Ambulance demand prediction via convolutional neural networks

Maximiliane Rautenstrauß, Maximilian Schiffer

TL;DR

This paper introduces a novel Convolutional Neural Network (CNN) architecture for predicting ambulance demand by transforming time-series data into heatmaps. The approach effectively integrates diverse external features and optimizes model performance through a Bayesian Optimization framework that simultaneously handles hyperparameter tuning and feature selection, outperforming existing methods by over 9%.

What Problem Does This Paper Address?

Emergency medical services (EMS) are critical for timely patient care, but face challenges with limited resources and increasing demand due to factors like aging populations and climate change. Efficiently allocating ambulances requires accurate demand forecasts. While various machine learning models exist for this, they often struggle to capture complex spatio-temporal patterns and integrate external factors effectively. This paper aims to bridge this gap by proposing a CNN-based approach that can learn from both historical demand and a wide range of external features, leading to more robust and accurate predictions.

Methods

The core of this work is a novel Convolutional Neural Network (CNN) architecture designed to predict ambulance demand. Unlike traditional methods that might divide regions and forecast locally, this CNN uses three-dimensional convolutional layers to capture spatio-temporal correlations across an entire region simultaneously. It also ingeniously transforms time-series data into heatmaps for prediction. A key innovation is the flexible integration of external features of varying dimensions (e.g., weather, events, holidays) through concatenation at different stages of the network.

To optimize this complex model, the authors employ Bayesian Optimization (BO). This is not just for hyperparameter tuning but also for intrinsic feature selection. They treat feature inclusion/exclusion as hyperparameters, creating a high-dimensional search space. To manage this, they introduce two advanced BO techniques: BO with dimension dropout and a novel hierarchical BO. These methods efficiently explore the search space and identify optimal model configurations and relevant features.

Datasets

The study utilizes Seattle's 911 call data for ambulance demand prediction, specifically focusing on "Medic Response" incidents. This data covers a two-year period (2020-2021) and is divided into 8-hour intervals. External data sources include:

  • Weather Underground data: Temperature, wind speed, humidity, dew point, sea level pressure, and precipitation (daily min, max, and average).
  • Public holidays, school holidays, and events: Information on these special days, including event locations.

Key Insights and Ideas

  • Spatio-temporal heatmaps for demand: Transforming time-series ambulance demand into heatmaps allows the CNN to learn spatial and temporal patterns more effectively than traditional methods.
  • Flexible external feature integration: The CNN architecture's ability to incorporate diverse external features with varying dimensions is crucial for improving prediction accuracy, as these factors significantly influence demand.
  • Joint hyperparameter tuning and feature selection: Treating feature selection as a hyperparameter within the BO framework is a powerful approach. It allows the model to simultaneously optimize its structure and the features it uses, leading to more efficient and effective models.
  • Hierarchical and dropout BO for high-dimensional spaces: These advanced BO techniques are essential for navigating the complex search space created by numerous hyperparameters and features, leading to better optimization outcomes.
  • CNN robustness: The CNN demonstrates greater robustness to changes in upsampled and three-dimensional input data compared to MLPs and random forests, likely due to its ability to consider all subregions simultaneously.

Key Results

  • The developed CNN architecture outperforms existing state-of-the-art methods and industry practice (Medic method) by more than 9% in terms of Mean Squared Error (MSE).
  • The CNN achieves a 9.83% lower MSE than the best MLP, and significantly outperforms decision trees and random forests.
  • Incorporating feature selection reduces the number of trainable parameters by 40.4% with a negligible performance decrease (< 0.05%).
  • The CNN demonstrates superior performance across various time granularities (2-hour to 24-hour intervals), outperforming the Medic method in all cases.
  • SHAP value analysis indicates that the CNN's predictions are primarily influenced by data characterizing the predicted period, suggesting higher robustness to historical data variations compared to MLPs and random forests.

Limitations and Future Work

  • Feature selection impact: While feature selection significantly reduces model complexity, the paper notes that some features (e.g., month, school holidays) are neglected by all models, potentially because their information is implicitly learned from historical demand.
  • Time granularity optimization: The authors suggest further optimizing the CNN architecture for fine space and time granularities, possibly by assigning more weight to non-zero demand instances.
  • BO convergence: Investigating the convergence behavior of the BO when feature selection is included is identified as an interesting area for future research.
  • Generalizability: While the study uses Seattle's data, further validation on diverse geographical areas and EMS systems would be beneficial.

Summary updated

Abstract

Minimizing response times is crucial for emergency medical services to reduce patients' waiting times and to increase their survival rates. Many models exist to optimize operational tasks such as ambulance allocation and dispatching. Including accurate demand forecasts in such models can improve operational decision-making. Against this background, we present a novel convolutional neural network (CNN) architecture that transforms time series data into heatmaps to predict ambulance demand. Applying such predictions requires incorporating external features that influence ambulance demands. We contribute to the existing literature by providing a flexible, generic CNN architecture, allowing for the inclusion of external features with varying dimensions. Additionally, we provide a feature selection and hyperparameter optimization framework utilizing Bayesian optimization. We integrate historical ambulance demand and external information such as weather, events, holidays, and time. To show the superiority of the developed CNN architecture over existing approaches, we conduct a case study for Seattle's 911 call data and include external information. We show that the developed CNN architecture outperforms existing state-of-the-art methods and industry practice by more than 9%.
BibTeX
@misc{rautenstrau2023ambulance,
  author = {Maximiliane Rautenstrauß and Maximilian Schiffer},
  title = {{Ambulance demand prediction via convolutional neural networks}},
  year = {2023},
  eprint = {2306.04994},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.04994},
}
Figure