Temporal, Probabilistic and Surprise Question Approaches of Clinician Predicted Survival in Advanced Hepatopancreaticobiliary Cancers: A Prospective Cohort Study

INTRODUCTION

Hepatopancreaticobiliary (HPB) cancers, comprising primary and metastatic cancers of the liver, pancreas, gallbladder and biliary system, have historically been deemed to have a poor prognosis.^[1] The innocuous and latent symptomatic phase, which delays medical attention, results in a tardy diagnosis and an advanced stage at detection with limited curative-intent treatment options. This profoundly affects the quality of life of patients, increases the symptom burden, and leads to financial and time-related toxicities. In previous studies, the median overall survival (OS) rates after diagnosis of advanced hepatocellular, gallbladder and pancreatic cancers were 2–3, 3–6 and 6–11 months, respectively.^[2-4] For these patients, accurate prognostication is a crucial step in incorporating timely specialist palliative care (SPC) services, supporting patient-centred care and facilitating decision-making processes regarding treatments and the preferred place of care/death, thereby ensuring an adequate quality of End-of-Life (EOL) care and provision of dignified death. As death is a probabilistic event, it may not be possible to ascertain the exact time of death.^[5] Moreover, with disease progression, the illness trajectory of HPB cancers could be interspersed with complications such as obstructive jaundice, encephalopathy, infections, hepatic failure, malignant ascites, peritoneal carcinomatosis and gastrointestinal haemorrhage, the outcomes of which would further limit prognosis.^[6]

Research on prognostication has predominantly focused on cancers in general, relying on clinician prediction of survival (CPS) and various scales and indices.^[7-10] However, there is a paucity of studies specifically addressing the prognostication of HPB cancers within the context of SPC. CPS is commonly used by physicians as it is simple, prompt and convenient.^[11] It considers various patient- and disease-related factors, including illness trajectory, performance status, symptom burden and clinicians’ assessment of laboratory markers. Although it incorporates known prognostic factors, variability in the weighting given to each parameter, in conjunction with prior clinical knowledge, clinical experience and personality, often results in heterogeneous and inaccurate estimations.^[12,13] Hui et al. have demonstrated that the accuracy of CPS is comparable to or exceeds that of other prognostic indices across various timeframes.^[14] Perez-Cruz et al. compared probabilistic and temporal approaches in patients with advanced cancer during the final days of life. The findings showed the probabilistic approach was more accurate than the temporal approach.^[15] With the establishment of early integration of palliative care as a standard practice for metastatic cancers and advancements in cancer therapeutics, SPC physicians must deploy the best available prognostication tool to efficiently guide patients and caregivers with respect to informed decision making, better healthcare utilisation, and advance care planning, and to align the goals of treatments with what matters the most to patients.

MATERIALS AND METHODS

RESULTS

Among the 186 patients evaluated for the study, 160 were found to meet the eligibility criteria and were subsequently enrolled. One hundred and thirty-four (83.8%) of the patients died at the end of 90 days. Of the remaining 26 patients who were censored, 19 (11.9%) were alive (right-censored) and 7 (4.4%) were lost to follow-up (left-censored).

The mean age of the patients in our study was 56 (±11.98) years, 81 (50.6%) were female, and 142 (88.8%) were married. The predominant sites of primary malignancy were the gallbladder (n = 80, 50%), liver (n = 60, 37.5%) and pancreas (n = 20, 12.5%). The median ECOG-PS score on presentation to the OPD was 3 (IQR = 1). The median number of years of experience of the seven Palliative Medicine physicians participating in our study was 5 years (Range: 3–25 years). The detailed characteristics are listed in Table 1.

The median OS was 24 days (95% confidence interval [CI]: 21.16–26.83). The median OS for CPS range estimates for ≤7, 8–30, 31–60, 61–90 days were 36 days (95% CI: 0.00– 78.94), 16 days (95% CI: 12.20–19.79), 30 days (95% CI: 21.77–38.23) and 70 days (95% CI: 57.60–82.39), respectively (P < 0.001) [Figure 1].

Close

Figure 1:

Kaplan–Meier curve describing survival curve stratified for clinician prediction of survival categories. CPS: Clinician predicted survival.

Export to PPT

Temporal clinician-predicted survival (CPS)

The median continuous temporal CPS was 40 days (95% CI: 33.75–46.25 days). The median difference between AS and temporal CPS was 16 days (95% CI: 12.59–19.42). The correlation coefficient between temporal CPS and AS is 0.48 (95% CI: 0.33–0.60; P < 0.001). Among 134 patients included in the analysis, 54 (40.3%) were accurate, 57 (42.5%) were optimistic/overestimated, and 23 (17.2%) were pessimistic/underestimated. The mean difference in the overestimated group was 23.4 days. The mean difference in the underestimated group was 29.3 days. The c-statistic value for CPS at day 60 was 0.62 (95% CI: 0.52–0.72, P = 0.03) and at day 90 was 0.73 (95% CI: 0.60–0.86, P = 0.001) [Table 2].

Table 2: AUROC analysis for temporal CPS at 7, 30, 60 and 90 days.

Survival assessment timepoints	c-statistic (95%CI)	P-value
7-day	0.25 (0.11-0.38)	<0.001
30-day	0.40 (0.30-0.49)	0.05
60-day	0.62 (0.52-0.72)	0.03
90-day	0.73 (0.60-0.86)	0.001

AUROC: Area under receiver operating curve, CI: Confidence interval. Bold and italic values specify significant P-values (P<0.05)

Receiver operating characteristics for 7, 30, 60 and 90 days have been described in Figures 2-5.

Close

Figure 2:

Receiver operating Clinician predicted survival at day 7. Blue line signifies the model’s curve for Clinician predicted survival for the specified timepoint. Red line signifies a diagonal line that represents a classifier based on random guessing.

Export to PPT

Close

Figure 3:

Receiver operating characteristic curve for Clinician predicted survival at day 30.

Export to PPT

Close

Figure 4:

Receiver operating characteristic curve for Clinician predicted survival at day 60.

Export to PPT

Close

Figure 5:

Receiver operating characteristic curve for Clinician predicted survival at day 90.

Export to PPT

The OA of the categorical temporal CPS prediction was 65(48.5%). The optimistic/overestimated estimates were 45 (33.6%), and the pessimistic/underestimated estimates were 24 (17.9%). McNemar’s test, used to compare the OA of continuous and categorical predictions, showed that the categorical approach was significantly more accurate than continuous temporal CPS prediction (48.5% vs. 40.3%, P = 0.04).

DISCUSSION

Prognostication is pivotal in palliative care, particularly for advanced HPB cancers, which have a poor prognosis and limited survival. Accurate CPS predictions are linked with increased adherence to patient care preferences, less aggressive EOL care, timely hospice enrolment, initiation of ACP, reduced intensive care unit (ICU) admissions and length of stay in the ICU and increased compliance with Do-Not-Resuscitate outcomes, resulting in enhanced family satisfaction and improved patient outcomes.^[19-22] We found that CPS was most accurate at day 90, followed by day 7- 86.5% and 83.5% for temporal, 86.3% and 65.4% for probabilistic, 93.4% and 89.8% for SQ approaches, respectively. The SQ approach was more accurate than temporal and probabilistic approaches. The least accurate prognostic estimation was recorded on day 30 for all three approaches. Temporal CPS displayed a significantly moderate diagnostic accuracy on days 60 and 90, as determined by AUROC analysis. The median survival for patients in our cohort was 24 days.

The median difference between AS and CPS was 16 days, with values ranging between −29.3 and +23.4 days. The continuous temporal estimates demonstrated various metrics according to the method of estimation-40% accuracy, as per the AS ± 33% method, a moderate correlation between AS and continuous temporal CPS and c-statistic values ranging from 0.25 to 0.73. The SN for 7-day prediction was 5.3%, but the accuracy was 83.5%, largely contributed by the ability to detect true negatives. In the AUROC analysis, temporal CPS displayed moderate diagnostic accuracy on days 60 and 90. However, the accuracy varied depending on the estimation method, as reported recently.^[23] Hiratsuka et al., in their study, recorded a higher AUROC estimation (0.86 at 7 days, 0.82 at 30 days) as compared to estimation by the AS ± 33% method-30%. The accuracy of the categorical estimates at different time points ranged from 58.9% to 86.5%. The temporal CPS yielded OAs of 40% and 49% for the continuous and categorical estimates, respectively. The categorical estimates were more accurate than the continuous estimates (P < 0.001). The overall temporal CPS accuracy was similar to that reported in previous studies.^{[12,15,23,24]} The possible reasons for low accuracy may be tied to the subtle nature of the temporal question, lack of upper threshold of clinician response, inclination to err on the side of caution to maintain hope and protect patients and caregivers from psychological distress contributed by poor prognosis, stringent accuracy criteria for patients with short, expected survival,^[12] increased likelihood of acute complications in our subgroup, the non-compliance to horizon effect^[25] and high clinician dependence, making it less reliable.^[16,26] Physicians with experience who are not directly involved in patient management tend to provide more accurate prognostication than those actively engaged in the patients’ care. Inaccuracy attributed to underestimation is frequently associated with inherent unpredictability stemming from overall general conditions, comorbidities and sudden unexpected deaths.^[27] In the Indian setting, the coping mechanisms of patients and caregivers are influenced by philosophical and religious beliefs, including stoicism, acceptance of fate, karma and spiritual convictions.^[28] These factors contribute to a reduction in overall distress, possibly enhancing survival outcomes and changing the prognostic course for patients. Other factors contributing to CPS inaccuracy include resource constraints, high patient volume and inadequate training in the application of prognostic instruments. Although AUROCs are recommended for estimation, they may not capture clinically significant differences in prognostic discrimination, and the percentages shown in the AUROC are not the same as the predicted probability that the patient would die in the given period.^[23] Notably, the low sensitivity and high specificity of day-7 prediction are concordant with the appearance of physical signs of death in the last week of life.^[29]

The accuracy of probabilistic prediction exhibited a nonlinear pattern, with the lowest accuracy and maximum uncertainty observed on day 30. Conversely, the highest accuracy and least uncertainty were documented on day 90. Survival time was predicted to be accurate in the short term (days of survival) or long term (months of survival). In contrast, prediction in the intermediate time frame necessitates comprehending the risk of developing acute life-threatening events (infections, refractory dyselectrolytemia and acute thromboembolism), which frequently occur in the weeks preceding death. The outcomes of these events are challenging to predict, even in patients not on any DMT, due to inter-patient variability and pre-existing disease burden. This aligns with other studies that demonstrate the challenging nature of predicting intermediate survival.^[12,30] Notably, previous studies did not exclusively consider patients receiving BSC only. The use of prognostic tools in this timeframe could potentially enhance the accuracy of clinicians. In addition, probabilistic estimation had the least accuracy among the three. This finding is contradictory to that of previous studies by Hui et al. and Perez-Cruz et al., who compared temporal and probabilistic predictions and found that probabilistic predictions were more accurate than temporal predictions.^[12,15]

Our results showed that SQ demonstrated an increasing trend in sensitivity and PPV, with the highest values recorded on day 90. Meanwhile, specificity was highest on day 7, and accuracy was highest on day 90, followed by day 7. The lowest accuracy was recorded for the prediction on day 30. We found that sensitivity and PPV decreased as the timeframe for the SQ decreased, similar to the findings reported in previous studies.^[31,32] The possible reason could be the hesitancy of physicians associated with answering the SQ with shorter time frames, to avoid ‘taking risks’, as being wrong could have a significantly undesirable impact on clinical decision making. Our results for day-7 SQ prediction are similar to those of a study by Kim et al., who assessed SQ at 7, 21 and 42 days in a multicentre prospective cohort study in a palliative care inpatient unit.^[33] The specificity at different time points followed a nonlinear pattern, contrary to the finding observed by Hamano et al.^[34] that as the duration of the prediction period decreases, the SQ demonstrates a lower accuracy in identifying patients who will die. Our study was performed primarily in an outpatient clinic setting with shorter SQ timeframes in a cohort of patients with advanced HPB cancer, which differs from previous studies, which focused on patients with diverse advanced stage cancers in hospital-based settings.^[32-34] Further studies with a shorter timeframe SQ are recommended in patients with advanced cancer presenting to different settings, preferably multi-centric trials.

Prognostication ultimately translates to prognostic communication, which can be significantly influenced by inaccuracies. This is particularly evident in the facilitation of goals-of-care discussions and decision-making processes in the EOL phase, consequently affecting overall family satisfaction and confidence in the healthcare team.^[35] Harding et al. highlight sociocultural factors affecting prognostic communication in India. Healthcare professionals demonstrate a greater inclination towards disclosing prognostic information, while families prefer a collaborative approach to decision-making centred around the family for patient care to alleviate anticipatory psychological stress. A paternalistic approach, in conjunction with patients’ passive reception of information and reliance on physicians for decision-making, is prevalent. Financial constraints and caregivers’ opportunity costs significantly impact information-giving and decision-making processes. Consequently, patients often lack disease insight, autonomy in decision-making and informed consent to treatment, which results in them accessing futile treatments and experiencing treatment poverty.^[36]

The SQ approach was more accurate than the temporal and probabilistic approaches across all time points, highlighting the critical role of question framing in enhancing clinicians’ predictive abilities. To the best of our knowledge, this is the first study that focuses on prognostication in patients with advanced HPB cancers and on BSC in a low-middle income country (LMIC) setting. Notably, a novel feature of our study is that we have determined and compared all three approaches of CPS in variable timeframes, and an effort to determine which modality would be the most accurate approach for CPS, within the same disease-specific patient cohort, in diverse palliative care settings at a large tertiary cancer centre. In addition, we recorded continuous and categorical estimates as two separate questions, which is one of the limitations of the previous studies.

CONCLUSION

Prognostication is a dynamic process, as opposed to a single event. Recent clinical practice guidelines recommend using CPS in all patients with a survival of a few months or less, regardless of their DMTs. We further emphasise through the means of this study to employ SQ as the preferred approach for determining CPS. Supplementing the 30-day and 60-day estimations with serial CPS or employing prognostic tools can potentially improve accuracy. In addition, more research focusing on prognostication should be conducted, especially in LMIC settings where survival estimates differ from the developed world and among diverse cancer groups.