P680 Machine learning (ML) in RCU-operated patients: Can we predict postoperative complications?

Abstract

Background: The study analyses short-term outcomes of laparoscopic and open total abdominal colectomy (TAC) for medically refractory ulcerative colitis (UC). A computer program based on Machine Learning (ML) was developed to predict postoperative complications. Methods: 32 patients who underwent TAC either laparoscopic or laparotomic for UC between December 2010 and October 2017 were retrospectively identified from an hospital database. Biographical data, preoperative therapy (steroids and biological drugs), nutritional status, surgical technique, operative time, blood transfusion, hospital length of stay, morbidity and mortality were recorded. Univariate analysis by unpaired Student t-test and multivariate analysis were conducted. A feature selection was performed to use selected data as input of the ML algorithm. Results: A total of 32 patients underwent TAC, 24 by laparoscopic (75%) and 8 (25%) by open approach. 17 patients presented an acute severe UC not responding to rescue therapy requiring urgent or semiurgent colectomy: 70.5% treated by laparoscopic and 29.5% by open approach. 15 patients presented a chronic refractory UC in treatment with steroids: 86% treated by laparoscopic and 14% by open approach. Operative time was higher in the laparoscopic compared with open group (267.5 ± 27 vs. 193 ± 37.5 min, p < 0.01). Progressive decrease of operative time was observed in laparoscopic patients during recent years due to improvement of skills (220 min in last year). Infectious as well as no infectious complications observed in laparoscopic group were lower when compared with open group (respectively 10% vs. 50%, p < 0.05 and 15% vs. 58%, p < 0.05). Postoperative length of stay was shorter in laparoscopic than open group (11.25 ± 4.97 vs. 15.9 ± 7.69 days, p = NS). Progressive decrease of hospital length of stay was observed in laparoscopic group (7 days in last year). ML of evaluation analysing age, sex, surgical technique, nutritional status, preoperative therapy with steroids/biological drugs and blood transfusion has predicted infectious minor complications with high strike rate (84%), high sensitivity (87.5%), and high specificity (83%). Conclusions: Data confirm that laparoscopic technique is a safe and valid approach in the surgical treatment of UC also in urgent setting with patient in critical conditions, reducing the global morbidity and hospital stay. ML has been proven to be useful in predicting rate of postoperative complications in patients operated for UC, despite of a small group of patients. ML used to wider samples or multi-centre experiences could became a valid help to choose the best treatment in order to avoid dangerous or simply annoying complications in RCU patients.