Prediction of the clinical course of low-back trouble using multivariable models

Abstract

The inability to predict outcome in patients with low-back pain seriously impedes clinical trials and leads to inappropriate or unnecessary treatment. This prospective study investigated the value of multivariable mathematical models to predict the 1-year clinical course of 109 patients with low-back trouble (LBT). Discriminant analysis was used to determine predictive models for outcome groups at 1 month, 3 months and 1 year. The variables selected in the analyses were subsets of 29 items from a clinical interview at presentation. These included anamnestic features of the first episode as well as symptomatic details and results from clinical tests for the current spell. The derived models successfully discriminated outcome groups with estimates of sensitivity and specificity ranging from 63 to 99%. When models from one set of patients were tested for predictive accuracy by the application of them to a different set, nonrecovery and satisfactory improvement were predicted with a 76-100% success rate. The results affirmed the importance of considering combinations of interrelated variables for prediction and discrimination in LBT. This work has demonstrated that outcome can be predicted successfully by the use of mathematic models based just on presentation data. The ability to determine homogenous groups in respect to outcome is seen as an important aid to therapeutic research; further work will enable refinement of these models for general clinical use and for incorporation into computer-based interview systems.