Diagnostic models for impending death in terminally ill cancer patients: A multicenter cohort study

Abstract

Abstract Background Accurately predicting impending death is essential for clinicians to clarify goals of care. We aimed to develop diagnostic models to predict death ≤3 days in cancer patients. Methods In this multicenter cohort study, we consecutively enrolled advanced cancer patients admitted to 23 inpatient hospices in 2017. Fifteen clinical signs related to impending death were documented daily from the day when the Palliative Performance Scale (PPS) declined to ≤20–14 days later. We conducted recursive partitioning analysis using the entire data set and performed cross‐validation to develop the model (prediction of 3‐day impending death‐decision tree [P3did‐DT]). Then, we summed the number of systems (nervous/cardiovascular/respiratory/musculoskeletal), where any sign was present to underpin P3did score (range = 0–4). Results Data following PPS ≤20 were obtained from 1396 of 1896 inpatients (74%). The mean age was 73 ± 12 years, and 399 (29%) had gastrointestinal tract cancer. The P3did‐DT was based on three variables and had four terminal leaves: urine output (u/o) ≤200 ml/day and decreased response to verbal stimuli, u/o ≤200 ml/day and no decreased response to verbal stimuli, u/o >200 ml/day and Richmond Agitation‐Sedation Scale (RASS) ≤−2, and u/o >200 ml/day and RASS ≥−1. The 3‐day mortality rates were 80.3%, 53.3%, 39.9%, and 20.6%, respectively (accuracy = 68.3%). In addition, 79.6%, 62.9%, 47.2%, 32.8%, and 17.4% of patients with P3did scores of 4, 3, 2, 1, and 0, respectively, died ≤3 days. Conclusion We successfully developed diagnostic models for death ≤3 days. These may further help clinicians predict impending death and help patients/families prepare for their final days.

Keywords

• advanced cancer
• diagnostic models
• impending death
• recursive partitioning analysis