Harnessing Machine Learning for Hospital Efficiency: A Study on Mortality and Length of Stay Predictions
In the ever-evolving landscape of healthcare, the ability to accurately predict patient discharge timing and in-hospital mortality is crucial for improving hospital efficiency. Traditionally, physicians have relied on their expertise to make these predictions, yet their estimates often lack consistency and accuracy. A recent study explores the potential of machine learning models to deliver more reliable predictions for both in-hospital mortality and length of stay (LoS) simultaneously.
Methods and Analysis
The study utilized electronic health records from Oxfordshire, UK, spanning from November 1, 2021, to October 31, 2024. By leveraging two years of data to train models and evaluating them on the last year’s data, the research aimed to assess the efficacy of various machine learning models. The models employed included task-specific Extreme Gradient Boosting (XGB), Logistic Regression (LR), Multilayer Perceptron (MLP), and TabNet, each designed to tackle two distinct tasks: mortality prediction and LoS prediction.
These models were compared to a single multiclass XGB model that predicted combinations of LoS and mortality, as well as an MLP and TabNet-based multitask learning model designed to predict both outcomes simultaneously. The predictions generated by the most effective models were then contrasted against the discharge predictions made by clinicians.
Results
The findings revealed that physicians provided relevant discharge prognoses for only 3–5% of admissions, primarily just before discharge. Task-specific XGB models demonstrated impressive performance, achieving an area under the receiver operating curve of 0.92 for mortality prediction and 0.83 for predicting LoS quartiles in elective admissions. For emergency admissions, the models reached an area of 0.72 for LoS predictions, outperforming task-specific LR, MLP, and TabNet models.
Interestingly, the multiclass XGB model and the multitask MLP or TabNet models did not consistently enhance performance. The best-performing task-specific XGB models matched the LoS prediction accuracy of physicians in elective admissions and significantly surpassed physicians in emergency admissions, with a statistical significance of p < 0.001.
Conclusion
Machine learning models have demonstrated their potential to predict in-hospital mortality and LoS as accurately, if not more so, than physicians, particularly in emergency departments. These models offer the promise of consistent predictions, which can improve discharge planning and hospital resource management. By integrating such technologies into hospital systems, healthcare providers can enhance operational efficiency and ultimately improve patient care outcomes.
For further reading, access the full study here.
“`
