TL;DR
This paper introduces a Bayesian Autoregressive (AR) model to improve the prediction of monthly outpatient visits, addressing the inherent uncertainties and irregular fluctuations often missed by traditional methods. By integrating Bayesian inference with an AR model and optimizing lag orders, the proposed framework demonstrates superior accuracy and stability in forecasting, offering a valuable tool for hospital management.
What Problem Does This Paper Address?
Hospitals are increasingly facing challenges in managing patient flow due to continuous population growth and the difficulty in accurately predicting outpatient visit volumes. Traditional statistical methods often struggle with the irregular fluctuations and unquantifiable factors influencing these numbers, leading to suboptimal resource allocation and potentially impacting patient satisfaction. While multivariate and machine learning approaches exist, this paper focuses on improving univariate time series forecasting by addressing the inherent uncertainty in historical data and model parameters, a crucial aspect for effective hospital management and operational planning.
Methods
The core of this research lies in a Bayesian Autoregressive (AR) model. Here's a breakdown of the key techniques:
- Autoregressive (AR) Model: This is a standard time series technique used to model the linear and autocorrelation relationships within historical data. It predicts future values based on a linear combination of past values.
- Bayesian Inference: This is the novel aspect. Instead of treating model parameters as fixed, Bayesian inference treats them as random variables with probability distributions. This allows for a more robust way to handle uncertainty.
- Gibbs Sampling: To overcome the computational complexity of Bayesian methods (especially for estimating posterior distributions), the researchers employed Gibbs sampling. This Markov Chain Monte Carlo (MCMC) method iteratively samples from conditional distributions to approximate the joint posterior distribution of the model parameters, providing stable and reliable estimates.
- Parameter Optimization: Bayesian inference is used to correct and optimize the AR model parameters, incorporating prior beliefs and updating them with observed data.
- Lag Order Adjustment: The paper emphasizes the importance of selecting the appropriate "lag order" (how many past time steps to consider) for the AR model. They used the Akaike Information Criterion (AIC) to determine the optimal lag order for the traditional AR model and explored a wider range of lag orders for the Bayesian AR model to improve fitting and accuracy.
- STL Decomposition: Seasonal and Trend decomposition using LOESS (STL) was used to break down the time series into seasonal, trend, and remainder components, helping to understand underlying patterns and explain model behavior.
Datasets
The study utilized historical monthly outpatient visit data from seven hospitals in China. The data was publicly available and sourced from the CNKI (China National Knowledge Infrastructure) website. Specific details about the data collection and criteria for selection are provided, ensuring transparency. No new benchmark datasets were introduced; rather, existing real-world data from multiple hospitals was used to evaluate the model's generalizability.
Key Insights and Ideas
- Uncertainty is Key: The paper strongly advocates for explicitly modeling uncertainty in outpatient visit predictions, arguing that traditional methods often overlook this crucial aspect, leading to less reliable forecasts.
- Bayesian Power: Bayesian inference, when applied to AR models, offers a powerful way to refine parameter estimates and account for uncertainty, leading to improved predictive performance.
- Lag Order Matters (A Lot): The choice of lag order is critical. The study found that increasing the lag order, especially for seasonal data, significantly improved the Bayesian AR model's ability to capture patterns and enhance prediction accuracy. This highlights that simply using a standard AR model with a default or low lag order might miss important historical context.
- Beyond Seasonality and Trend: The STL decomposition revealed that remainder components in outpatient visit data often exhibit complex dynamic patterns, suggesting that factors beyond simple seasonality and trends are at play, which a more sophisticated model like the Bayesian AR can better address.
Key Results
- Improved Accuracy: The Bayesian AR model consistently outperformed the traditional AR model across all seven hospitals.
- MAPE Reduction: The Mean Absolute Percentage Error (MAPE) saw significant reductions, ranging from 0.1431% to 0.0342% in the experimental data. For instance, in hospital H3 and H4, MAPE decreased by over 0.10%.
- Optimal Lag Orders: For the Bayesian AR model, optimal lag orders were found to be generally higher (between 12-16) compared to the traditional AR model, suggesting that more historical data is needed to capture the complex dynamics.
- Stable Predictions: The Bayesian AR model demonstrated stable prediction accuracy, especially when the training set proportion was between 70% and 80%.
Limitations and Future Work
- Unmeasured Confounding Factors: Like other univariate models, the Bayesian AR model's predictions can be influenced by unmeasured external factors (e.g., policy changes, public health events).
- AR Model Stability Check: The traditional AR model's stability check can limit the exploration of optimal lag orders. The authors addressed this by eliminating the stability check for the Bayesian AR model to explore higher lag orders.
- Nonlinear Dynamics: While the Bayesian AR model improved predictions, it still struggled with capturing long-term nonlinear information, particularly near the end of prediction horizons in some hospitals (H3-H5).
- Future Directions: The authors suggest incorporating deep learning models with long-term memory (e.g., LSTMs) in future research to better handle complex nonlinear dynamics and potentially improve prediction accuracy further.
Summary updated