Behavioral health providers and utilization management organizations control costs by minimizing the amount of treatment their clients receive. However, shorter treatment stays often mean shorter times to relapse that can lead to recurrent episodes of crisis stabilization, emergency commitment, or additional inpatient care. Currently, the amount of treatment clients receive is determined by loose practice standards or guidelines interpreted through subjective clinical judgment and cost driven utilization management. This study seeks to develop evidenced based decision making tools to improve the quality of decisions about length and intensity of behavioral health treatment. We propose to develop statistical models of the amount of inpatient mental health care needed by specific clients, models which are accurate enough to inform clinical decisions. Interjecting empirical information into the lengths of stay decision making process could lead to two outcomes: 1) more efficient allocation of resources, and 2) more confidence in utilization manager's length of stay authorizations, thereby reducing the need for continuing stay reviews and time consuming oversight of clinical care. In this pilot study, we will determine the feasibility of statistical models linking care intensity and time to readmission and their ability to aid clinical decision makers in predicting outcomes for individual patients. The actual effects of increased accuracy on utilization management and clinical practice will be pursued in a later study. To generate models accurate enough to aid in clinical decisions, we will apply a technique new to behavioral health services research, neural network modeling, and compare its predictive ability with hierarchical linear regression techniques. Because data quality may limit predictive ability, we will build both sets of models using two sets of data: (1) research quality data on veterans with post traumatic stress disorder; and (2) clinical electronic data from the Veterans Affairs national administrative databases. Accuracy of the modeling techniques and data will be compared by area under the curve analysis of receiver operator characteristic curves. This pilot study will serve as the basis for a subsequent study where these models (or similar models built on other data sets) will be introduced to clinicians and utilization management staff, and applied to real or simulated clinical decisions.
