We have previously described a physiological model of glucose-insulin interaction in insulin-dependent (type 1) diabetes mellitus which has been developed for patient and medical staff education about diabetes mellitus, as well as possibly for clinical use. The model attempts to reflect the underlying (patho)physiology of insulin action and carbohydrate absorption in quantitative terms such as insulin sensitivity, volume of glucose and insulin distribution and maximal rate of gastric emptying. The model's predictions also allow a 24 h simulation of patient blood glucose profiles to be generated. Advice is provided by a qualitative knowledge based system which suggests what the next step in improving glycaemic control might be for a given patient, e.g. 'increase before breakfast long-acting insulin by 2 units'. Validation work performed on a previous version of the knowledge based system has demonstrated that it can provide qualitative advice comparable to that of a clinician. Furthermore, bench testing of the predictive accuracy of the model has yielded encouraging results. We therefore set out to perform a preliminary retrospective medical validation of the physiological model using data collected by 30 insulin-dependent diabetic patients attending diabetes out-patient clinics at various centres throughout Europe. We found that the physiological model could only be parameterized for data from 24 (80%) of the 30 patients in the study. Comparison of observed and predicted blood glucose data from these 24 patients over a period of 5-6 days following parameter estimation revealed a mean (+/- SD) root mean square deviation between measured and simulated blood glucose values of 1.93 +/- 0.86 mmol l-1. The implications of these results are discussed.