Public Access Defibrillator Deployment for Cardiac Arrests: A Learn-Then-Optimize Approach with SHAP-based Interpretable Analytics

Out-of-hospital cardiac arrest (OHCA) survival rates remain extremely low due to challenges in the timely accessibility of medical devices. Therefore, effective deployment of automated external defibrillators (AED) can significantly increase survival rates. Precise and interpretable predictions of OHCA occurrences provide a solid foundation for efficient and robust AED deployment optimization. This study develops a novel learn-then-optimize approach, integrating three key components: a machine learning prediction model, SHAP-based interpretable analytics, and a SHAP-guided integer programming (SIP) model. The machine learning model is trained utilizing only geographic data as inputs to overcome data availability obstacles, and its strong predictive performance validates the feasibility of interpretation. Furthermore, the SHAP model elaborates on the contribution of each geographic feature to the OHCA occurrences. Finally, an integer programming model is formulated for optimizing AED deployment, incorporating SHAP-weighted OHCA densities. Various numerical experiments are conducted across different settings. Based on comparative and sensitive analysis, the optimization effect of our approach is verified and valuable insights are derived to provide substantial support for theoretical extension and practical implementation.

💡 Research Summary

This paper tackles the critical public‑health problem of low out‑of‑hospital cardiac arrest (OHCA) survival by proposing a novel “learn‑then‑optimize” framework that integrates machine learning, interpretable SHAP analytics, and integer programming for the optimal placement of automated external defibrillators (AEDs).
Data and preprocessing – The authors rely exclusively on publicly available geographic information from OpenStreetMap: 76 types of points‑of‑interest (POI) and 39 building categories for the city of Virginia Beach. OHCA incident records from Jan 2017 to Jun 2019 are aggregated onto Uber’s H3 hexagonal grid system (level 7, average edge length ≈ 1.41 km), a spatial resolution chosen because an emergency responder can traverse the grid radius within the “golden four minutes”.
Prediction model – A multilayer perceptron (MLP) regression is trained to map the counts of POIs and building types in each grid cell (Xᵢ) to the observed OHCA count (yᵢ). The model achieves an R² > 0.75 on a held‑out test set, demonstrating that geographic features alone can explain a substantial portion of the spatial variance in cardiac‑arrest risk.
Interpretability with SHAP – To open the black‑box model, the authors compute SHAP values for every feature. Because SHAP is additive, they distribute the SHAP contribution of a feature type across all individual buildings of that type within a grid (φₚ = φᵢⱼ · Xᵢⱼ / Xᵢⱼ). This yields a risk score for each building, highlighting high‑impact locations such as hospitals, senior‑housing complexes, and sports facilities. The SHAP analysis not only validates the model (e.g., positive contributions from health‑related POIs) but also provides decision‑makers with transparent, evidence‑based insights.
Optimization model (SIP) – Candidate AED sites (K) are randomly sampled among the POIs and buildings. For each candidate k, the authors define a coverage radius Aₖ (e.g., 500 m) and compute a SHAP‑weighted OHCA density Sₖ by summing the building‑level SHAP scores inside Aₖ, adjusted for overlapping grid areas. The integer programming formulation maximizes the total weighted coverage Σₖ Sₖ yₖ (yₖ ∈ {0,1}) subject to: (i) a minimum spacing constraint between any two deployed AEDs (to avoid redundancy) and (ii) a preset budget on the number of devices.
Experimental evaluation – The SIP model is tested under three deployment scales (10, 20, 30 AEDs) and three minimum‑spacing settings (200 m, 300 m, 400 m). Compared with a random‑placement baseline, SIP consistently yields higher predicted OHCA coverage (≈ 23 % increase) and higher estimated survival rates (≈ 15 % increase). Sensitivity analysis shows that spacing below 300 m markedly improves performance, while larger spacings lead to diminishing returns.
Strengths – (1) The approach works with only geographic data, making it applicable to regions lacking detailed demographic or historical OHCA records. (2) SHAP provides a clear, quantitative link between physical urban features and risk, bridging prediction and policy. (3) The integration of SHAP scores directly into the objective function creates a seamless, interpretable optimization pipeline.
Limitations – The linear allocation of SHAP values assumes equal influence among buildings of the same type, potentially overlooking complex interactions. Candidate sites are selected randomly, ignoring practical constraints such as power supply, accessibility, or legal permissions. The model is static; it does not adapt to real‑time incident streams or seasonal variations.
Future directions – The authors suggest extending the framework to multi‑objective optimization (cost, equity, maintenance), developing dynamic re‑allocation algorithms that ingest live EMS data, and conducting field pilots to validate the theoretical gains in real urban settings.
In summary, the paper delivers a compelling, data‑efficient, and interpretable methodology for AED deployment that could substantially improve OHCA outcomes, especially in data‑sparse environments, while also laying out clear pathways for further refinement and real‑world implementation.