🔢 mathematics

Selective Ambulance Dispatch Under Contextual Travel-Time Uncertainty

This paper proposes IDEAL, a selective dual-dispatch framework that leverages a weakly supervised bilevel representation network to learn context-specific travel-time uncertainties and dynamically dispatch a second ambulance only when the optimistic gap between primary and secondary paths exceeds a threshold, thereby optimizing the trade-off between response time and fleet resource utilization for out-of-hospital cardiac arrest cases.

Original authors: Zikun Lin, Daniel Zhuoyu Long, Viet Anh Nguyen

Published 2026-05-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zikun Lin, Daniel Zhuoyu Long, Viet Anh Nguyen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Second Car" Dilemma

Imagine you are the dispatcher for an ambulance service. A heart attack happens. You need to get help there as fast as possible. You have two main problems:

Traffic is a liar: You can't predict exactly how long a trip will take because traffic jams, accidents, and road closures happen suddenly.
You have limited cars: If you send two ambulances to the same spot, you have one fewer car available for the next emergency.

Usually, dispatchers follow one of two rules:

The "Neighborhood" Rule: "This area belongs to Depot A, so send the car from Depot A." This is simple, but if there's a massive traffic jam on Depot A's usual route, the car gets stuck.
The "Always Send Two" Rule: "Send two cars just in case." This is safe, but it wastes resources. If the first car arrives in 5 minutes, the second one was just sitting there doing nothing, unable to help someone else.

The IDEAL solution is a smart middle ground. It asks: "Is there a realistic chance that a second car, taking a different route, could beat the first car by a significant margin?" If the answer is yes, it sends the second car. If the answer is no (or the second car is just a backup that will likely arrive at the same time), it keeps the second car on standby for a different emergency.

How IDEAL Works: The Three-Step Brain

The paper describes IDEAL as a three-step thinking process that happens in milliseconds when a call comes in.

1. Learning the "Ghost Map" (Weakly Supervised Learning)

The Analogy: Imagine trying to learn the best routes through a city, but you don't have a map of the roads the drivers actually took. You only know the start time, the end time, and the total time it took. You don't know which streets they drove down.

What IDEAL does: It looks at thousands of past ambulance trips. Even though it doesn't know the exact path taken, it uses a special math trick (called bilevel learning) to guess the speed of every single street segment. It essentially says, "If the total trip took 10 minutes, and the car started here and ended there, the streets in between must have been moving at this average speed." It builds a "ghost map" of how fast roads usually are at specific times and weather conditions.

2. Drawing the "Uncertainty Bubble" (Contextual Scenario Set)

The Analogy: Imagine you are betting on a race. You don't just guess one time; you draw a "bubble" of possibilities.

Small Bubble: "It's a sunny Tuesday morning; traffic is predictable. The car will arrive between 4 and 5 minutes."
Huge Bubble: "It's a rainy Friday evening with a known accident nearby. The car could arrive in 3 minutes or 15 minutes."

What IDEAL does: It doesn't just guess one travel time. It creates a "bubble" of possible scenarios. It uses a mathematical tool called Burg divergence (think of it as a "distance meter" for uncertainty) to figure out how big that bubble should be.

If the history shows that predictions are usually wrong during rain, the bubble gets bigger.
Crucially, it understands that traffic is correlated. If one road is jammed, the road right next to it is probably jammed too. IDEAL simulates these "bubbles" so that if the first road slows down, the second road slows down in the same way, rather than treating them as totally random.

3. The "Optimistic Gap" Check (Difference-of-Convex Programming)

The Analogy: You have a primary runner (Ambulance A) and a backup runner (Ambulance B). You want to know: "In the worst-case scenario for Runner A, is there a realistic scenario where Runner B could beat Runner A by, say, 2 minutes?"

What IDEAL does:

It takes the primary route (the one the system usually picks).
It looks inside its "Uncertainty Bubble" to find the worst-case scenario for that primary route (e.g., "What if every light turns red?").
It then looks for the best-case scenario for a different route (e.g., "What if the backup car takes a side street that stays clear?").
It calculates the gap: How much faster could the backup car be in that best-case scenario compared to the primary car in its worst-case?

The Decision:

If the gap is small (e.g., 10 seconds), IDEAL says, "Not worth it. Keep the second car on standby."
If the gap is large (e.g., 2 minutes), IDEAL says, "Send the second car immediately!" because there is a real chance it will save critical time.

Why This is Better (The Results)

The researchers tested this system using real data from the Hong Kong Fire Services Department. They simulated thousands of heart attack emergencies and compared IDEAL against:

Old Rules: Just sending the nearest car based on a fixed map.
Google Maps Rules: Just sending the car with the shortest predicted time.
Always-Send-Two: Sending two cars every single time.

The Findings:

Better Safety Net: IDEAL found the fastest possible arrival time more often than the other methods. It was better at avoiding the "worst-case" delays.
Saving Resources: Unlike the "Always-Send-Two" rule, IDEAL didn't waste ambulances. It only sent the second car when it was truly necessary.
Tail Risk: This is the most important part. In emergency medicine, a few seconds matter. IDEAL drastically reduced the number of times an ambulance was severely late. It made the "worst-case" scenarios much less scary.

Summary in One Sentence

IDEAL is a smart dispatcher that learns from past traffic patterns to predict not just how long a trip will take, but how uncertain that prediction is, and only sends a backup ambulance when the math proves there's a real chance it will arrive significantly faster than the first one.

Technical Summary: Selective Ambulance Dispatch Under Contextual Travel-Time Uncertainty

Problem Statement
Out-of-hospital cardiac arrest (OHCA) requires immediate medical intervention, where survival rates drop significantly with even minute-long delays. Ambulance dispatchers face a critical trade-off: balancing the need for rapid response against limited fleet capacity. Traditional dispatch strategies rely on either static geographic territories or deterministic point predictions of travel time. However, static regions fail to adapt to dynamic congestion, and point predictions can be misleading when traffic conditions change en route. Conversely, "always-dual" dispatch (sending two ambulances for every high-priority call) ensures redundancy but consumes excessive fleet capacity, potentially leaving other areas uncovered.

The core challenge addressed is selective dual dispatch: determining when and which secondary path justifies the capacity cost of deploying a second ambulance. This decision is complicated by correlated travel-time uncertainties; congestion often affects multiple nearby road links simultaneously, meaning a secondary ambulance following a different route may still be delayed by the same underlying traffic state.

Methodology: The IDEAL Framework
The authors propose IDEAL (Intelligent Dual dispatch of Emergency AmbuLances), a scenario-aware framework that dispatches a second ambulance only when the "optimistic gap" between the primary path and a plausible secondary path exceeds a clinically meaningful threshold. The framework consists of three technical components:

Weakly Supervised Bilevel Representation Learning:
- Challenge: Historical dispatch records provide trip-level travel times but do not reveal the specific routes taken by ambulances.
- Solution: IDEAL employs a bilevel representation network. The lower-level problem solves a shortest-path problem based on predicted edge travel times, while the upper-level updates network parameters to minimize the loss between the predicted shortest-path time and the observed trip time.
- Training: The model uses a mini-batch conservative gradient method to handle the nonsmoothness arising from the shortest-path selection (where ties can occur). The authors prove an asymptotic convergence guarantee for this training procedure.
Contextual Correlated Scenario Sets via Metric Perturbations:
- Uncertainty Modeling: Instead of independent edge uncertainties, IDEAL models travel times by perturbing a shared metric in the learned representation space. Edge travel times are derived from the squared Euclidean norm of learned edge embeddings under a positive-definite matrix $X$ .
- Correlation: By perturbing the metric matrix $X$ (measured via Burg divergence from the identity), the model induces correlated changes across all edges. Edges with similar contextual representations respond similarly to the same metric perturbation.
- Radius Prediction: A separate network predicts a context-specific uncertainty radius $\rho_T$ based on historical underprediction errors, defining the size of the scenario set.
Efficient Oracle for Optimistic Gap Computation:
- Formulation: Computing the maximum plausible arrival-time advantage (optimistic gap) over the Burg scenario set is formulated as a Difference-of-Convex (DC) program over the metric matrix.
- Algorithm: The authors derive an efficient oracle using the Difference-of-Convex Algorithm (DCA). The key subproblem is reduced to a one-dimensional scalar search for a Lagrange multiplier, avoiding high-dimensional matrix optimization at dispatch time.
- Complexity: The authors provide complexity guarantees for the DCA procedure, showing it can be solved efficiently in real-time.

Key Contributions

Novel Learning Framework: The development of a weakly supervised bilevel representation network that learns context-specific edge travel times from trip-level data without requiring observed route labels.
Correlated Uncertainty Modeling: A method for constructing scenario sets via metric perturbations in a learned representation space, capturing the spatial correlation of traffic delays more effectively than independent edge models.
Computational Efficiency: An efficient oracle for the optimistic gap computation that transforms a complex matrix optimization into a tractable scalar search, enabling real-time decision-making.
Theoretical Guarantees: Proofs of asymptotic convergence for the bilevel training algorithm and convergence guarantees for the DCA-based dispatch-time optimization.

Experimental Results
The framework was evaluated in collaboration with the Hong Kong Fire Services Department using historical OHCA records (2014–2023) and real-time adaptive simulations on the Hong Kong Island road network.

Performance vs. Baselines: IDEAL was compared against region-based dispatch, Google Maps-based primary dispatch, Google-based dual dispatch (always sending two), and Google-based interval-selective dispatch.
Trade-off Optimization: IDEAL achieved a superior response-time/resource trade-off. At similar dispatch volumes, IDEAL reduced mean travel time and significantly lowered tail risk (severe delays).
Tail Risk Reduction: In the "always-dual" operating point, IDEAL reduced the 95th-percentile regret (time lost compared to the best possible candidate path) from 60.9 seconds (Google dual) to 20.5 seconds.
Statistical Significance: One-sided paired Wilcoxon signed-rank tests confirmed that IDEAL's improvements over all baselines, including the strong "Google dual" baseline, were statistically significant ( $p < 4.0 \times 10^{-7}$ ).
Transferability: A transfer test in the Sha Tin district (different road network and depot pair) demonstrated that the learned policy remains robust without retraining, outperforming local baselines.

Significance and Claims
The paper claims that IDEAL provides a rigorous, data-driven approach to the "selective redundancy" problem in emergency services. By explicitly modeling correlated travel-time uncertainty and learning context-specific representations, the framework avoids the inefficiency of always-dual dispatch while mitigating the risks of single-path reliance. The authors emphasize that the policy operates as an overlay on existing workflows, using external routing engines for primary paths and local capacity data to set thresholds, ensuring interpretability and auditability. The results suggest that selective dual dispatch, guided by the IDEAL framework, can significantly improve survival outcomes for OHCA patients by reducing the frequency and severity of response delays without proportionally increasing fleet strain.