Development and validation of an interpretable Markov-embedded multilabel model for predicting risks of multiple postoperative complications among surgical inpatients: a multicenter prospective cohort study

Abstract

Background: When they encounter various highly related postoperative complications, existing risk evaluation tools that focus on single or any complications are inadequate in clinical practice. This seriously hinders complication management because of the lack of a quantitative basis. An interpretable multilabel model framework that predicts multiple complications simultaneously is urgently needed. Materials and Methods: The authors included 50 325 inpatients from a large multicenter cohort (2014–2017). The authors separated patients from one hospital for external validation and randomly split the remaining patients into training and internal validation sets. A MARKov-EmbeDded (MARKED) multilabel model was proposed, and three models were trained for comparison: binary relevance, a fully connected network (FULLNET), and a deep neural network. Performance was mainly evaluated using the area under the receiver operating characteristic curve (AUC). The authors interpreted the model using Shapley Additive Explanations. Complication-specific risk and risk source inference were provided at the individual level. Results: There were 26 292, 6574, and 17 459 inpatients in the training, internal validation, and external validation sets, respectively. For the external validation set, MARKED achieved the highest average AUC (0.818, 95% CI: 0.771–0.864) across eight outcomes [compared with binary relevance, 0.799 (0.748–0.849), FULLNET, 0.806 (0.756–0.856), and deep neural network, 0.815 (0.765–0.866)]. Specifically, the AUCs of MARKED were above 0.9 for cardiac complications [0.927 (0.894–0.960)], neurological complications [0.905 (0.870–0.941)], and mortality [0.902 (0.867–0.937)]. Serum albumin, surgical specialties, emergency case, American Society of Anesthesiologists score, age, and sex were the six most important preoperative variables. The interaction between complications contributed more than the preoperative variables, and formed a hierarchical chain of risk factors, mild complications, and severe complications. Conclusion: The authors demonstrated the advantage of MARKED in terms of performance and interpretability. The authors expect that the identification of high-risk patients and the inference of the risk source for specific complications will be valuable for clinical decision-making.