients, four related to patients with heart failure, one study related to patients with sepsis, one study related to chronic obstructive pulmonary disease and six studies included patients with a variety of different diagnoses. The majority of studies (15, 58%) specified 12 months as the relevant period for the SQ. Eight papers specified a time frame of 6 to 12 months. Three papers specified shorter time periods.

We were able to extract data from 22/26 papers either directly from the paper or after contacting the author (Table 3).

The outcomes of 25,718 estimates were reported in the 22 studies with complete data. Patients died within the specified timeframe (whatever it was in that particular study) on 4217 occasions (16%). A response of “No, I would not be surprised” was given on 6495 occasions (25%) and clinicians’ intuitions were ‘correct’ (about whether they should or should not be surprised) in 20,964/25,718 cases (82%). Most of the correct attributions occurred when clinicians indicated that they would be surprised if the patient died within the specified time period (19,223 occasions) and they did in fact survive for that length of time (17,985 occasions).

The results across the studies (Table 3) showed a wide variation in the reported accuracy of the SQ. The sensitivity ranged between 11.6% and 95.6% and a range of 13.8% to 98.2% was reported for specificity. The PPV ranged from 13.9% to 78.6%, and the NPV ranged from 61.3% to 99%. The AUROC score (c-statistic) across the studies ranged from 0.512 to 0.822.

There was no indication of publication bias from the funnel plot (Additional file 1: Figure S1).

On meta-analysis, the pooled level of accuracy, that is the number of times the clinician correctly predicted the outcome of a patient, was 74.8% (95% CI 68.6–80.5; I ² = 99.1%, 95% CI 99–99). The studies were sorted by date of publication and there appeared to be no trend by year (Fig. 2). After a sensitivity analysis for lower quality rated scores, in which eight papers were removed [19, 31, 32, 34, 35, 41, 42, 44], the pooled accuracy level increased to 75.4% (95% CI 70.8–79.7; I ² = 96.8, 95% CI 96–98). Those studies that used a shorter time frame for the SQ (up to 6 months) had a pooled estimate of 76.6% (95% CI 61.6–88.8%; I ² 99.6%, 95% CI 100–100), and when the time frame was reduced to imminent death (i.e. 7 days, 30 days or ‘this admission’) the pooled accuracy estimate was 76.4% (95% CI 52.4–93.8; I ² 99.8%, 95% CI 100–100) (Fig. 3). The meta-regression indicated that the increase in time frame did not impact on the diagnostic accuracy of the SQ: comparing up to 30 days with 12 months (difference in accuracy = 0.8%, 95% CI –12.8 to 14.5, P = 0.901) and comparing up to 6 months with 12 months (difference = 4.3%, 95% CI –10.8 to 19.4, P = 0.561).

One unpublished paper [33] reported the results of the SQ and a modified version of the SQ, by asking clinicians “would you be surprised if this person was to be alive in a year’s time?” By rewording the question, they found that specificity was improved (i.e. correctly identifying those who do not die) but sensitivity was reduced (i.e. less correct predictions about those who will die).

One paper reported the difference in performance of the SQ when used by nurses and doctors [40]. As a result of additional data supplied by the author, it was possible to calculate the sensitivity, specificity, PPV, and NPV for both doctors and nurses (Table 4).

Doctors appeared to be better at predicting dying within 12 months, with a sensitivity of 73% (95% CI 63–82) and specificity of 74% (95% CI 70–78) compared to a sensitivity of 45% (95% CI 40–49) and specificity of 82% (96% CI 80–83) for nurses. The c-statistic for doctors was 0.735 (95% CI 0.688–0.783) compared to 0.632 (95% CI 0.608–0.655) for nurses.

Of the 22 papers that included data, eight reported research conducted within a haemodialysis context [18‐21, 27, 29, 31, 40], five were within oncology [22, 23, 25, 26, 34], and nine papers were on other areas (five on all diagnoses [28, 33, 35, 41, 44], one on sepsis [42], two on heart disease [15, 30], and one on COPD [32]). Table 5 highlights the range of scores from the individual studies reporting from each specialty.

On meta-analysis, the pooled accuracy for oncology was 78.6% (95% CI 69.7–86.3; I ² = 99.0%, 95% CI 99–99). The pooled accuracy for renal was 76.1% (95% CI 73.9–78.3%; I ² = 36.0%, 95% CI 0–72). The pooled accuracy estimate for the other group was 72.3% (95% CI 58.0–84.6; I ² = 99.1%, 95% CI 99–99) (Fig. 4).

Four studies had insufficient data to be included in the meta-analysis. Three authors responded to requests, but were unable to provide additional data to that presented in the paper [37, 39]. One author was not contactable [38].

One study [37] did not report data on the accuracy of a ‘Yes’ response to the SQ. In this study, it was reported that specialist heart failure nurses would not be surprised if 88/126 patients would have died within 12 months. In fact, 78/88 (89%) of those patients did die within 12 months. However, no data were reported about the outcomes for the 38 patients where the heart failure nurses would have been surprised if the patient had died. One study only contained data relating to the responses from the clinicians but not the outcome [36]. Gardiner et al. (2013) described the responses from doctors and nurses for two survival time points: 12 months and “death during this admission”. Of 297 patients, doctors would not have been surprised if 123 had died within 12 months, or if 50 had died during the admission to hospital. Out of a total of 473 patients, nurses predicted that 180 would die within the next 12 months, and 74 would die during the admission. The actual survival figures were not reported [36]. Lilley et al. [39] reported on the accuracy of 28 clinicians who provided responses for 163 patients. Their results show a ‘No’ SQ response was given in 93 cases (60%). They reported a sensitivity of 81% (95% CI 71–91%), a specificity of 51% (41–61%), a PPV of 52% (42–61%), and a NPV of 82% (72–91%). The exact breakdown of outcome by SQ prediction was not presented [39]. Lefkowits et al. [38] report the accuracy of the SQ within gynaecological oncology and across physicians. They asked 22 clinicians (18 gynaecologic oncology physicians and advanced practice providers (APP); four chemotherapy nurses) the SQ with a timeframe of 12 months. They reported the unadjusted odds ratio for death within a year associated with a ‘No’ SQ response; physicians 5.6 (P < 0.001), APP 9.2 (P < 0.001) and nurse 6.9 (P < 0.001). They reported that the APP group had the highest sensitivity (79.5%) and the nurses had the best specificity (75.6%). No further data was presented [38].

The reported accuracy of the SQ varied considerably between studies. The c-statistic ranged from 0.512 (poor) to 0.822 (good). The PPV ranged from 13.9% to 78.6%. This degree of heterogeneity was not uncommon in studies assessing the accuracy of clinicians’ prognostic estimates [6].

On meta-analysis, the overall accuracy of the SQ was approximately 75%. However, this overall estimate should be reviewed with caution given the low proportion of people who died within each study (16%) and the low number of high quality studies included (I ² = 99%). There was virtually no difference in level of accuracy when considering studies in which the timeframe of the SQ had been reduced, which suggests that even when the patient is thought to be imminently dying, there is only moderate accuracy and continued uncertainty. A major limitation of the meta-regression analysis completed in this review was the lack of power due to the small sample size (n = 24). Therefore, a significant difference between time frames of the SQ was less likely to be observed.

One study presented data about the differences in accuracy between different professional groups at using the SQ. This suggested that doctors’ responses to the SQ (c-statistic 0.735) may be more accurate than nurses’ (c-statistic 0.632); however, more research is needed to fully address this question.

The variation in the accuracy of the SQ might be due as much to variations in disease trajectory in the last year of life as it is due to variations in the prognostic ability of the clinicians. There is some evidence that the SQ may be slightly better when used in oncology patients rather than in renal or other disease groups. The pooled accuracy for oncology was 79% compared to 76% for renal and 72% for other disease groups. This supports the idea that patients with a cancer diagnosis have a more predictable disease trajectory [45]. However, there was little variation between the disease groups and so these data should be interpreted cautiously. Another recent review on the SQ (using different inclusion and exclusion criteria to our own) [46] also found that accuracy was slightly better in oncology patients than in other disease groups.

When proposing the original definition of the SQ, Lynn suggested that the accuracy of the outcome was not actually that relevant as all patients identified by this question would typically need the services of palliative care such as advance care planning, home care assistance or financial support [13]. However, it is often the case that clinicians delay referring to palliative care services because they feel that the judgment should be made on the basis of a reasonably accurate and relatively short prognosis. This review highlights that, intuitively, clinicians are actually quite good at excluding patients with longer survival times but that use of the SQ alone is likely to lead to identification of a substantial number of patients who are not necessarily approaching the ends of their lives. However, the review could not provide evidence about how many patients who were identified (or missed) by the SQ actually had palliative care needs.

What is apparent from the data presented is that using the SQ to identify patients with a limited prognosis will detect at least as many ‘false positives’ as ‘true positives’. In most circumstances, the consequences of misclassification by the SQ are rarely likely to be clinically important. For instance, erroneously including patients on a palliative care register who are not actually in the last year of life is unlikely to adversely affect their care, and indeed may result in better provision of holistic care. However, the recognition that half of patients included on such registers, as a result of the SQ, may not actually be in the last year of life has resource implications (e.g. additional staff time, care planning and multi-disciplinary involvement). It is thus not clear whether the SQ is a cost-effective way of identifying patients potentially suitable for palliative care. A careful balance is needed between identifying more people with unmet palliative care needs in a timely way while not over-burdening limited resources with too many patients in need of good care for long-term conditions over a much longer period.

It should be acknowledged that the SQ is usually used as part of a wider prognostic assessment that includes both general measures of performance status and disease burden along with disease-specific indicators [1, 2]. It is possible that the combination of the SQ with these other prognostic measures may well be more accurate than the SQ used in isolation. This, however, was not the focus of this systematic review and further work is needed in order to evaluate the accuracy of these approaches and to determine whether other prognostic tools (e.g. Prognosis in Palliative Care Study [47], Palliative Prognostic Score [10] or Palliative Performance Scale [48]) would be a more accurate way of identifying patients approaching the end of life, or whether accuracy could be further improved by using the SQ alongside other tools such as the Palliative Outcome Scale, which can identify and document changes in the patient’s condition over time [49].

Our study had a number of strengths. This was the first systematic review of the SQ that has attempted a meta-analysis of all studies reporting data on the SQ, including shorter time scales such as 7 days. We adopted a broad search strategy to identify any potentially relevant papers. We also obtained missing data by contacting relevant authors and requesting unpublished results. We have also appraised the quality of the papers included in our study and have conducted a sensitivity analysis to determine whether our conclusions are robust. Finally, each stage of the review process was undertaken by two independent reviewers to ensure rigor.

Our study had a number of limitations. It was difficult to devise a search strategy to capture all of the relevant papers. There is no agreed methodology for conducting a search of the literature for this type of research question. It is therefore possible that some papers may have been missed. In order to perform a meta-analysis of data, we combined studies that had evaluated the accuracy of the SQ over different time-frames, in widely differing patient groups and with different groups of clinicians. It may be argued that this is not a legitimate approach in terms of clinical heterogeneity as the performance of the SQ may be very different in each of these different circumstances. Further work is clearly needed to investigate any such differences (particularly our preliminary finding that a doctor’s estimates may be more informative than nurses’ estimates).

This review has highlighted the wide degree of accuracy reported for the SQ as a prognostic tool. Further work is required to understand the processes by which clinicians arrive at their prognostic estimates, to refine the accuracy of the SQ and to compare its performance against other more sophisticated prognostic tools, particularly in populations where a higher proportion of deaths occur.