Machine learning prediction model for postoperative outcome after perforated appendicitis

PURPOSE: Appendectomy for acute appendicitis is one of the most common operative procedures worldwide in both children and adults. In particular, complicated (perforated) cases show high variability in individual outcomes. Here, we developed and validated a machine learning prediction model for postoperative outcome of perforated appendicitis. METHODS: Retrospective analyses of patients with clinically and histologically verified perforated appendicitis over 10 years were performed. Demographic and surgical baseline characteristics were used as competing predictors of single-patient outcomes along multiple dimensions via a random forest classifier with stratified subsampling. To assess whether complications could be predicted in new, individual cases, the ensuing models were evaluated using a replicated 10-fold cross-validation. RESULTS: A total of 163 patients were included in the study. Sixty-four patients underwent laparoscopic surgery, whereas ninety-nine patients got a primary open procedure. Interval from admission to appendectomy was 9 ± 12 h and duration of the surgery was 74 ± 38 min. Forty-three patients needed intensive care treatment. Overall mortality was 0.6 % and morbidity rate was 15%. Severe complications as assessed by Clavien-Dindo > 3 were predictable in new cases with an accuracy of 68%. Need for ICU stay (> 24 h) could be predicted with an accuracy of 88%, whereas prolonged hospitalization (greater than 7–15 days) was predicted by the model with an accuracy of 76%. CONCLUSION: We demonstrate that complications following surgery, and in particular, health care system-related outcomes like intensive care treatment and extended hospitalization, may be well predicted at the individual level from demographic and surgical baseline characteristics through machine learning approaches. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s00423-022-02456-1.