Abstract
BACKGROUND: The density of information in digital health records offers new potential opportunities for automated prediction of cost-relevant outcomes. AIMS: We investigated the extent to which routinely recorded data held in the electronic health record (EHR) predict priority service outcomes and whether natural language processing tools enhance the predictions. We evaluated three high priority outcomes: in-patient duration, readmission following in-patient care and high service cost after first presentation. METHOD: We used data obtained from a clinical database derived from the EHR of a large mental healthcare provider within the UK. We combined structured data with text-derived data relating to diagnosis statements, medication and psychiatric symptomatology. Predictors of the three different clinical outcomes were modelled using logistic regression with performance evaluated against a validation set to derive areas under receiver operating characteristic curves. RESULTS: In validation samples, the full models (using all available data) achieved areas under receiver operating characteristic curves between 0.59 and 0.85 (in-patient duration 0.63, readmission 0.59, high service use 0.85). Adding natural language processing-derived data to the models increased the variance explained across all clinical scenarios (observed increase in r2 = 12-46%). CONCLUSIONS: EHR data offer the potential to improve routine clinical predictions by utilising previously inaccessible data. Of our scenarios, prediction of high service use after initial presentation achieved the highest performance.