Improving pressure ulcer prevention trial design and analysis using multi-state modelling of existing data

Pressure ulcer (PU) prevention trials are challenging due to low incidence leading to large sample size requirements. Longitudinal data at multiple skin sites per patient are collected, but commonly aggregated to a single endpoint. Multi-state models (MSM) have potential to improve efficiency of trials but there is little published on MSM as the primary analysis method.

The aim was to understand PU development natural history and better use longitudinal data for PU research design and analysis.

After fitting a 4-state progression MSM to existing trial datasets, a simulation study explored impact on power of using MSM instead of methods based on a single endpoint. This required a hypothesis test definition for multiple effect estimates in the MSM setting. State misclassification was explored using Hidden Markov Models (HMM) applied to trial data, with impact on power and bias of misclassified states assessed through simulations. Candidate state definitions in the presence of missing data were proposed and analysed using a selection model.

MSM led to increased power in some situations. When the intervention was effective in reducing onset and development across all states, follow-up could be halved from 60 to 30 days and assessments reduced from daily to every 2 - 3 days compared to the base case. State misclassification, when analysed appropriately, led to little loss of power and unbiased effect estimates, but there were convergence and identifiability concerns. Selection models were shown to be a special case of HMM and can be implemented using readily available software. Descriptive summaries of trial data suggested non-ignorable missing data, however analysis results were insensitive to different state definitions.

For disease prevention trials where participants pass through a series of health states, MSM may lead to efficient trial designs. Missing data is easily accommodated. Further work is required to develop robust modelling strategies for misclassified data.