A novel machine learning approach for spatiotemporal prediction of EMS events: A case study from Barranquilla, Colombia

Abstract

Anticipating the timing and location of future emergency calls is crucial for making informed decisions about vehicle location and relocation, ultimately reducing response times and enhancing service quality. A predictive model for EMS (Emergency Medical Services) events is proposed to address this need. The proposed spatiotemporal approach integrates machine learning, signal analysis, and statistical features, capturing geographical, temporal, and event-specific factors. The model identifies patterns associated with the occurrence or absence of emergency calls, using clustering techniques for demand spatial splitting and then training an XGBoost model on the multivariate time series. The model uses signal analysis to extract valuable insights from time-series data, enhancing understanding of temporal patterns, while statistical features enhance predictive capabilities. Principal Component Analysis (PCA) enhances convergence and integrates diverse time series features. As a result, this novel integrated approach improves the estimation of spatiotemporal probabilities of emergency events, effectively addressing data sparsity challenges. This framework adapts effectively, predicting EMS zones and guiding system configuration. The model outperforms a Random Forest trained solely on time-series data, boosting accuracy by up to 26.9 % in Barranquilla's case study zones, with a mean improvement of 16.4 %. Accuracy improvement makes the model helpful in assisting city authorities in ambulance location/relocation and dispatching decisions.

Keywords

• PCA (principal component analysis)
• Clustering
• XGboost
• Signal processing
• Statistical features
• Spatiotemporal classification