Optimizing Healthcare Resilience: Advanced Machine Learning for Predicting Patient Length of Stay

In line with the ongoing global focus and demand for efficient healthcare and resilient operations, this study emphasizes the criticality of predictive analytics in healthcare service delivery. As ‘prevention is better than cure,’ we employed retrospective data to anticipate disruptions in healthcare systems. Our primary indicator, patient length of stay (LOS), serves as a pivotal metric for resource allocation in hospital settings, especially in densely populated regions. Accurate LOS predictions are crucial for optimizing the use of hospital beds, medical supplies, and physician teams. We utilized the MIMIC III real-world dataset, undertaking extensive preprocessing to enhance the predictive capabilities of our machine learning (ML) models. Feature engineering revealed key factors influencing patient hospital stays. Our analysis compared the efficacy of linear and non-parametric ML algorithms, applied on clinical and demographic patient data. The results demonstrated the superior performance of non-parametric models, notably the random forest (RF) and decision tree (DT) algorithms, which yielded a low mean squared error (1.34) and a high R-squared score (0.91). These models outperformed the multi-variable linear regression (MVLR) and support vector regression (SVR) models, as well as baseline measures like average (7.23) and median LOS (6.37). Our findings have significant implications for healthcare service operations, offering a robust tool for decision-makers in resource allocation and bed management. This proactive approach can prevent emergency situations, contributing to more resilient healthcare operations. Furthermore, the insights gained from digital data analysis can enhance service operations in healthcare.