Dynamic prediction of early discharge after major pancreatic surgery: Derivation and cross-validation of an automatable electronic medical record-based model.

BACKGROUND

Length of hospital admission after major oncologic surgery is often highly variable. Although in carefully selected patients, early discharge can be safe, few automated systems exist to prospectively identify eligible patients. We aimed to develop and validate a predictive model to provide dynamic discharge predictions through each postoperative day.

METHODS

Electronic medical record data from the day of operation through to postoperative day 3 from adult patients who underwent elective pancreas resections between 2001 and 2021 were used. The final model used tabular prior data fitted networks. Early discharge was defined as length of stay <6 days with 90-day readmission as a counterbalance measure. Models were assessed via 10-fold cross validation in an 80% training and validation set and applied to a 20% hold-out test set.

RESULTS

A total of 3,081 consecutive patients (median age 64; 46.5% female) were included. All metrics improved as information accrued from postoperative day 0 to 3, with the tabular prior data fitted network performing best: area under the receiver operating characteristic curve of 0.90 (95% confidence interval 0.89-0.91), average precision of 0.80 (0.77-0.83), and Brier score of 0.12 (0.11-0.13) during cross-validation. The area under the receiver operating characteristic curve was 0.93 (95% confidence interval 0.91-0.94), average precision was 0.84 (0.80-0.89), and Brier score was 0.10 (0.08-0.11) in hold-out testing. Readmission rates were lower in those predicted suitable for early discharging (23.9% vs 34.2%).

CONCLUSION

This dynamic predictive model to predict early discharge after major oncologic surgery based on automatically abstractable electronic medical record data is now suitable for prospective evaluation.