Non-inferiority of retrospective data collection for assessing perioperative morbidity

Background. Postoperative morbidity has immediate and delayed consequences for surgical patients, including excess risk of premature death. Capturing these data objectively and routinely in large electronic databases using tools such as the Postoperative Morbidity Survey (POMS) would offer tremendous clinical and translational potential. However, POMS has thus far only utilised prospective data collection by research staff. We hypothesised that retrospective data collection from routinely collated hospital data from paper and electronic charts, medical and nursing notes was non-inferior to prospective data collection requiring research staff capturing POMS-defined morbidity in real-time.Methods. Morbidity was recorded by a trained investigator as defined by POMS prospectively on postoperative days 3 and 7. Separately, an independent investigator blinded to prospectively acquired data retrospectively assessed the same patients’ morbidity as defined by POMS criteria, using medical charts, nursing summaries and electronic data. Equivalence was accepted when the confidence limits for both modes of data collection fell completely inside the equivalence bounds, with the maximum equivalence difference (i.e., the largest value of the difference in sensitivities deemed to reach a conclusion of equivalence) set a priori at 0.2. Differences for confidence limits between retrospective and prospective data collection were based on Nam’s RMLE method. The relationship between morbidity on postoperative day 3 as recorded by each data collection method on time to become morbidity free and length of hospital stay was compared using the log-rank test.Results. POMS data from 85 patients undergoing elective or emergency surgery were analyzed. At postoperative day 3, POMS-defined morbidity was similar regardless of whether data were collected prospectively or retrospectively (95% CI [−0.13–0.013]; p < 0.001). Non-inferiority for sensitivity was observed for all other POMS domains and timepoints. Time to become morbidity free Kaplan–Meier plots were indistinguishable between POMS obtained prospectively or retrospectively (hazard ratio: 1.09 (95% CI [0.76–1.57]); p = 0.33, log rank test). Similarly, the mode of data collection did not alter the association between early postoperative morbidity on postoperative day 3 and delayed hospital discharge.Conclusions. Postoperative morbidity as defined by the Post Operative Morbidity Survey can be assessed retrospectively. These data may therefore be easily captured using electronic patient record systems, thereby expanding the potential for bioinformatics approaches to generate new clinical and translational insights into recovery from surgery