BACKGROUND AND AIMS: The development of the Banff classification of renal transplant pathology has allowed the standardisation of approaches to transplant biopsy histology and reduced interobserver and interdepartmental variation. The usefulness of the Banff classification in the diagnosis of acute rejection has previously been tested by sending sections from 21 "difficult" biopsies to almost all of the renal transplant pathologists in the UK. Although the Banff classification improved reproducibility, the accuracy of diagnosis of early acute rejection was unchanged from the "conventional" approach. Perhaps this is because in making a diagnosis of acute rejection, the Banff classification uses only two features: tubulitis and intimal arteritis. To include more features on a systematic basis would be laborious for a human observer. Therefore, a Bayesian belief network was developed for this task. METHODS: The network was initialised with observations from 110 transplant biopsies. Its performance was then tested on 21 biopsies that had been seen by 37 different renal transplant pathologists in an earlier study. These biopsies had been selected to represent histologically difficult problems but, in retrospect, they all had clear diagnoses of rejection or non-rejection on clinical grounds. RESULTS: Using the Bayesian belief network, a relatively inexperienced pathologist made 19 of 21 correct diagnoses, better than had been achieved by any of the pathologists who had seen the same sections previously (17 of 21), and considerably better than the average proportion of correct diagnoses provided by all 37 renal transplant pathologists (65%). Application of the system by a second pathologist produced a tendency to overdiagnosis of acute rejection, illustrating the consequences of interobserver variation. CONCLUSIONS: In the diagnosis of acute rejection, further useful information can be extracted from features that are currently not considered in the Banff classification. Integration of data by a computer can give a more reliable diagnosis of early acute rejection, but routine application will require the development of a more sophisticated system that can also accommodate clinical data, perhaps one that can continue to "learn" as more data are entered.