Near-End-of-Life Trajectory in Patients with Advanced Lung Cancer

ABSTRACT

Background

Performance status is a useful prognostic index in palliative care. The objectives of this study were to investigate the prognosis of near-end-of-life patients with advanced lung cancer admitted due to deteriorated performance status, evaluate the utility of performance status measurement in predicting life expectancy, and identify prognostic factors using Bayesian proportional hazards models.

Methods

This prospective observational cohort study measured at admission, the Performance status, Palliative Performance Scale, Karnofsky Performance Status (KPS), Palliative Prognostic Index scores, and clinical biochemistry parameters of the included patients. Ninety-two patients with advanced lung cancer and deteriorated performance status admitted to our department at the NHO Yonago Medical Center between January 2016 and December 2018 were analyzed. Bayesian proportional hazards models were employed to classify remaining life prognoses. The Watanabe–Akaike information criteria (WAIC) with the lowest score was used to choose the final model.

Results

The median survival time was 25 (range, 1–107) days. The final prognostic model for the dying trajectory of patients was selected as an M-spline proportional hazards model according to the least WAIC. Independent prognostic factors for shorter remaining life were hypercalcinemia [hazard ratio (HR) 5.37, 95% confidence interval (CI) 2.11–12.9] and low KPS (HR 1.77, 95%CI 1.30–2.46).

Conclusion

Assessing hypercalcinemia and KPS can be advantageous for predicting near-end-of-life prognosis in terminally ill patients with advanced lung cancer.

Patients with cancer often experience a sharp decline in functional status^1,2,3,4 and quality of life (QOL)⁵ during the final weeks of life. These individuals frequently utilize acute-care hospitals to manage acute issues and receive symptomatic treatment,^6,7,8 especially older patients, in their final month of life.⁷ Consequently, patients and their families frequently seek guidance from physicians regarding the expected survival time. A clearer understanding of near-end-of-life trajectories can assist clinicians in planning care to address patients’ evolving needs, thereby supporting both patients and caregivers in navigating this critical period.⁴ Undernourishment is recognized as an important marker of near-end-of-life in patients with cancer.^8,9,10,11 Additionally, the Karnofsky Performance Status (KPS) is an essential tool for palliative and hospice care and is widely accepted as the gold standard for determining performance status in patients with cancer.¹² KPS provides healthcare providers, caregivers, and family members with valuable insights into a patient’s functional status.

In an evaluation of predictive validity, Mor et al. identified a significant correlation between initial KPS scores and survival time.¹³ In patients with advanced cancer, median survival times correlates with a decline in KPS (80–100: 215 days; 50–70: 119 days; 40–50: 49 days; and 10–30: 29 days).¹⁴ Patients with primary malignant neoplasms located in the central nervous system and lung exhibited a less pronounced reduction in KPS.¹⁵

Cases of cancers with multiple metastases to the brain, liver, or lung are often associated with organ failure, refractory hypercalcemia, persistent tumor bleeding, or bone marrow failure without transfusions. Despite these associations, the most intuitively valid indicator for prognosticating terminal illness near the end of life remains elusive. Glare et al. demonstrated that clinicians’ survival estimates are frequently inaccurate and overly optimistic.¹⁶ We hypothesized that in patients with advanced lung cancer approaching end of life, specific factors, including physiological decline and performance status, may serve as predictors of poor prognosis.

This study aimed to investigate the association between poor prognosis and symptoms, performance status, and biological findings at the time of final hospital admission for near-end-of-life care in patients with lung cancer. While Cox regression analysis is a well-established method for examining the effects of independent variables as prognostic factors, its predictive accuracy may be limited when applied to small datasets and/or censored data. Conversely, Bayesian proportional hazards models are considered highly effective because they can construct models with smaller datasets without losing statistical power and maintain an independent posterior distribution of parameters.^{17, 18}

In this study, we applied Bayesian proportional hazards models to predict prognostic factors in hospitalized patients with advanced lung cancer near the end of life.

MATERIALS AND METHODS

Study setting and participants

This investigation is an open, real-life, single-center retrospective epidemiological study. Data from January 2016 to December 2018 were collected from the medical records of all patients with advanced lung cancer (stages IIIB and IV) who were consecutively admitted for their final hospitalization before death, at the Department of Respiratory Medicine, NHO Yonago Medical Center in Japan. Patients whose cause of death was not associated with advanced lung cancer were excluded. The medical records of deceased patients with advanced lung cancer at our hospital were extracted from the Discharge Abstract Database.

Data regarding age, sex, and ongoing or previous tumor-directed treatment (chemotherapy, radiotherapy, biological agents, and hormonal therapy) of 102 patients were obtained from electronic or written charts. According to automated hospital record updates from the Cause of Death Registry, the date of death was entered retroactively. The study protocol was approved by the institutional review board of NHO Yonago Medical Center (approval number: 0511-01) and adhered to the principles of the Declaration of Helsinki.

Performance status measures

Data on performance status, Palliative Performance Scale (PPS), KPS, and Palliative Prognostic Index (PPI) were collected on the day of admission.

The PPS is a validated and reliable tool used to assess the performance or functional status of individuals undergoing palliative treatment and monitor end-of-life progression.¹⁹ The PPS comprises 11 performance categories delineated in 10% increments ranging from 0 (deceased) to 100% (fully functional).¹⁹

The KPS evaluates the ability of patients with cancer to tolerate chemotherapy and assists healthcare providers, caregivers, and family members in understanding a patient’s functional status.²⁰ The KPS employs a 100-point scale to rate patients’ ability to perform daily activities and their need for assistance, ranging from normal activity (100) to death (0). Scores of 80–100 indicate that the patient can perform normal activities and work; 50–70 indicate that the patient requires varying levels of assistance but can care for most personal needs; and 0–40 (defined as low) indicate that the patient is disabled and requires higher levels of care.

The PPI incorporates the PPS, oral intake, and presence or absence of delirium, edema, and dyspnea as the main parameters to evaluate performance status.²¹ We utilized the KPS because prior reports have established the performance status as a superior prognostic index.^{13, 14}

Statistics

Patient characteristics at the time of admission were summarized using standard descriptive statistics. Frequencies of categorical variables were calculated. Kaplan–Meier estimates were used to measure survival time post-admission, and log-rank tests were used to evaluate variations in survival distributions among different patient categories. The trajectory curve was plotted using an S-spline.

Proportional hazards models were deployed using Bayesian inference. The posterior distribution of the model parameters was simulated based on the data, with an assumed prior distribution for the parameters. This method improves model interpretation by offering flexibility in computing specific posterior probabilities, beyond the use of p-values. Such probabilities provide additional insights into the role of corresponding predictors within the model. Analysis was conducted using R statistical software (version 3.5.2; R Foundation for Statistical Computing, Vienna, Austria) and Bayesian inference with the rstanarm package (version 2.18.2; R Foundation).²² This Bayesian survival analysis accommodates several standard parametric distributions for baseline hazards (exponential, Weibull, Gompertz, B-spline, and M-splines).

Non-informative prior distributions were selected for the corresponding prognostic parameters. To achieve non-informativeness, normal distributions were employed as prior distributions for all risk metrics, with a commonly applied prior mean of zero and a most frequent prior SD of 100. The posterior distribution was drawn using a Markov Chain Monte Carlo (MCMC) with 2,000 replicates for the variables.

A non-informative independent framework was assumed, with a normal prior N (0, 1000) for the regression coefficients (β) and an independent gamma prior for the baseline hazard parameter (λ), with hyperparameter values. Parallel MCMC chains were initiated and continued for a sufficient number of iterations until convergence was achieved, which occurred after 51,000 replications with a burn-in of 500 iterations. Hazard ratios (HRs) with 95% confidence intervals (CIs) are used to present the results. Models were calibrated using Watanabe–Akaike information criterion (WAIC) and Leave-One-Out information criterion (LOOIC).^{23, 24} The optimal model was identified by the lowest WAIC, with a difference of 5 in WAIC considered significant.²⁵ Gelman–Rubin R-hat convergence diagnostics and trace plots were utilized to graphically evaluate the convergence of each MCMC,²⁶ with an R-hat < 1.1 indicating chain convergence.

RESULTS

Patient characteristics

From January 2016 to December 2018, 102 patients with advanced lung cancer who were admitted to our hospital and subsequently died were identified. Of these, 10 patients were excluded from the analysis due to causes of admission unrelated to advanced lung cancer (n = 5), comprising interstitial pneumonia (n = 2), ileus (n = 1), chronic obstructive pulmonary disease (n = 1), and brain infarction (n = 1), and five patients who had been treated at the time of admission. A total of 92 patients were eventually analyzed [men, n = 76; women, n = 16; mean age: 75 (range, 40–98) years], with 58 patients (63%) being > 70 years old.

The characteristics, pathology, and treatment of patients are presented in Table 1. Histological examinations revealed that 39 patients (42.4%) had adenocarcinoma, 21 (22.8%) had squamous cell carcinoma, 12 (13.0%) had small-cell carcinomas, and 32 (34.8%) had other histological types, including large-cell neuroendocrine cells or undefined non-small cells. Among patients with adenocarcinoma, nine underwent mutation analysis of the epidermal growth factor receptor gene. For ethical reasons, histological diagnoses were unavailable for 10 patients. Most patients (70.7%) were diagnosed at stage IV. The primary metastatic sites were the bones (27.7%) and lungs (24.6%), with patients potentially having multiple metastatic sites. Chemotherapy was the predominant treatment, being administered to 42.4% of patients, whereas 35 patients (38.0%) were treated exclusively with palliative support.

Table 1.  Patient characteristics.

Variables	Total patients (n = 92)
Age, years	74.9 ± 11.0
%Female	17.4
Pathology, n (%)
Non-small cell carcinoma	69 (75.0)
Adenocarcinoma	39 (42.4)
Squamous cell carcinoma	21 (22.8)
Undefined	7 (7.6)
Large cell neuroendocrine carcinoma	2 (2.2)
Small cell carcinoma	12 (13.0)
Unidentified	11 (12.0)
Overall survival time, days	383.8 ± 514.2
Treatment, n (%)
Chemotherapy	39 (42.4)
Radiation	26 (28.3)
Primary lesion and lymph node	10 (10.9)
Palliative radiation	9 (9.8)
Brain radiation	7 (7.6)
Best supportive care	35 (38.0)
Main complaints, n (%)
Shortness of breath	39 (42.4)
Appetite loss	35 (38.0)
Pain	19 (20.7)
Fatigue	14 (15.2)
Gait disturbance	8 (8.7)
Cough and/or phlegm	8 (8.7)
Fever	4 (4.3)
Paralysis	5 (5.4)
Performance
PPS level	40.4 ± 20.9
Ambulation	53.6 ± 20.5
Active and evidence of disease	51.6 ± 20.3
Self-care	52.4 ± 20.2
Intake	43.7 ± 22.6
Consciousness level	81.7 ± 24.1
KPS	47.8 ± 20.7
Low: 10–40, n (%)	39 (42)
Intermittent: 50–70, n (%)	43 (47)
High: 80–100, n (%)	10 (11)
PPI	5.8 ± 4.2
SpO2, %	91.9 ± 7.9
Biological analysis
WBC, x 109/L	10.9 ± 6.7
Neutrophil count, x109/L	9.7 ± 9.8
Lymphocyte count, X109/L	1.1 ± 1.6
Platelet count, x 106/μL	25.6 ± 11.3
BUN, mmol/L	8.5 ± 4.8
ALT, IU/L	26.5 ± 39.2
ALP, IU/L	488.9 ± 471.3
Albumin, mg/dL	3.0 ± 0.6
CRP, mg/L	8.9 ± 8.4

ALP, alkaline phosphatase; ALT, alanine transaminase; CRP, C-reactive protein; KPS, Karnofsky Performance Status; PPI, Palliative Prognostic Index; PPS, Palliative Performance Scale; SpO₂, oxygen saturation in the peripheral vessels; WBC, white blood cell.

The primary complaints upon admission are presented in Table 1, with the most prevalent being dyspnea (n = 39; 42.4%), loss of appetite (n = 35; 38.0%), and pain (n = 19; 20.7%). All eligible patients exhibited deteriorated performance status, with a mean PPS level of 40.4%.

The leading cause of death was tumor burden, accounting for 42.4% of cases, with infection (31.5%) being the second most prevalent cause. Various complications of metastases accounted for 19.8% of deaths, including multiple organ failure (10.9%), hypercalcemia (8.7%), and unspecified causes (1.1%). Hypercalcemia is diagnosed by a serum calcium concentration > 10.4 mg/dL, adjusted by serum albumin according to formula: corrected calcium = serum calcium + 0.8 * (4 - serum albumin).

Survival analysis

Median overall survival (OS) from the first hospital visit was 357 days (range: 573–924; 95% CI: 243–562) for treated patients with non-small cell lung cancer (NSCLC; n = 50) and 341 days (range: 15–563; 95% CI: 17–479) for treated patients with small cell lung cancer (SCLC; n = 6). The median OS for patients with untreated NSCLC (n = 19) and SCLC (n = 6) was 64 (range: 17–1,252, 95% CI 41–264) and 99 (range: 11–49, 95% CI 12–43) days, respectively. Analysis of the treated subgroups (treated vs. untreated) revealed statistically significant differences in OS (P = 0.034). The corresponding Kaplan–Meier curves are shown in Fig. 1a.

Fig. 1.

 Kaplan–Meier curves. (a) Overall survival, and (b) survival of last admission grouped according to treated patients with NSCLC (n = 50; thick solid line) and SCLC (n = 6; thin solid line), as well as untreated patients with NSCLC (n = 19; thick hatched line) and SCLC (n = 6; thin hatched line). Eleven patients: one treated patient and ten untreated patients, were excluded from this assessment due to unidentified pathological patterns. NSCLC, non-small cell lung cancer; SCLC, small cell lung cancer.

Survival was analyzed from the date of the last admission until death. The median survival for all patients was 25 (range: 1–107, 95% CI: 12–46) days. For patients with treated NSCLC (n = 50) and SCLC (n = 6), the median survival was 31 (range: 1–77, 95% CI: 24–40) and 19 (range: 11–49, 95% CI: 12–43) days, respectively. The median survival for patients with untreated NSCLC (n = 19) and SCLC (n = 6) was 22 (range: 1–107, 95% CI 18–34) and 17 (range: 9–45, 95% CI: 12–26) days, respectively (Fig. 1b). No significant differences in survival were observed between the subgroups. Figure 2 illustrates that the KPS was associated with the remaining life prognosis. Low KPS (≤ 40) correlated with a shorter remaining life span, as fitted by the S-spline curve.

Fig. 2.

 Trajectory curve based on KPS in near-end-of-life patients with advanced lung cancer.

KPS, Karnofsky Performance Status.

Bayesian proportional hazard model

The estimation using the rstanarm package is based on Bayesian inference. For baseline hazard models, this package supports several standard parametric distributions, including exponential, Weibull, Gompertz, B-splines, M-spline 1 (df = 1), and M-spline 2 (df = 10). Figure 3 shows the estimated baseline hazards for each of the six models. We compared the fits of these models using the LOOIC and WAIC methods (Table 2). The WAIC and LOOIC scores are widely used goodness-of-fit measures in the Bayesian proportional hazards model, aiding in the prevention of overfitting by evaluating a model’s generalizability and its predictive accuracy for future values. The optimal model with the lowest WAIC that best explained the data was the M-spline 1 (df = 1) model. Finally, we employed the M-spline 1 (df = 1) model for the actual inference.

Fig. 3.

 Estimated baseline hazards (posterior median and 95% confidence interval) for each of the six models.

Table 2.  Model performance using WAIC and LOOIC scores.

	WAIC	LOOIC
Exponential model	773.5	773.9
Weibull model	746.9	747.8
Gompertz model	748.0	749.1
B-splines model	749.7	756.3
M-splines (df = 1) model	745.6	746.5
M-splines (df =10) model	750.0	750.9

LOOCI, Leave-One-Out information criteria; WAIC, Watanabe–Akaike information criteria.

Four chains were utilized, each with 2,000 iterations. The first 500 iterations in each chain were excluded as burn-in. Table 3 shows the means, standard deviations, marginal effects, and quintiles of the posterior distributions. The number of effective MCMC samples and the convergence of MCMC chains were evaluated using Gelman–Rubin convergence statistics, with the R-hat for all variables being 1.1 or less. Figure 4 shows trace plots and posterior density distributions. Regarding the prediction of poor prognosis, the Bayesian proportional hazard model identified hypercalcemia and KPS score as significant factors. According to the results of this model, the risk of death increased with hypercalcemia (HR: 5.37, 95% CI: 2.11–12.9) and low KPS score (HR: 1.77, 95% CI: 1.30–2.46). Kaplan–Meier analysis revealed a rapid decline in prognosis for patients with hypercalcemia and a low KPS score (Fig. 5).

Table 3.  Results of M-spline (df = 1) proportional hazards model.

Variable	Mean	SD	10%	50%	90%	R-hat*
(Intercept)	1.28	0.68	0.42	1.28	2.16	1.00
CRP	0.00	0.02	–0.02	0.00	0.02	1.00
Alb	0.01	0.20	–0.26	0.02	0.26	1.00
KPS	0.57	0.26	0.24	0.57	0.90	1.00
Hypercalcinemia	1.68	0.71	0.75	1.71	2.56	1.00
Dyspnea	0.33	0.26	0.00	0.33	0.67	1.00

*R-hat is calculated as the potential scale reduction factor on split chains. Alb, albumin; CRP, C-reactive protein; KPS, Karnofsky Performance Status.

Fig. 4.

 Trace plots (a) and posterior density distribution (b) for the Weibull proportional hazards model. META, metastasis; HyperCa, hypercalcemia.

Fig. 5.

 Kaplan–Meier estimates of the survival functions for (a) KPS and (b) hypercalcemia (HyperCa). The p-value of the log-rank test comparing the survival distribution of the two groups is also provided. KPS, Karnofsky Performance Status. KPS = 0: low, 1: not low. HyperCa = 0: (–), 1: (+).

DISCUSSION

In this study, the progression of advanced lung cancer showed a rapid decline, typically within 1 month from admission until death, which was strongly associated with deteriorated performance status. Accelerated progression was more evident in patients with hypercalcemia and low KPS scores.

Although this was an open, real-life study, the median OS aligned with those of previous reports,^27,28,29 demonstrating that patients with advanced NSCLC who received chemotherapy and optimal supportive care survived longer (median OS, 8–14 months) than those who received best supportive care only (median OS, 5 months). This study revealed no difference in median survival from last admission due to deteriorated symptoms and/or performance status among the groups classified by pathology and treatment.

In cancer’s dying trajectory, a stable status is followed by a rapid decline near the end of life.^1,2,3,4 Our results suggest that the decline commences at approximately 1 month near the end of life. The present study revealed results similar to those of previous studies, which reported that the general trajectory pattern and onset of functional deterioration occur approximately 3–4 months before death, with disability levels decreasing to 49% at 6 weeks preceding death.^{3, 30} Eligible patients admitted to our hospital for deteriorated symptoms and/or performance status exhibited a median PPS level of 40%, reflecting significant functional decline.

Generally, clinicians’ estimates of survival are often inaccurate and over-optimistic,¹⁶ and the prognosis following a decline in performance status among patients with advanced lung cancer remains poorly understood. Our results potentially addressed this limitation, revealing that these patients survive for approximately 1 month until death.

We employed a Bayesian model to predict this shortened prognosis of patients with advanced lung cancer near the end of life. Bayesian models offer several advantages over classical Cox hazard models, including the ability to simultaneously process both quantitative and qualitative data and the absence of a priori hypotheses regarding the nature of the modeled relationship. The use of classical Cox hazard models to determine biomarkers yields biased results for sample sizes under 600 patients, with overestimation of the standard error of the interaction coefficient.³¹ Compared with the classical Cox hazard models, the Bayesian approach to sample size determination requires fewer participants when appropriate prior information exists.³²

Some lung symptoms were reported as independent predictors of death within 90 and 180 days.^33,34,35 However, our results demonstrated that no symptoms predicted a short prognosis. Conversely, performance status is related to survival duration.^36,37,38 A KPS score ≤ 70% has been reported as an independent prognostic factor for OS.^{36, 37} The PPS is a modification of the KPS, designed specifically to measure physical status in palliative care,¹⁹ and the PPI is a tool for predicting prognosis, using the PPS along with oral intake, edema, dyspnea at rest, and delirium.³⁸ For PPI > 6.0, survival is less than 3 weeks. Our Bayesian proportional hazard model demonstrated the utility of KPS in predicting functional decline.

Reports have highlighted accelerated deterioration during the final months of life, with notable variability in the trajectories of cerebrovascular disease, respiratory failure, congestive heart failure, and diabetes.^{3, 30} However, we excluded patients with these comorbidities. In our study, patients with poor prognosis exhibited specific characteristics of the cause of death, such as hypercalcemia. Similarly, previous studies report that 2.5–12.5% of patients with lung cancer developed hypercalcemia.^{39, 40} In a study on the survival of patients with solid tumors, including lung cancer, the median OS among patients with hypercalcemia was 40 days (95% CI 33–47).⁴¹ A review reported that patients with hypercalcemia had a median hospital stay of 4 days and an in-hospital mortality rate of 6.8%.⁴² This study revealed that hypercalcemia, in addition to KPS, is also an important factor for predicting prognosis.

This study had some limitations. First, we measured a single performance status at the time of admission to evaluate factors predicting life to death, thus limiting our ability to assess longitudinal changes. The longitudinal assessment of performance status over time is suggested to be highly beneficial for predicting survival in terminally ill patients with cancer.⁴³ Second, our study population comprised patients admitted to the acute illness unit. In patients with advanced cancer nearing the end of life, early palliative care dramatically decreased hospital and intensive care unit admissions, decreased hospital mortality, and increased hospice membership.⁴⁴ Future studies in other settings, such as palliative care units, are therefore required. Third, we evaluated the KPS, which focuses on physical function and the need for assistance but not on QOL. Cui et al. identified QOL as an independent factor predicting the survival of patients with advanced lung cancer.⁴⁵ Further studies in this direction are therefore warranted.

In conclusion, identifying prognostic factors is essential for guiding clinicians, patients, and families in planning optimal care as the end-of-life approaches. The study highlights that patients with advanced lung cancer in the near-end-of-life phase had an estimated survival time of approximately 1 month, with low KPS and hypercalcemia being a significant prognostic factor. This information may support clinicians, patients, and families in optimizing end-of-life care.