TL;DR
This review surveys the application of machine learning to optimize patient flow in hospitals, breaking down the complex problem into four key areas: predicting demand, managing transfers, optimizing inpatient care, and forecasting length-of-stay/discharge. The authors highlight the need for standardized datasets and more generalizable, dynamic ML models to truly impact real-world clinical decision-making.
What Problem Does This Paper Address?
Hospitals are facing increasing strain due to aging populations and the lingering effects of the COVID-19 pandemic, leading to inefficiencies in patient flow and longer wait times. Traditional management techniques have proven insufficient. While electronic health records (EHRs) provide a wealth of data, applying machine learning to improve patient flow across different hospitals remains a challenge due to a lack of standardized data, varying methodologies, and models that are often specific to a single institution. This paper aims to provide a structured overview of the current research landscape, identify gaps, and suggest future directions for machine learning in this critical area.
Methods
This paper is a review, not a presentation of novel methods. It categorizes existing research into four main areas of patient flow:
- Prediction of Demand: Forecasting patient admissions to the hospital, including emergency and elective admissions, and predicting readmissions.
- Emergency-to-Inpatient Interface: Predicting which patients from the Emergency Department (ED) will require hospital admission and where they will be admitted.
- Intra-hospital Management: Predicting inpatient transfers (e.g., to the ICU), resource utilization, and forecasting resource needs.
- Discharge Prediction: Estimating patient length-of-stay (LOS) and predicting readiness for discharge.
The review discusses various machine learning techniques employed across these areas, including regression, classification, decision trees, Naive Bayes, gradient boosting machines, recurrent neural networks (RNNs), deep neural networks, and reinforcement learning.
Datasets
The paper emphasizes the lack of standardized, publicly available datasets for patient flow research. Most studies utilize data from the specific hospitals where the research is conducted. The MIMIC-III dataset is mentioned as a publicly available benchmark, but its focus on intensive care patients limits its direct applicability to broader patient flow problems. The authors strongly advocate for the creation of a shared, public dataset with standardized definitions for patient types, co-morbidities, and pre-defined tasks to enable better benchmarking and model comparison.
Key Insights and Ideas
- Holistic vs. Granular Approach: The authors suggest that optimizing patient flow benefits from considering both the hospital as a whole system and individual patient journeys, ideally through a combination of approaches.
- The "Bed-Blocker" Problem: A significant challenge in inpatient settings is patients who are medically ready for discharge but lack appropriate destinations, leading to prolonged hospital stays and reduced capacity.
- Data Standardization is Crucial: The lack of standardized data definitions (e.g., for "elderly" patients or "long-stay" patients) and consistent data collection across institutions hinders the development of generalizable models.
- ED vs. Inpatient Focus: Research is more heavily concentrated on ED-related predictions (admissions, ED-to-inpatient transfers) compared to the inpatient journey, likely due to the greater unpredictability of emergency admissions.
- The Need for Dynamic Models: Future models should be dynamic, capable of real-time decision-making to assist clinical staff on the "shop floor."
Key Results
As this is a review paper, it does not present new experimental results. Instead, it synthesizes findings from numerous studies. The key takeaway is that machine learning has demonstrated feasibility across all four subcategories of patient flow. However, the field is still nascent, with many studies being institution-specific and lacking direct comparison or validation across different healthcare settings.
Limitations and Future Work
Limitations: * Lack of Generalizability: Most models are trained and validated on data from a single hospital, limiting their applicability elsewhere. * Data Challenges: Patient privacy concerns, lack of standardized data recording, and the dynamic nature of healthcare data (covariate shift) pose significant hurdles. * Incomplete Information: Many prediction algorithms provide limited information beyond a basic prediction, hindering comprehensive clinical planning. * Deployment Challenges: Issues like model generalizability and continuous use over time are significant challenges in real-world deployment.
Future Work: * Standardized Datasets: Creation of a public, comprehensive dataset with strict definitions for patient types, co-morbidities, and pre-defined tasks. * Multitask Learning: Developing single models that can predict multiple related outcomes simultaneously to provide a more holistic patient view. * Transfer Learning: Leveraging knowledge from models trained on large datasets (even non-specific ones) to improve performance on smaller or specialized datasets (e.g., for rare diseases or specific patient cohorts). * Continual Learning: Enabling models to learn and adapt over time as new data becomes available, accommodating changes in healthcare contexts, patient populations, and emerging diseases. * Feature Engineering: Incorporating novel data sources like wearable sensors for more granular patient information. * Dynamic, Real-time Models: Developing algorithms that can support real-time decision-making for clinical staff.
Summary updated