A new automatic external defibrillator was tested first against a tape-recorded data base of rhythms and then during use by first-responding fire fighters in a tiered emergency system. The sensitivity for correctly classifying ventricular fibrillation and ventricular tachycardia was substantially less during clinical testing in 298 patients than would have been predicted from preclinical results: 52% of ventricular fibrillation analyses in patients were correctly classified versus 88% of episodes in the data base, and 22 versus 86%, respectively, for ventricular tachycardia (p less than 0.001). The detection algorithm was modified and evaluated further in another 322 patients. The modified detector performed substantially better than did the one that had been designed from prerecorded rhythms: with its use, 118 (94%) of 125 patients in ventricular fibrillation were counter-shocked compared with 91 (77%) of 118 similar patients with use of the initial algorithm (p less than 0.001). No inappropriate shocks were delivered. This improvement resulted in a shorter time to first shock (p less than 0.01) and more shocks being delivered for persistent or recurrent episodes of ventricular fibrillation (p less than 0.05). Of 620 patients treated with the automatic defibrillator, 243 (39%) had ventricular fibrillation; 57 (23%) of the 243 regained pulse and blood pressure before paramedics arrived, 141 (58%) were admitted to hospital and 71 (29%) were discharged.(ABSTRACT TRUNCATED AT 250 WORDS)