This study aimed to develop machine learning and deep neural network based-models, which could predict the LOS for a severely at-risk patient who had undergone an operation with general anesthesia. We demonstrated models using only preoperative features to predict postoperative LOS. We compared ridge regression, XGBoost, multi-layered perceptron neural networks, and random forest models. The XGBoost model showed the outperformance result on LOS prediction evaluated on the RMSE score. Our framework also established that model learning can be used concurrently with the model’s explainable techniques to support a deeper understanding of the risk predictors involved in determining the LOS.
Our study demonstrates several main findings. Consisting of 422 preoperative predictive features among 67,077 observations, our dataset deduced that the XGBoost was the best-performing model. Moreover, XGBoost predicted that following an operation using general anesthesia, severe patients spend an average of 3.56 days at the hospital after the operation. The findings encourage the use of XGBoost tree models for forecasting the duration of severe patients undergoing these operations. In addition, the population distribution plot has shown that the length of stay for these severe patients is significantly crowded between 3 to 7 days following surgery (Fig. 4). Thus, using this model, the hospital management could adapt early and efficient resources allocations that would allow them to adequately equip for crowding and financial burdens. Moreover, the liver transplant surgery departments consist of the longest LOSs following surgery (Fig. 5). Therefore, they should be supplied with additional support to systematically and physically assist in the outcome of post-operative treatments. These predictions will ultimately lead to shortening the overall duration outcomes.
In healthcare, decision support systems are significantly important to both the providers and the patients38,39. Effectively reducing time management by supporting the decision-making for clinicians will eventually provide high-quality care and drive the right clinical outcomes for the patients40,41. Therefore, implementing an accurate but deeper analysis system of interpreting the risk predictors to improve the decision-making prior to an event plays a crucial role in anticipating the patient’s stay. In this respect, in agreement with Bertsimas’s findings42, we demonstrate that by adding explainability to machine learning results the analysis in predicting risk factors for delays in discharge can be further enhanced. To accomplish this, SHAP analyzes the model learning more accurately and consistently through global and local explanation approaches to further investigate our model.
Initially, the global explanation approach presented the top twenty preoperative predictors contributing to forecasting the LOS outcome for critical individuals (Fig. 6). These provide sound evidence of the SHAP technique being able to sufficiently provide clinical importance through the use of the XGBoost model. Among the top twenty most influential features, ten emerged from the medication data points relating to the severe patients' prescription prior to the surgery being predominantly associated with the subsequent LOS. Moreover, this could lead to the physicians taking a high level of selected features as risk factors prior to arranging the pre-surgery procedures. We found similar key markers for predicting the LOS as previously detailed by Iwase and co-authors’, where albumin and direct bilirubin were associated with the LOS for critically ill patients7. Especially, the observed magnitude of direct bilirubin test attribution, with lower than 0.1 units (Fig. 7-a), alongside the removal of other organs diagnoses indicates an association. Furthermore, these defined predictors will reduce the discrepancy in selecting the irrelevant features pre-surgery providing only the essential resources to the physicians. Thus, to effectively reduce the hospitalization days, we suggest four-fold implications: Firstly, a patient’s contribution to avoiding the pre-surgery medication intake according to the lists presented by our findings; secondly, physicians attentively devote more attention to specific blood laboratory results paired with the patient’s prescribed medication (Fig. 7); thirdly, the hospital operational team strategically manages patients receiving treatment at certain departments and avoids additional, impractical departmental changes, particularly for females (Fig. 7-d).
Furthermore, the knowledge gained from the local explanation approach (Fig. 8), affirms that patients may also progress into hospital management. Thus, results in financial improvements since the discovered features for each patient can detail the contribution from the whole model output. Primarily, the operations management team in a hospital may take the findings of these positively contributing preoperative predictors from patients, which relate to true positive prediction performance and accurately predict the stay (Fig. 8-c). To a greater extent, these patient-inspired findings present an opportunity to perform patient-specific care, whereby individuals are encouraged to engage with their medication to reduce unnecessary healthcare costs and inefficient clinical trials, to ultimately shorten their postoperative stays43. Nevertheless, because the most impactful predictors display variances between the patient observations, the experiment should be conducted globally to further refine the cohorts.
Using the SHAP interpretation, our study achieved better decision-making from patients' visits by detecting the risk factors coupled with the predictors. Overall, based on our feature identification, the LOS could be more highly and accurately designated at an earlier stage of the treatment process (Figs. 6, 7). In summary, an accurate analysis of the importance and contribution of the XGBoost model’s preoperative predictors to the operative LOS will both support the facilitation of the operation department and provide efficient resource allocation toward advancing overall hospital management.
Previous studies have developed algorithms for predicting the LOS focused on disease-specific surgeries44–47. Predicting the risk factors of critically ill patients is significant since a prolonged stay for patients can increase the risk of hospital-acquired infections and hinder other patients’ access to the operation and medical resources48. A longer LOS is reported to be related to the illness severity5,49. Additionally, Naessens proposed higher-risk populations are likely to incorporate considerably more resources50. Hence, we focused on cohorts with critical patients, narrowing the focus to patients who had undergone general anesthesia. Consequently, our study offers insightful clinical suggestions to the operation’s department as well as to patients experiencing a postoperative crisis due to an extended LOS.
Yet, our study contains several limitations. Firstly, only a single-centered data analysis was used, which limits any further validation from external resources. Future works should consider taking the external validation development from multiple sites to enhance the model’s predictive performance. Secondly, no socio-economic and behavioral data were included in the study, which could have impacted postoperative LOS. Further study is suggested to involve socio-economic, genetic testing, and behavioral data, which can potentially affect a patient's recovery outcome. Such data can increase the precision in predicting the stay duration tailored to individual patients’ circumstances51,52.