Predicting future unwell-being of hospice patient using real world patient reported outcome measures, evaluation of the practical application of a frequentist and Bayesian approach.

Abstract

Background In palliative care, Patient-Reported Outcome Measures (PROMs) gauge symptom severity. These PROMs help develop predictive models supporting proactive care. Longitudinal data prediction models in palliative care need to incorporate survival, and joint models, though appropriate, have not been applied in this context. Objective Evaluate the practical application of frequentist and Bayesian joint modeling for predicting future hospice patient unwell-being using real-world PROMs. Methods Design: a mixed-methods approach, combining prospective cohort design for practical application and cross-sectional design for evaluation. Utilize the Sympal cohort from Dutch hospices (August 2015 – May 2023). Data collected in clinical practice were entered into the research database including, patient characteristics, illness characteristics and symptoms and concerns assessed by means of the Utrecht Symptom Diary – four dimensional (USD-4D) clinical practice. Outcomes: Assess processing, congeniality, software implementation, practical application, and estimate comparisons. Predictors include pain, sleep disturbance, dry mouth, dysphagia, anorexia, constipation, nausea, dyspnea, fatigue, anxiety, depressed mood, time for oneself, bearing life events, letting go of loved ones, feeling harmony in life, being at peace with the end of life, unwell-being and value of life, assessed by means of the USD-4D. In addition, patient characteristics and illness characteristics were added to the models. Analysis: Conduct descriptive analysis for the evaluation and employ frequentist and Bayesian joint modeling in three steps: 1)multiple imputation, 2)linear mixed modeling for model optimization, and 3) joint modeling combining linear mixed model and Cox proportional hazards model. Results Joint models using frequentist and Bayesian approaches predicted future unwell-being, life value, pain, dry mouth, anorexia, fatigue, depressed mood, and life balance with comparable estimates. Bayesian approach required ten times more computation time. Frequentist approach analysis lacked congeniality,requiring additional programming during multiple imputation, linear mixed models, and joint modeling. Bayesian approach necessitated a specific statement on imputation covariates. Both methods are not fully implemented in statistical software, limiting validation and use for future patients. Discussion / Conclusion Future unwell-being prediction, accounting for survival, involves physical, psychological symptoms, and sociospiritual concern. Despite longer computational times, the Bayesian approach fits the analysis better. Additional implementation is essential for applying Bayesian models in developing predictive models for future palliative care.