It is clear that in the healthcare system, medical transport plays a fundamental role in ensuring both care and the efficient planning and management of services. While its role in emergency situations stands out, it is equally necessary in the transfer of patients requiring specialized care. Achieving a correct operation is possible thanks to different factors, including an adequate infrastructure of personnel and resources, as well as effective coordination between the agents involved.
Hence, integrating data mining techniques into ambulance service planning, alongside insights from organ allocation studies, could offer a holistic framework for optimizing resource allocation, predicting service needs, and enhancing the overall efficiency of healthcare transport systems.
2. Health Transport Demand Prediction
Over the past few years, several research studies have been conducted with the aim of developing predictive models that can accurately estimate the demand for medical assistance in the ambulance setting [
10,
11]. These studies have used different methodological approaches and predictor variables to identify the factors that influence demand and provide a solid basis for the development of more accurate and reliable models.
One of the key issues addressed in the scientific literature is the consideration of temporal patterns of demand. Studies [
12,
13] have investigated the variability of demand throughout the day, week, or year, identifying specific demand patterns at different time points. These analyses reveal the existence of peak demand at certain times of the day or days of the week, as well as seasonal variations that may be influenced by external factors.
In addition to temporal patterns, geographic variables have been explored as another influential factor in the demand for ambulance care [
14,
15]. Studies have analyzed the influence of geographic location, highlighting that areas with a higher population density or with specific geographic characteristics, such as mountainous or rural areas, may have a higher demand for ambulance services. Furthermore, geographic variability in ambulance use is large and is associated with variations in the health status and socioeconomic situation of patients [
16].
The relationship between demographic variables and the demand for ambulance services has also been investigated [
17]. Previous studies have examined the impact of demographic variables, such as population, age, and gender, on the demand for ambulance services. These demographic factors may be associated with certain types of medical emergencies, which implies that their consideration is essential to develop accurate predictive models.
In terms of the methodologies used, the scientific literature has covered a wide spectrum of approaches. They have ranged from traditional statistical models, such as linear regression [
18] or time series [
19], to more advanced techniques, such as neural networks [
20], support vector machines, and machine learning algorithms [
21]. These approaches have proven effective in predicting care demand, with promising results in terms of accuracy and generalizability.
3. Predicting Ambulance Demand for Care Using Time-Series Techniques
In [
22], the importance of using time series prediction techniques for adequate health planning is highlighted. The cited study employs the Holt-Winters exponential smoothing model, a time series prediction technique, to detect seasonality patterns and demand evolution, allowing quality predictions in the short term. This method has advantages such as straightforward interpretation and implementation, as well as high reliability, being surpassed only by procedures that require a more complex and detailed comparative analysis. Therefore, this approach is recommended for routine use. Several adjustment measures are evaluated, such as RMSE, MAE, or MAPE, achieving short-term predictions with a MAPE of
5.9% at one week and
10.4% at three weeks.
In [
17], a modified clustering genetic algorithm is applied to compare optimal ambulance locations, predict future ambulance locations, and determine the required number of vehicles. The study predicts variations in care demand, in this case for emergencies, by reassigning the location of ambulances to other nearby ones. This reduces the average response time by 57 s. The importance of the age variable in considering the number of services is highlighted, using demographic predictions to infer future cases of long-term emergency health services.
A similar approach is explored in [
23], wherein neural networks incorporating predefined trajectories are used to predict the location of future demand for ambulance services. Leveraging these forecasts obtained, ambulances are relocated before actual emergencies occur.
These studies demonstrate the value of time series techniques, genetic algorithms, and neural networks in forecasting ambulance demand for care, allowing for efficient resource allocation and improved response times in emergency situations.
4. Fleet Management
Ambulance fleet management has been the subject of most of the studies reviewed in the literature, which have focused mainly on vehicle location and relocation, often sidelining the predictions regarding the number of services. A review of modeling approaches used in ambulance fleet management was conducted in [
24]. This examination considered factors such as objectives, coverage and location constraints, number of ambulances, and geographic region. Among the most commonly used techniques are branch and bound, branch and cut, heuristic methods, tabu search, the ant colony algorithm, and genetic algorithms. In addition, Bayesian approaches are noted to have been proposed for predicting the number of emergency calls in each area.
The Bayesian approach has proven to be an important feature in health management applications as it allows the combining of available data with prior information in a sound theoretical framework. In this way, subsequent inference can be used as prior information when new data become available, as reported in [
25]. For example, a Bayesian model has been used to estimate the distribution of ambulance travel times in different road segments of a city [
26], as well as to predict the demand of patients attended by the home care service [
27] and emergency calls [
28].
5. Spatio-Temporal Prediction
Another strand of research has centered on both temporal and spatio-temporal prediction of medical transport demand. To forecast the spatio-temporal demand for ambulances in Toronto, Ref. [
29] considered weekly seasonality, daily seasonality, and short-term serial dependence during some specific hours. Notably, they addressed the seasonality of the area without considering the exact routes. Prior to applying machine learning (ML) algorithms in their work, a prediction was formed using an averaging formula over a spatial region of 1 km² for a duration of one hour. Previous research has used different methods to predict aggregate ambulance demand as a temporal process, including autoregressive moving averages [
30], factor models considering hourly and daily seasonality [
31], spectral analysis [
19], and grid-based neural networks in discrete time and space [
32]. In the case of [
29], a discrete-time and continuous-space Gaussian model was adopted to predict emergency call volumes.
More recent research has used decision models, such as a hybrid decision tree using a naive Bayes classifier, to predict ambulance offload delay [
33]. In this case, the predictor variables used in the decision tree include the day of the week, time of day, call volume, free ambulance rate, and total number of ambulances.
Figure 1 shows the significance of these variables obtained after applying the hybrid decision tree with a naive Bayes classifier [
33].
Figure 1. Importance of variables obtained after applying a hybrid decision tree with a naive Bayes classifier to three different models. It is observed that, in all models, the most important variable is the number of ambulances at ED, followed by the hour of the day or the number of calls per hour, and the values provided by the National Emergency Department Overcrowding Scale (NEDOCS). Less important are the day of the week and the ambulance clear rate.
However, this is not the first time that ML techniques have been used to estimate ambulance demand. In [
18], these techniques were employed to quantify the characteristics that influence demand. Among these characteristics, age plays an important role, as a region is likely to experience higher demand due to a larger elderly population. For this purpose, actual patient data sets and demographic data from the past 10 years were used. Other relevant factors include the day of the week and month, as demand tends to have a periodic pattern, as well as short- and long-term historical demands in each region (e.g., outbreaks, sporting events). The variables considered in this study were classified into spatial (region ID), temporal (day of the week, day of the month, day of the year), demographic (number of people over 50 years old in a specific year), short-term historical demands (7 variables corresponding to the demands of the last 7 days in that region), and long-term historical aggregate demand (total demand for the last 30 days, last 7 days, week to sample date, month to sample date).
In addition, socioeconomic variables have been added in this research to analyze whether the demand for ambulances is related to the socioeconomic characteristics of the inhabitants of a region. For this purpose, methods such as regional moving average, linear regression, support vector regression (SVR), multilayer perceptron (MLP), radial basis function neural network (RBFN), and light gradient boosting machine (LightGBM) were used. After comparing the different methods, it was concluded that the best solution was LightGBM (regression tree). The most important characteristics were the ID of the region in which the demand was predicted, the demand of the previous 30 days, and the demand of the previous 7 days. The number of people over 50 years of age in the region was also considered important.