Source-linked AI summary
Beyond Heuristics and Commercial Navigation: An AI-Augmented Decision-Support for Ambulance Travel Time Estimation
Muhammad Adil Abid, Johan Holmgren, Fabian Lorig, Jesper Petersson
TL;DR
Ambulance travel-time estimates support emergency dispatch, but common navigation, fixed-ratio, and lookup-table approaches may not reflect priority and temporal context. Using historical Skåne ambulance trips, the paper develops a context-aware machine-learning system and reports that it outperforms Google Maps estimates while travel-time differences vary by urgency.
Problem
Common ambulance travel-time approaches use generic navigation, fixed-ratio assumptions, or lookup tables, while evidence on how real trips vary by priority and temporal context is limited in the described setting.
Source: §1 · p. 1Method
The authors train a context-aware machine-learning estimator on historical Skåne ambulance trips and explore SHAP and permutation importance to assess operationally meaningful patterns.
Source: §1 · p. 2, §2.1 · p. 2Results
The proposed system reduces MAE versus Google Maps by 79.2%, 76.7%, and 52.7% for Priority 1, Priority 2, and Priority 3 trips, respectively, and both Google Maps and the 5% faster baseline overestimate travel time, particularly for higher-priority trips.
Source: §4.2 · p. 6Takeaways & Limitations
Ambulance travel time varies with urgency and temporal context, so a single fixed-speed adjustment does not represent the observed differences across priorities.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Accurate ambulance travel time estimation is critical for emergency medical dispatch, yet existing decision-support practices often rely on static lookup tables or simplified assumptions regarding how ambulance travel time differs from ordinary passenger-car travel. For instance, they assume ambulances simply travel a fixed percentage (e.g., 5%) faster than ordinary passenger cars. However, such assumptions are rarely validated using real ambulance operational data and may fail to capture the effects of trip priority, temporal context, and origin-destination. We address these limitations using historical ambulance trips from the Sk{\aa}ne region in Sweden. Our analysis reveals that ambulance travel times vary substantially by hour, day, and trip priority, and that commercial navigation systems fail to capture emergency-specific dynamics. We contribute a machine learning system to more accurately estimate ambulance travel time. To make the system interpretable, we explored SHapley Additive exPlanations and permutation importance techniques to ensure that our system captures operationally meaningful patterns.
1 Introduction
Ambulance dispatch estimates matter for timely prehospital care, but common approaches rely on generic navigation, fixed-speed adjustments, or lookup tables that may miss operational context. The paper proposes a data-driven, interpretable system using real ambulance operations and context-sensitive trip features.
1 Introduction
- Travel-time estimates matter because they affect how quickly prehospital care reaches patients and inform ambulance selection, hospital routing, response planning, and resource allocation.The introduction identifies temporal conditions, trip urgency, and local road-network characteristics as relevant context.Source: §1 · p. 1
- Ignoring operational context can produce estimates that do not reflect actual conditions, especially in time-sensitive cases.The paper presents this as a concern for emergency medical services.Source: §1 · p. 2
- Existing systems often use generic passenger-car navigation, fixed-ratio speed assumptions, or predetermined lookup tables that may miss priority, temporal variation, and local conditions.The paper motivates a data-driven alternative for more accurate estimates that support dispatch decisions.Source: §1 · p. 2
- The system learns from real ambulance operations and uses origin, destination, priority, and departure-time context to estimate travel time.SHAP and permutation importance are explored to assess which variables influence estimates and whether learned patterns are operationally meaningful.Source: §1 · p. 2
2 Dataset and Exploratory Data Analysis
The study uses completed ambulance trips from Skåne, Sweden, and finds travel-time variation across temporal, priority, and route-specific contexts. These patterns motivate retaining context rather than relying on static lookup tables or fixed-ratio assumptions.
2.1 Dataset
- The dataset covers 2020–2021 completed trips and uses dispatch-available origin, destination, priority, weekday, and hour to predict elapsed minutes from departure to arrival.Priority levels distinguish emergency, urgent, and routine trips; cancelled trips and non-completed mission segments were filtered.Source: §2.1 · p. 2
2.2 Exploratory Data Analysis
- Ambulance trips from Skåne, Sweden, show substantial variation by time, day, and priority, challenging static lookup tables and fixed-ratio assumptions.The exploratory analysis also identifies route-specific factors as part of the variation.Source: §2.2 · pp. 2–3
- 27.90 min at 05:00 and 43.35 min at 09:00 mark the lowest and highest hourly mean travel times, a 15.45-minute difference.Travel times rise in the morning and remain elevated through daytime.Source: §2.2 · p. 3, Figure 1 · p. 3
- 38.18 min on weekdays versus 34.38 min on weekends summarizes the reported weekday–weekend mean travel-time difference.Thursday has the highest reported daily mean at 38.71 min.Source: §2.2 · p. 3, Figure 1 · p. 3
- Weekday hourly peaks reach 48.5 minutes on Mondays, while weekend values are generally lower and the observed range is 21.9–48.5 minutes.The joint patterns support retaining hour of day and day of week as separate features.Source: §2.2 · p. 3, Figure 1 · p. 3
- Priority 1 emergency trips average 29.3 minutes versus 39.1 minutes for Priority 3 routine trips, a nearly 10-minute difference.The analysis attributes the difference to emergency-driving practices and warns that estimates not conditioned on priority would systematically mispredict the highest-acuity trips.Source: §2.2 · p. 3, Figure 1 · p. 3
3 AI-Augmented Decision Support for Ambulance Travel Time Estimation
The system combines dispatch-available trip context with supervised machine learning, then uses explainability analyses and comparisons with navigation and fixed-ratio baselines. Its inputs and evaluation are designed around operational trip context and priority.
3 AI-Augmented Decision Support for Ambulance Travel Time Estimation
- The workflow overview introduces the study’s system, with subsequent subsections detailing its steps.The overview is presented in Fig. 2.Source: §3 · p. 3, Figure 2 · p. 4
3.1 Data Preprocessing
- Five dispatch-available features—origin area, destination area, trip priority, weekday, and hour—form the model input after spatial categories are one-hot encoded and numeric features standardized.The standardizer is fitted only on the corresponding training partition to avoid data leakage.Source: §3.1 · p. 4
3.2 Prediction
- The study explores an ANN, polynomial linear regression, two CatBoost variants, and random forest for supervised ambulance travel-time estimation.The candidates represent statistical, nonlinear tabular, and interaction-learning approaches.Source: §3.2 · p. 4
3.3 Explainability Methods
- Permutation importance and SHAP are applied on held-out data to assess feature influence and whether predictions reflect operationally meaningful patterns.Permutation importance measures the MAE increase after feature shuffling, including joint origin–destination block permutation; SHAP values for one-hot spatial columns are aggregated to parent features.Source: §3.3 · pp. 4–5
3.4 Baseline Travel Time Estimation Methods
- The evaluation compares model estimates with Google Maps and a fixed-ratio baseline, stratifying comparisons by trip priority.Google Maps estimates use matched weekday-hour queries for the next future occurrence, not historical departure times; the fixed-ratio baseline scales Google Maps durations by 0.95, following.Source: §3.4 · p. 5
4 Experiments and Results
The ANN-based system outperformed the Google Maps and fixed-ratio baselines, while explainability analyses highlighted spatial context alongside priority and temporal features.
4.1 Comparison of Candidate ML Models
- The ANN achieved the best overall performance, with the lowest MAPE, MAE, and RMSE and the highest R2, and was selected for the proposed system.Table 1 presents the machine-learning model comparisons.Source: §4.1 · p. 5, Table 1 · p. 6
4.2 Proposed System Comparison with Baseline Methods
- 79.2%, 76.7%, and 52.7% lower MAE than Google Maps for Priority 1, Priority 2, and Priority 3 trips, respectively; the proposed system also achieved similar reductions against the 5% faster baseline.Both baselines overestimated ambulance travel time, particularly for higher-priority trips, while the proposed system stayed closer to observed times.Source: §4.2 · p. 6, Figure 3 · p. 6
4.3 Explainability Analysis
- Origin area and destination area were the most influential permutation features, increasing MAE by 14.6 and 14.4 minutes, respectively; priority and time features also contributed.A joint origin–destination permutation preserved valid routes and supported the interpretation that spatial importance reflects route-level structure.Source: §4.3 · p. 7, Figure 4 · p. 7
- Global SHAP importance likewise ranked origin and destination highest, with mean absolute values of 7.8 and 7.3 minutes, followed by priority, hour of day, and day of week.The reported values for the remaining features were 2.1, 1.1, and 0.4 minutes, respectively.Source: §4.3 · p. 7, Figure 4 · p. 7
5 Conclusions
Using Skåne dispatch data, the paper developed a context-aware machine-learning system and reports that ambulance travel time varies across spatial, temporal, and priority contexts. Its findings argue against a single fixed-speed adjustment and support ambulance-specific, interpretable, priority-aware estimation.
5 Conclusions
- The Skåne study developed a data-driven, context-aware machine-learning system and reports substantial variation in ambulance travel time across spatial, temporal, and priority contexts.The conclusion contrasts this variation with static lookup tables, fixed-ratio heuristics, and generic commercial navigation estimates.Source: §5 · pp. 7–8
- The implied speed advantage over commercial estimates was roughly 46% for emergency trips versus 13% for routine trips, contradicting a single fixed-speed adjustment.The gap between commercial estimates and actual ambulance travel varied systematically by urgency.Source: §5 · p. 8
- The authors conclude that ambulance-specific, interpretable, priority-aware AI components support travel-time estimation, transparency, and operational responsiveness in cooperative emergency dispatch systems.They describe the proposed system as a practical, low-latency component that can be integrated into EMS systems to reduce cognitive load and improve priority decisions.Source: §5 · p. 8