Statistical interpretation of a veterinary hospital database : from data to decision support
Research was undertaken to investigate whether data maintained within a veterinary hospital database could be exploited such that important medical information could be realised. At the University of Glasgow Veterinary School (GUVS), a computerised hospital database system, which had maintained biochemistry and pathology data for a number of years, was upgraded and expanded to enable recording of signalment, historical and clinical data for referral cases. Following familiarisation with the computerised database, clinical diagnosis and biochemistry data pertaining to 740 equine cases were extracted. Graphical presentation of the results obtained for each of 18 biochemistry parameters investigated indicated that the distributions of the data were variable. This had important implications with respect to the statistical techniques which were subsequently applied, and also to the appropriateness of the reference range method currently used for interpretation of clinical biochemistry data. A percentile analysis was performed for each of the biochemistry parameters; data were grouped into ten appropriate percentile band intervals; and the corresponding diagnoses tabulated and ranked according to frequency. Adoption of a Bayesian method enabled determination of how many times more likely a diagnosis was than before the biochemistry parameter concentration had been ascertained. The likelihood ratio was termed the "Biochemical Factor". Consequently, a measurement on a parameter, such as urea, could be classified on the percentile scale, and a diagnosis, such as hepatopathy, judged to be less or many times more likely, based on the numerical evaluation of the Biochemical Factor. One issue associated with the interrogation of the equine cases was that the diagnoses were clinical in origin, and, because they may have been made with the assistance of biochemistry data, this may have yielded biased results. Although this was considered unlikely to have affected the findings to a large extent, a database containing biochemistry and post mortem diagnosis data for cattle was also assessed.