Minimum Data Needed on Patient Preferences for Accurate, Efficient Medical Decision Making

Abstract

Involving patients in their health care decisions improves patient satisfaction and outcomes, but can be costly because of the materials and time needed to discuss the many issues that constitute a medical problem. The authors present a framework for identifying the minimum data needed on patient preferences for accurate medical decision making. The method is illustrated for the decision of whether patients with end-stage renal disease should undergo short or long hemodialysis treatments. The value of health states to patients was modeled as a function of six outcomes: survival, uremic symptoms, hospital days per year, the inconvenience associated with long dialysis treatment duration, presence of hypotension during dialysis, and presence of other symptoms during dialysis. The relative importance of each outcome was characterized in a value function by weights referred to as preference-scaling factors. These factors were varied at random over a uniform distribution to simulate different patterns of patient preferences on the six outcomes. The decision model's recommendation was recorded for each simulation. Classification and regression-tree (CART) and stepwise logistic regression analyses were applied to these recommendations to determine the scaling-factor levels that predict short or long treatments. Knowledge of scaling factors on only the inconvenience of long dialysis treatment duration, the worst alive state of health on hemodialysis, and presence of hypotension identified the correct treatment in more than 97% of simulations. Fifty-five patients undergoing hemodialysis were then surveyed for their scaling factors on the six dimensions of well-being. When patients' scaling factors were applied to the predictive rule generated by CART using simulated scaling factors, more than 94% of treatment decisions were classified correctly—sensitivity and specificity of predicting long dialysis were 89% and 100%, respectively. These statistical techniques applied to results of a decision model help identify the minimum data needed on patient preferences to involve patients in efficient and accurate decisions about their health care.