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

This study proposes a novel learn-then-optimize framework that leverages geographic data and SHAP-based interpretability to guide an integer programming model for the strategic deployment of public access defibrillators, thereby enhancing out-of-hospital cardiac arrest survival rates.

The Gist

Imagine a city as a giant, living organism. Sometimes, this organism suffers a sudden, life-threatening "heart attack" right in the middle of the street. This is called an Out-of-Hospital Cardiac Arrest (OHCA). When this happens, every second counts. If a person doesn't get a shock from a defibrillator (a device called an AED) within four minutes, their chances of survival drop dramatically.

The problem? AEDs are like fire extinguishers; they are useless if they are locked in a basement three blocks away when the fire starts. Many cities have AEDs, but they are often placed randomly or in the wrong spots, leaving dangerous "blind spots" where help arrives too late.

This paper presents a smart, three-step recipe to solve this problem: Predict, Explain, and Place.

Step 1: The Crystal Ball (Prediction)

Traditionally, to find where heart attacks happen, experts needed detailed, hard-to-get data like census records (who lives where, their age, income) or a perfect history of every past heart attack. But what if a city doesn't have those records?

The authors built a digital crystal ball (a Machine Learning model) that doesn't need those complicated records. Instead, it looks at the city's "skeleton":

• Where are the buildings? (Apartments, schools, offices).
• Where are the points of interest? (Restaurants, parks, gas stations).

The Analogy: Think of it like predicting where a traffic jam will happen. You don't need to know the name and age of every driver. You just need to know that "rush hour" happens near "office parks" and "high schools." Similarly, the model learned that heart attacks are more likely to happen in areas with lots of apartments (where many people live) and less likely in retail stores or graveyards.

They tested this on a city in Virginia and found their "crystal ball" was surprisingly accurate, guessing the right spots over 75% of the time using only building maps.

Step 2: The Translator (Interpretation)

Usually, computer models are "black boxes." They give an answer, but you don't know why. In life-or-death situations, we need to trust the answer.

The authors used a tool called SHAP (which sounds like a friendly robot) to act as a translator. It opened the black box and said:

• "Hey, the model thinks this neighborhood is high-risk because there are 500 apartments here."
• "It thinks this park is low-risk because it's mostly open space."

The Analogy: Imagine a doctor diagnosing a patient. Instead of just saying "You have a fever," the doctor explains, "You have a fever because of this specific infection." SHAP does this for the city map. It tells city planners exactly which buildings are driving the risk, so they know exactly where to focus.

Step 3: The Chess Master (Optimization)

Now that we know where the risk is and why, we need to place the AEDs. But we can't just put one on every corner; they are expensive, and we need to space them out so they don't overlap too much.

The authors created a mathematical chess master (an Integer Programming model).

• The Goal: Place a specific number of AEDs (say, 100) to cover the most "at-risk" buildings.
• The Rule: No two AEDs can be too close together (to avoid waste), but they must be close enough to reach a victim in 4 minutes.

The Analogy: Imagine you are placing sprinklers in a garden to water the wildest, driest flowers. You don't want two sprinklers watering the same patch of grass (waste), but you also don't want a dry patch in the middle. The "Chess Master" calculates the perfect layout so that every drop of water (or every AED) counts.

The Results: Why It Matters

The team tested their method against "random placement" (just throwing darts at a map to decide where to put AEDs).

• Random placement missed a lot of heart attacks.
• Their "Learn-then-Optimize" method covered significantly more ground.
• The Magic Number: They found that spacing AEDs about 1.2 kilometers (0.75 miles) apart was the "sweet spot." This is the distance a person can run in 4 minutes.

The Bottom Line:
This paper shows that you don't need perfect, expensive data to save lives. By looking at simple maps of buildings and using smart math, cities can figure out exactly where to put life-saving machines. It turns a chaotic guessing game into a precise strategy, potentially saving hundreds of lives by ensuring help is always just a 4-minute run away.

Technical Summary

Here is a detailed technical summary of the paper "Public Access Defibrillator Deployment for Cardiac Arrests: A Learn-Then-Optimize Approach with SHAP-Based Interpretable Analytics."

1. Problem Statement

Out-of-Hospital Cardiac Arrest (OHCA) is a leading cause of death globally, with survival rates as low as 1.2% due to delays in medical intervention. The "golden four minutes" window for effective treatment using Automated External Defibrillators (AEDs) is frequently missed because AEDs are not optimally located relative to where cardiac arrests occur.

Existing strategies for identifying high-risk OHCA areas and optimizing AED deployment face two major limitations:

1. Data Scarcity: Traditional methods rely heavily on demographic data (e.g., census data) or historical OHCA records, which are often unavailable, outdated, or difficult to collect in many regions.
2. Lack of Interpretability: While machine learning (ML) models can predict risk, "black box" models do not explain why certain areas are high-risk, making it difficult for policymakers to trust and act on the recommendations.

The paper aims to develop a framework that predicts OHCA risk using only easily accessible geographic data (Point-of-Interest and building distribution) and uses interpretable analytics to guide an optimization model for AED placement.

---

2. Methodology: The "Learn-Then-Optimize" Framework

The authors propose a three-stage framework integrating Machine Learning, Explainable AI (XAI), and Operations Research.

Stage 1: Machine Learning Prediction (The "Learn" Phase)

• Input Data: Instead of demographics, the model uses geographic features extracted from OpenStreetMap (OSM), specifically:
  • POI Types: 76 categories (e.g., restaurants, community centers).
  • Building Types: 39 categories (e.g., apartments, schools, hotels).
• Spatial Unit: The study area (Virginia Beach) is divided into H3 Level 7 hexagonal grids (average edge length 1.41 km). This size was chosen because a responder can run from the center to the edge in ~4 minutes.
• Model: A Multilayer Perceptron (MLP) (Neural Network) is trained to predict OHCA density ($y_i$) based on the count of geographic features ($X_i$) within each grid.
  • Goal: To validate that geographic features alone can serve as proxies for demographic risk factors.

Stage 2: SHAP-Based Interpretable Analytics

• Technique: SHAP (Shapley Additive Explanations) values are calculated for the trained MLP.
• Function: SHAP quantifies the contribution of each specific geographic feature (e.g., number of apartments) to the predicted OHCA risk in a specific grid.
• Derivation:
  • SHAP values are aggregated to determine the "risk score" of individual buildings/POIs.
  • A key hypothesis is applied: The SHAP value of a feature type in a grid is distributed equally among all instances of that feature type within the grid.
  • This transforms grid-level predictions into building-level risk scores, creating a granular map of OHCA risk density.

Stage 3: SHAP-Guided Integer Programming (SIP) (The "Optimize" Phase)

• Objective: Maximize the total SHAP-weighted OHCA density covered by deployed AEDs.
• Model Formulation: A Binary Integer Programming model is constructed with:
  • Decision Variables: $z_k \in \{0, 1\}$ (Select candidate location $k$ or not).
  • Objective Function: $\max \sum_{k \in K} S_k \cdot z_k$, where $S_k$ is the SHAP-weighted density around candidate $k$.
  • Constraints:
    1. Minimum Spacing ($D_{min}$): Ensures AEDs are not placed too close to each other (avoiding redundancy).
    2. Budget ($N$): Limits the total number of AEDs to be deployed.
• Evaluation Metrics:
  • Historical Coverage: Number of past OHCA incidents within the coverage radius ($CR = 0.96$ km) of at least one deployed AED.
  • Survival Rate: Estimated using a logistic function based on the simulated time to AED delivery ($t_{AED}$) and CPR time.

---

3. Key Contributions

1. Data Efficiency: Demonstrated that geographic data (POIs/Buildings) alone is sufficient to predict OHCA risk with high accuracy ($R^2 > 0.75$), bypassing the need for scarce demographic or historical medical data.
2. Interpretability: Successfully bridged the gap between ML prediction and decision-making by using SHAP to identify specific building types (e.g., apartments) that drive risk, providing actionable insights for urban planners.
3. Optimization Framework: Developed a SHAP-guided Integer Programming (SIP) model that translates interpretability insights into a mathematically optimal deployment strategy.
4. Robust Validation: Validated the approach through extensive numerical experiments, sensitivity analysis, and comparison against random baselines.

---

4. Experimental Results

The study was conducted using data from Virginia Beach, USA (2017–2019).

• Prediction Performance:
  • The Neural Network achieved an $R^2$ of 0.975 on the training set and 0.752 on the test set, confirming a strong correlation between geographic features and OHCA events.
• Interpretability Insights:
  • Positive Correlation: High density of apartments and residential buildings significantly increases OHCA risk (aligning with population density theories).
  • Negative Correlation: High density of retail and graveyards correlates with lower risk.
• Optimization Performance (SIP vs. Random Baseline):
  • Coverage: The SIP model significantly outperformed random deployment. At $N=100$ AEDs with optimal spacing ($D_{min}=1.2$ km), SIP covered 1,388 cases (92.4%), a 27% improvement over the random baseline.
  • Survival Rate: SIP improved the average survival rate by up to 48% (at low $N$) and maintained a 16% minimum improvement at $N=100$ compared to random placement.
  • Robustness: The SIP model showed lower standard deviation across multiple random candidate sets, indicating stable performance.
• Sensitivity Analysis:
  • Optimal Spacing: The best performance occurred at a minimum spacing ($D_{min}$) of 1.2 km.
    • Too small (<1.2 km): AEDs overlap, wasting resources.
    • Too large (>1.2 km): Gaps appear in high-risk coverage.
  • Saturation Point: Deploying beyond 100 AEDs yielded diminishing returns, as most high-risk areas were already covered.

---

5. Significance and Future Work

• Practical Impact: The framework provides a scalable, cost-effective solution for cities lacking detailed demographic data to optimize emergency resource allocation. It moves AED deployment from "random" or "heuristic" to "data-driven and interpretable."
• Policy Relevance: The identification of specific building types (e.g., apartments) as high-risk zones gives city planners concrete targets for infrastructure investment.
• Generalizability: The "Learn-Then-Optimize" approach is not limited to AEDs; it can be adapted for other medical resource distributions (e.g., vaccines, ambulances).
• Limitations & Future Directions:
  • Heterogeneity: The correlation between geography and OHCA may vary across different cities due to cultural or urban planning differences.
  • Future Work: The authors propose validating the model across diverse global cities and developing benchmarks to determine when a model needs retraining for a new region.

In conclusion, this paper presents a rigorous, interpretable, and highly effective methodology for saving lives by strategically placing AEDs based on the built environment, overcoming significant data barriers in public health emergency planning.