Five Survival Tools Tested in Children Dying of Advanced Cancer

Researchers in China have compared five widely used survival prediction tools in children with advanced cancer receiving palliative care, and found that two of the simplest measures performed best. The single-center retrospective cohort study, published in BMC Palliative Care, evaluated the Karnofsky Performance Scale (KPS), the Palliative Performance Scale (PPS), the Eastern Cooperative Oncology Group Performance Status (ECOG), the Palliative Prognostic Index (PPI), and the Palliative Prognostic Score (PaP) for predicting death within 7, 14, and 30 days.

The study included 90 children with stage IV malignant tumors admitted to the palliative care unit of West China Fourth Hospital, Sichuan University, between January 2020 and March 2026. The population was extremely ill: median survival was 7.0 days, with an interquartile range of 2.75 to 16.0 days. Mortality rates were 52.2 percent at 7 days, 72.2 percent at 14 days, and 92.2 percent at 30 days, reflecting a cohort already in the final days of life.

Survival differences were analyzed using Kaplan-Meier curves and Cox proportional hazards regression, with the proportional hazards assumption verified by Schoenfeld residuals and time-dependent interaction terms. To assess how well each tool distinguished children who died within each time window, the researchers constructed receiver operating characteristic (ROC) curves and compared areas under the curve using DeLong’s test. Precision-recall curves and AUPRC values were added to address the imbalance in outcomes, and calibration was examined with the Hosmer-Lemeshow test and Brier scores. All cutoffs were explicitly exploratory and data-driven.

Kaplan-Meier and multivariable Cox analyses identified KPS and PPS as independent protective factors, meaning higher scores were associated with longer survival. The hazard ratios were 0.492 for KPS and 0.504 for PPS, each with a P value of 0.002. For the ROC analyses, the researchers reversed these scores (100 minus the original value) so that higher values aligned with mortality risk. Using the reversed scores, the 7-day AUCs were 0.691 for KPS and 0.720 for PPS.

PaP showed a different pattern. After multivariable adjustment, PaP was an independent risk factor, with a hazard ratio of 1.724 and a P value of 0.048, and it achieved a 30-day AUC of 0.705. This suggests the tool, which incorporates clinician-estimated survival, may hold preliminary potential for individualized 30-day mortality risk prediction in this specific cohort.

A key component of PaP is clinician-predicted survival (CPS), the clinical team’s own estimate of how long a patient will live. In a sensitivity analysis, CPS of 8.5 or higher was independently associated with mortality after adjustment, with an adjusted hazard ratio of 2.35 and a P value below 0.001. CPS achieved AUCs of 0.650, 0.627, and 0.651 for 7-, 14-, and 30-day mortality respectively, and its discriminative ability was statistically comparable to PaP in all DeLong comparisons. The authors suggest CPS may partly account for PaP’s predictive performance.

The two remaining tools fared less well. ECOG and PPI demonstrated limited discriminative ability in this highly homogeneous end-of-life cohort, where nearly all patients had similarly severe functional impairment. The authors caution that this finding should not be generalized to broader pediatric palliative care populations without further validation.

Calibration, the agreement between predicted and observed probabilities, was assessed for PaP using predicted probabilities derived from separate multivariable logistic-regression models for each time window. Hosmer-Lemeshow tests were non-significant for PaP, with P values of 0.523, 0.800, and 0.399 for 7-, 14-, and 30-day mortality. However, the researchers emphasize that given the limited sample size, these non-significant results cannot serve as definitive evidence of adequate model calibration. For KPS, PPS, ECOG, and PPI, the Hosmer-Lemeshow test was not applicable, and Brier scores ranged from 0.070 to 0.249.

The authors conclude that within this single-center cohort of children with advanced cancer in the final days of life, KPS and PPS show promise for macro-level prognostic stratification, while PaP demonstrates preliminary potential for 30-day individualized mortality risk prediction. They stress that all proposed exploratory cutoffs require external validation before any clinical translation, and that the findings are specific to this terminally ill, highly homogeneous population rather than pediatric palliative care in general.

Subject of Research: Comparison of five survival prediction tools for short-term mortality risk in children with advanced cancer receiving palliative care

Article Title: Comparison of five commonly used survival prediction tools (KPS, PPS, ECOG, PPI, and PaP) for 7‑day, 14‑day, and 30‑day mortality risk in children with advanced cancer receiving palliative care: a single‑center retrospective cohort study

Article References: Tan, F., Wang, Y., Wu, Y., Deng, Q., Gong, Q., Jiang, J., & Li, N. (2026). Comparison of five commonly used survival prediction tools (KPS, PPS, ECOG, PPI, and PaP) for 7‑day, 14‑day, and 30‑day mortality risk in children with advanced cancer receiving palliative care: a single‑center retrospective cohort study. BMC Palliative Care. https://doi.org/10.1186/s12904-026-02398-1

Image Credits: AI Generated

DOI: 10.1186/s12904-026-02398-1

Keywords: palliative care, pediatric oncology, survival prediction, Karnofsky Performance Scale, Palliative Performance Scale, Palliative Prognostic Score, clinician-predicted survival, retrospective cohort study, advanced cancer, prognostic models, mortality risk, BMC Palliative Care