A set of 79 European clinical and 13 blood test data was collected in each EC and COST country. Twenty-three other clinical data were tested on 400 cases. A network of over 128 centres throughout Europe collected 9500 cases according to the protocol. These were entered in a relational database, analysed in 2 centres and distributed to the collectors at intervals pro rata to cases supplied. The 16 diseases causing >1% jaundice in Europe were defined. They were described numerically in terms of the frequencies of occurrence of their symptoms in 100-1500 cases per disease. The most useful diagnostic items were identified. Similar descriptions were made on their 38 sub-diseases. The database covers adequately at least 45 disease conditions. The agreement among 4 observers examining a case according to the protocol was measured on 111 doctors in 8 countries. It was consistently 84-89% regardless of language. Norms for the 12 common blood tests used in diagnosing jaundice were based on the results of 75,000 tests. Normal ranges were obtained on 220 controls (11 centres) and laboratory variation on 6 standard samples tested in 22 centres. A problem in standards was discovered. Audit feedback on their cases in comparison to the database as a whole was provided to the centres, listing missing data, mix of diseases, symptom frequencies and the diagnostic accuracy of the program on their cases (a pan-European audit), and summaries of their cases. The use of the protocol and of feedback to the observers improved the diagnostic quality of the data by 12%, the diagnostic accuracy of the doctors by up to 50% and the omission rate of the data three-fold. Diagnostic programs were prepared by Bayesian, pattern-recognition, likelihood ratio, neural net and knowledge-bases techniques. The crude Bayesian version based on quality cases achieved 72% accuracy. Circulated for field test it attained 50% while a trial algorithm reached 75% (106 cases reported by 12 centres to date) Neural net also outperformed Bayes. The threshold for assessment of the value added by diagnostic technology was reproducibly set by computed diagnosis. The added value of each blood test was demonstrated. The residual scope for new technologies was also established, but may be affected by refinement of computed diagnosis.