Ventilation was measured in 31 difficult-to-wean patients while pressure support (PS) was reduced by 5 cm H2O every 20 min. Weaning had to be aborted in 14 of 31 patients (Group F) because they met predefined distress criteria. The remaining 17 patients who were able to complete the "weaning test" (Group S) had larger static respiratory compliances (Cstat = 0.08 +/- 0.02 versus 0.05 +/- 0.01 L/cm H2O, p < or = 0.05) and a lower dead space to tidal volume ratio (0.55 +/- 0.05 versus 0.64 +/- 0.06, p < or = 0.05). Group S patients had a larger tidal volume (VT) than did those of Group F at most PS settings. The groups differed with respect to VT maintenance during PS withdrawal (p < 0.01). In Group S, VT fell exponentially with machine support and stabilized at PS levels between 5 and 10 cm H2O. In contrast, Group F patients defended VT at higher PS settings but were unable to maintain VT during distress. Ventilatory response parameters such as the rapid shallow breathing index were of limited value in predicting weaning outcome and yielded receiver operator curve area values between 0.66 and 0.82 over the range of PS settings tested. We conclude that the gradual withdrawal of machine support does not facilitate the recognition of impending respiratory failure.