Towards Trustworthy AI-Enabled Emergency Medical Dispatch: Uncertainty-Aware Ambulance Travel Time Estimation

# Towards Trustworthy AI-Enabled Emergency Medical Dispatch: Uncertainty-Aware Ambulance Travel Time Estimation
Source: [https://arxiv.org/html/2610.31944](https://arxiv.org/html/2610.31944)
Muhammad Adil Abid🖂[https://orcid.org/0000-0002-0403-5353](https://orcid.org/0000-0002-0403-5353)††thanks:Corresponding author:E\-mail[muhammad\.adil\-abid@mau\.se](mailto:[email protected])Affiliation:Department of Computer Science and Media Technology, Malmö University, 21119 Malmö, SwedenAffiliation:Sustainable Digitalisation Research Centre, Malmö University, SwedenJohan Holmgren[https://orcid.org/0000-0001-7773-9944](https://orcid.org/0000-0001-7773-9944)Affiliation:Department of Computer Science and Media Technology, Malmö University, 21119 Malmö, SwedenAffiliation:Sustainable Digitalisation Research Centre, Malmö University, SwedenFabian Lorig[https://orcid.org/0000-0002-8209-0921](https://orcid.org/0000-0002-8209-0921)Affiliation:Department of Computer Science and Media Technology, Malmö University, 21119 Malmö, SwedenAffiliation:Sustainable Digitalisation Research Centre, Malmö University, SwedenJesper Petersson[https://orcid.org/0000-0003-3322-6383](https://orcid.org/0000-0003-3322-6383)E\-mail[jesper\.petersson@skane\.se](mailto:[email protected])Affiliation:Department of Health Care Management, Region Skåne, 21428 Malmö, SwedenAffiliation:Department of Neurology, Lund University, 22242 Malmö, SwedenE\-mail[\{muhammad\.adil\-abid,johan\.holmgren,fabian\.lorig\}@mau\.se](mailto:{muhammad.adil-abid,johan.holmgren,fabian.lorig}@mau.se)

###### Abstract

Emergency medical dispatch increasingly relies on AI\-enabled decision support systems when making ambulance dispatch decisions\. In this process, accurate ambulance travel time estimates are essential for timely resource allocation and prehospital care\. However, existing ambulance travel time estimation approaches typically produce only single point predictions\. Consequently, they do not indicate how reliable an individual estimate is or how much uncertainty it carries\. This limitation is particularly problematic in safety\-critical emergency medical service dispatch operations, where trips with similar expected travel times may exhibit substantially different levels of uncertainty\. Furthermore, real\-world ambulance trip logs may contain domain\-specific artefacts, such as co\-located healthcare facilities and physically implausible travel speeds\. These artefacts can degrade point prediction accuracy and affect the prediction task on which subsequent uncertainty estimation depends\. To address these challenges, we propose an uncertainty\-aware methodology for ambulance travel time estimation and evaluate it through a retrospective real\-world scenario study using two years of ambulance trip data from the Skåne region in Sweden\. The proposed methodology moves beyond average\-case point predictions by generating context\-adaptive prediction intervals with stable empirical coverage across the evaluated operationally important contexts\. The resulting prediction intervals provide explicit information about predictive uncertainty that can support more transparent and risk\-sensitive ambulance dispatch decision\-making\.

###### Keywords:

Ambulance travel time estimation Decision support systems Emergency medical dispatch Uncertainty quantification Trustworthy AI

## 1Introduction

Emergency medical dispatch is a high\-stakes decision\-making process that increasingly relies on AI\-enabled decision support systems that combine operational data, predictive models, and human judgement\. Within this public\-sector service, ambulance travel time estimation is an important component of dispatch decision support, as travel time estimates provide important information on which available unit should be assigned to an incident and how quickly prehospital care may reach the patient\. Even small estimation errors can lead to the assignment of an ambulance that arrives later than another available unit, which matters most for high\-priority incidents where delays may have clinical consequences\. Ambulance travel time estimation is challenging because travel times vary with local conditions, dispatch priority, route characteristics, and temporal context\[[1](https://arxiv.org/html/2610.31944#bib.bib1)\]\. To capture this complexity, recent research has increasingly adopted data\-driven machine learning approaches that can improve predictive accuracy compared with static lookup tables, fixed\-speed assumptions, and generic navigation estimates\[[2](https://arxiv.org/html/2610.31944#bib.bib2),[3](https://arxiv.org/html/2610.31944#bib.bib3),[4](https://arxiv.org/html/2610.31944#bib.bib4)\]\.

However, predictive accuracy alone is insufficient for safety\-critical dispatch support\. Most existing ambulance travel time estimation approaches provide only a single point prediction\. A single point prediction provides one estimated travel time, for example, that an ambulance will arrive in1212minutes, without indicating the uncertainty associated with that estimate\. This missing information is important in emergency medical service \(EMS\) decision support because predictions with similar expected travel times may differ substantially in their uncertainty\. For example, consider two ambulances available for a time\-critical emergency\. A point prediction model estimates travel times of1010minutes for Ambulance A and1111minutes for Ambulance B, which would favour Ambulance A based on the expected travel time alone\. Suppose, however, that an uncertainty\-aware methodology produces an interval of±5\\pm 5minutes for Ambulance A and±1\\pm 1minute for Ambulance B\. If an operational criterion requires arrival within1212minutes, Ambulance B may represent the lower\-uncertainty option under that criterion\. More generally, uncertainty information could support dispatchers when considering whether an individual travel time estimate is sufficiently reliable or whether additional operational information should be considered\. Thus, an uncertainty\-aware estimation complements the point estimate by explicitly representing the range of plausible travel times rather than providing only an expected value\. As our exploratory analysis in Section[4\.3](https://arxiv.org/html/2610.31944#S4.SS3)shows, travel time variability differs across temporal conditions and route frequency groups, while substantial variability remains even among trips with similar operational characteristics\.

Another important challenge concerns the operational validity of the*underlying data*on which predictive models should be developed\. Real\-world EMS datasets may contain domain\-specific artefacts that distort the prediction task\. First, operationally distinct facilities, such as co\-located hospitals and on\-campus clinics, may be geocoded to identical coordinates, creating ambiguity in origin–destination representations\. Second, a small subset of trips may have implied average speeds outside physically plausible ranges for ground ambulance transport\. Extremely low speeds may indicate idle or waiting time included in the recorded trip duration, whereas implausibly low or high speeds may also result from GPS inaccuracies or timestamp inconsistencies\. For example, departure or arrival events may be manually registered before or after the corresponding operational event, making the recorded trip duration too long or too short\. We therefore define two curation criteria a priori from domain knowledge: removing trips between facilities that share the same coordinates and removing trips with implausible average speeds\. Applying these criteria before model development provides a stronger basis for point prediction and subsequent uncertainty estimation\.

In this paper, we focus on three limitations that remain insufficiently addressed in ambulance travel time estimation research: \(i\)*deterministic predictions*, which do not quantify the uncertainty associated with individual travel time estimates; \(ii\)*operational data quality*, where domain\-specific artefacts in historical dispatch records may distort the prediction task; and \(iii\)*subgroup reliability*, where satisfactory aggregate calibration may conceal differences in empirical coverage across operationally relevant contexts\. To address these limitations, we propose an uncertainty\-aware methodology for ambulance travel time estimation and evaluate it through a retrospective real\-world scenario study using two years of ambulance trip data from the Skåne region in Sweden\. The proposed methodology first applies domain\-informed data curation to establish an operationally valid prediction task\. The resulting data are then prepared for model development, and a point prediction model is trained to estimate ambulance travel time\. Conformal calibration is subsequently used to generate prediction intervals that quantify predictive uncertainty\. Under the exchangeability assumption, the conformal methods provide finite\-sample marginal coverage guarantees\. By complementing point predictions with calibrated prediction intervals and evaluating their empirical coverage across operationally relevant contexts, the proposed methodology provides a basis for assessing predictive reliability beyond conventional point prediction accuracy alone\.

The key contributions of this paper are summarised as follows:

- •An uncertainty\-aware ambulance travel time estimation methodology that combines domain\-informed data curation, feature engineering and preprocessing, point prediction modelling, and conformal calibration to explicitly quantify predictive uncertainty for EMS decision support\.
- •A domain\-informed data curation process that establishes a more operationally valid prediction task than the uncurated formulation and provides a stronger basis for point prediction and subsequent uncertainty estimation\.
- •A retrospective real\-world scenario study in the Skåne region of Sweden that instantiates and evaluates the proposed methodology using two years of ambulance trip data, including a comparison of three conformal calibration strategies and their empirical coverage across operationally relevant contexts\.

## 2Related Work

Accurate ambulance travel time estimation is fundamental to EMS dispatch operations, as it influences ambulance allocation and compliance with response time targets\. The simplest approaches rely on fixed\-speed assumptions and lookup tables, which fail to capture temporal and operational variability\. Statistical approaches model ambulance travel times probabilistically, for example using Bayesian data augmentation\[[5](https://arxiv.org/html/2610.31944#bib.bib5)\]\. Other approaches use commercial navigation services, which systematically misestimate lights\-and\-siren ambulance travel times\[[4](https://arxiv.org/html/2610.31944#bib.bib4)\]\. More recent machine learning methods, including gradient boosting, random forests, and neural networks, substantially improve prediction accuracy by exploiting GPS trajectories and spatiotemporal features\[[2](https://arxiv.org/html/2610.31944#bib.bib2),[3](https://arxiv.org/html/2610.31944#bib.bib3),[6](https://arxiv.org/html/2610.31944#bib.bib6)\]\. However, these approaches almost exclusively produce deterministic point predictions and generally assume that historical dispatch records are free from operational artefacts\.

Predictive uncertainty receives more attention in the broader travel time prediction literature than in EMS research\. Approaches that account for uncertainty include Bayesian neural networks, quantile regression\[[7](https://arxiv.org/html/2610.31944#bib.bib7)\], ensemble\-based prediction intervals, and, more recently, conformal prediction\[[8](https://arxiv.org/html/2610.31944#bib.bib8)\]\. Conformal prediction provides distribution\-free prediction intervals with finite\-sample marginal coverage guarantees under the exchangeability assumption\. Although it shows promise for general travel time and trajectory prediction\[[9](https://arxiv.org/html/2610.31944#bib.bib9)\], its application to ambulance travel time estimation remains largely unexplored\. To address these limitations, we propose an uncertainty\-aware methodology that combines operational data curation with calibrated prediction intervals and evaluate it through a retrospective real\-world scenario study\.

## 3Methodology for Uncertainty\-Aware Ambulance Travel Time Estimation

Fig\.[1](https://arxiv.org/html/2610.31944#S3.F1)summarizes the proposed uncertainty\-aware ambulance travel time estimation methodology, which consists of four main stages\. Each stage is detailed below\.

### 3\.1Domain\-Informed Data Curation

Real\-world EMS operational data may contain recording artefacts that adversely affect predictive accuracy and the prediction task on which subsequent uncertainty estimation depends\. Therefore, we propose performing domain\-informed data curation before model development\. The first curation step addresses spatial ambiguities arising when operationally distinct healthcare facilities share identical geographical coordinates, such as hospitals and co\-located clinics\. Records affected by these ambiguities can yield degenerate or near\-zero origin–destination separations that do not correspond to meaningful ambulance transfers\. Such records are therefore removed before model development\. The second curation step addresses trips with physically implausible average speeds\. For each trip, the implied average speed is computed from the origin–destination route distance and the recorded travel time\. Trips with average speeds outside an a priori plausible range for ground ambulance transport are removed\. The plausible range should be defined based on domain knowledge for the operational setting in which the methodology is applied\. Extremely low speeds may indicate idle or waiting time included in the recorded trip duration, whereas implausibly low or high speeds may also result from GPS inaccuracies or timestamp inconsistencies\. For example, departure or arrival events may be manually registered before or after the corresponding operational event, causing the recorded trip duration to be too long or too short\. The two curation criteria are defined a priori based on EMS operational knowledge rather than predictive performance\. Together, they establish an operationally valid prediction task for subsequent point prediction and uncertainty estimation\.

Figure 1:High\-level workflow of the proposed uncertainty\-aware ambulance travel time estimation methodology\.
### 3\.2Feature Engineering and Preprocessing

The methodology uses information available at dispatch time to construct features relevant to ambulance travel time estimation\. These may include spatial, temporal, operational, and route\-related information\. Categorical features are represented using an encoding compatible with the selected prediction model, while numerical features are standardized when required by the modelling approach\. All preprocessing transformations are fitted using training data only to avoid information leakage\. The resulting feature representation is then used as input to the point prediction model\. The specific implementation used in the real\-world scenario study is described in Section[4\.4](https://arxiv.org/html/2610.31944#S4.SS4)\.

### 3\.3Point Prediction Model

The point prediction stage is formulated as a supervised regression task in which ambulance travel time is estimated from dispatch\-time features\. The proposed methodology is not tied to a particular regression algorithm\. Instead, one or more candidate models can be evaluated using a common validation protocol, and the best\-performing model can subsequently be used as the point prediction component of the methodology\. Model selection should be based on predictive performance measured on data that are independent of the final test set\. The specific model candidates and model selection protocol used in the real\-world scenario study are described in Section[4\.4](https://arxiv.org/html/2610.31944#S4.SS4)\.

### 3\.4Uncertainty\-Aware Ambulance Travel Time Estimation

The point prediction model generated in stage33provides a single estimated travel time for each ambulance dispatch\. However, a point estimate alone does not represent the uncertainty associated with an individual prediction\. The final stage of the proposed methodology therefore applies conformal calibration to complement point predictions with calibrated prediction intervals\. Conformal prediction does not require a parametric assumption about the prediction error distribution and provides finite\-sample marginal coverage guarantees under the exchangeability assumption\[[8](https://arxiv.org/html/2610.31944#bib.bib8)\]\. The methodology considers three conformal calibration strategies: global split conformal prediction, Conformalized Quantile Regression \(CQR\), and priority\-stratified Mondrian conformal prediction\. Fig\.[2](https://arxiv.org/html/2610.31944#S3.F2)illustrates the global split conformal workflow, comprising model development and calibration \(steps 1–5\) and prediction for new dispatch requests \(steps 6–9\)\.

Figure 2:Step\-by\-step workflow of global split conformal prediction for uncertainty\-aware ambulance travel time estimation\.Model development and calibration \(steps 1–5\)\.The global split conformal workflow begins with a curated historical ambulance trip dataset obtained through the domain\-informed data curation process described in Section[3\.1](https://arxiv.org/html/2610.31944#S3.SS1)\. The point prediction modelffis trained to estimate ambulance travel time from the dispatch\-time features \(steps 1–2\)\. To calibrate the prediction intervals, a portion of the historical trips is reserved exclusively for calibration and is not used for model training or selection \(step 3\)\. This*calibration set*,𝒟cal=\{\(𝐱i,yi\)\}i=1n\\mathcal\{D\}\_\{\\mathrm\{cal\}\}=\\\{\(\\mathbf\{x\}\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{n\}, where𝐱i\\mathbf\{x\}\_\{i\}denotes the dispatch\-time feature vector andyiy\_\{i\}the observed travel time, is used to quantify prediction errors on observations not used to train the point prediction model\. For each calibration trip, the absolute prediction error is computed \(step 4\), yielding the nonconformity score

ri=\|yi−f\(𝐱i\)\|,i=1,…,n\.r\_\{i\}=\\left\|y\_\{i\}\-f\(\\mathbf\{x\}\_\{i\}\)\\right\|,\\qquad i=1,\\ldots,n\.\(1\)The final calibration step \(step 5\) converts these historical prediction errors into a calibration margin, denoted byq^\\hat\{q\}\. For a target miscoverage levelα\\alpha, corresponding to nominal coverage1−α1\-\\alpha,q^\\hat\{q\}is chosen as the⌈\(n\+1\)​\(1−α\)⌉\\lceil\(n\+1\)\(1\-\\alpha\)\\rceil\-th smallest value amongr1,…,rnr\_\{1\},\\ldots,r\_\{n\}\. This finite\-sample\-adjusted quantile, rather than the empirical\(1−α\)\(1\-\\alpha\)quantile, provides the conformal marginal coverage guarantee under the exchangeability assumption\[[8](https://arxiv.org/html/2610.31944#bib.bib8)\]\.

Prediction for new dispatch requests \(steps 6–9\)\.For a new ambulance dispatch request \(step 6\), the trained point prediction model estimates the travel time from the dispatch\-time features \(step 7\)\. The calibration margin obtained during model development is then added to and subtracted from the point prediction \(step 8\), producing a calibrated prediction interval together with the point estimate \(step 9\):

C⁡\(𝐱\)=\[f⁡\(𝐱\)−q^,f⁡\(𝐱\)\+q^\]\.C\(\\mathbf\{x\}\)=\\left\[f\(\\mathbf\{x\}\)\-\\hat\{q\},\\;f\(\\mathbf\{x\}\)\+\\hat\{q\}\\right\]\.\(2\)
For example, a predicted travel time of3636minutes combined with a calibration margin ofq^=8\\hat\{q\}=8minutes produces the prediction interval\[28,44\]\[28,44\]minutes \(see Fig\.[2](https://arxiv.org/html/2610.31944#S3.F2)\)\.

Interpretation\.Under the exchangeability assumption, conformal prediction provides marginal coverage of at least1−α1\-\\alpha\. Thus, over repeated predictions, the proportion of future ambulance travel times contained within the reported prediction intervals is expected to be at least the specified nominal coverage level\.

Context\-aware calibration\.Global split conformal prediction applies the same calibration margin to every prediction and therefore does not adapt interval width to differences in predictive uncertainty across individual trips\. To allow the methodology to account for such differences, we additionally consider two context\-aware conformal calibration strategies\. \(i\) Conformalized Quantile Regression \(CQR\)\[[10](https://arxiv.org/html/2610.31944#bib.bib10)\]generates trip\-specific prediction intervals from estimated lower and upper conditional quantiles\. A subsequent conformal calibration step adjusts these intervals to attain the desired marginal coverage\. \(ii\) Priority\-stratified \(Mondrian\) conformal prediction\[[11](https://arxiv.org/html/2610.31944#bib.bib11)\]computes a separate calibration quantile for each dispatch priority group and applies the corresponding margin to new predictions within that group\.

The empirical performance of global split conformal prediction, CQR, and priority\-stratified Mondrian conformal prediction is evaluated in the real\-world scenario study presented in Section[4](https://arxiv.org/html/2610.31944#S4)\.

## 4Real\-World Scenario Study and Results

### 4\.1Scenario Description

To evaluate the proposed uncertainty\-aware methodology, we instantiated it in a retrospective real\-world scenario study based on ambulance operations in the Skåne region of southern Sweden using historical trips recorded during 2020–2021\. The scenario study examines: \(i\) the effect of domain\-informed data curation on point prediction performance; \(ii\) the predictive performance of alternative point prediction models; and \(iii\) the calibration and context\-specific reliability of the resulting prediction intervals\. The dataset, feature configuration, candidate models, and evaluation settings therefore represent one real\-world instantiation of the proposed methodology\.

### 4\.2Dataset

For the real\-world scenario study, we used historical ambulance trip records from the Skåne region in southern Sweden, covering the period 2020–2021\. The dataset was provided by the regional emergency dispatch authority and contained operational information about ambulance trips\. The scenario\-specific input features were origin area, destination area, trip priority, day of week, hour of day, origin–destination route distance, and a static Google travel time estimate\. The Google estimate was generated offline and stored as a precomputed route duration attribute, requiring no live API call during inference\. Trip priority was represented using three levels: Priority11denotes emergency trips, Priority22denotes urgent trips, and Priority33denotes routine or less time\-critical trips\. The prediction target was ambulance travel time in minutes, measured as the elapsed time between the recorded ambulance departure and arrival timestamps\. Cancelled trips and incomplete mission segments were excluded from the dataset\. For this scenario, the domain\-informed data curation procedure used an a priori plausible ground ambulance speed range of55–150150km/h\. Together with the removal of spatial ambiguities associated with co\-located healthcare facilities, the data curation procedure reduced the dataset from22,36022\{,\}360original trips to20,73820\{,\}738operationally valid trips\.

### 4\.3Exploratory Data Analysis

To characterize the evaluation scenario, we examined travel time variability across operational contexts\. The analysis yielded four observations that inform the subsequent evaluation of prediction interval calibration across operationally relevant contexts\.

Observation11: Travel time variability differs across operational contexts\.Panel \(a\) of Table[1](https://arxiv.org/html/2610.31944#S4.T1)reports travel time variability pooled across all origin–destination routes within each priority group\. These statistics therefore reflect both differences in route length distributions and variation between trips on the same route\. Relative variability, measured using the coefficient of variation \(CV\), is similar across priorities \(CV=0\.44\\mathrm\{CV\}=0\.44–0\.470\.47\), although Priority33trips, which are also the longest on average \(39\.539\.5min\), exhibit the highest absolute variability \(18\.0718\.07min\)\. Panel \(b\) shows clearer differences according to route frequency\. Very rare routes have a CV of0\.5920\.592, compared with0\.3060\.306–0\.4250\.425for routes observed more than100100times\. Temporal variability also differs across the day, with lower variability during nighttime hours and higher variability during weekday mornings, particularly around09:0009\{:\}00\(see Fig\.[3](https://arxiv.org/html/2610.31944#S4.F3)\)\. Variability among trips sharing similar operational characteristics is examined further in Observation33\.

Table 1:Travel time variability across operational contexts\.\(a\) Variability by ambulance priority

PriorityMean \(min\)Std\. Dev\. \(min\)CVTripsPriority 1 \(Emergency\)29\.7913\.840\.4652,345Priority 2 \(Urgent\)35\.2415\.470\.4393,972Priority 3 \(Routine\)39\.5118\.070\.45714,421
\(b\) Variability by OD route frequency

OD Frequency GroupTripsStd\. Dev\. \(min\)IQR \(min\)CVVery rare \(1–5\)10556\.8075\.630\.592Rare \(6–20\)14332\.8337\.710\.484Medium \(21–100\)52322\.5632\.600\.394Common \(101–500\)3,87115\.2517\.890\.306Very common \(\>500\>500\)16,09614\.3712\.920\.425

Observation22: Travel times exhibit a pronounced right tail\.The travel time distribution is strongly right\-skewed\. Half of all trips are completed within3131min, whereas5%5\\%exceed7070min and1%1\\%exceed9494min\. Thus, long\-duration trips form a non\-negligible upper tail of the observed distribution\. Although the data curation procedure removes trips with physically implausible average speeds, some long\-duration observations may still reflect delayed registration of departure or arrival events in addition to genuine operational variability\. A symmetric Gaussian approximation may therefore inadequately represent the observed asymmetric tail of ambulance travel times\.

Observation33: Substantial variability remains under similar operational conditions\.Even when trips are grouped by the same origin–destination route and hour of day, substantial travel time variability remains\. Across234234groups containing at least2020trips, the median within\-group standard deviation is approximately77min\. This remaining variability may reflect factors not represented by the available dispatch\-time features, including traffic conditions, weather, and incident\-specific circumstances\. These results indicate that trips with similar observed operational characteristics can still exhibit substantial differences in travel time, making context\-specific evaluation of prediction intervals important in the scenario study\.

Table 2:Empirical coverage of a naive global95%95\\%reference interval computed from the training set mean and standard deviation\.ContextCoverageAll trips \(average case\)96\.5%Morning peak \(09:00\)94\.4%Rare routes \(6–20 trips\)69\.9%Very rare routes \(1–5 trips\)44\.2%Observation44: Aggregate coverage can conceal subgroup differences\.As a simple diagnostic, we applied a single global95%95\\%reference interval across all operational contexts \(Table[2](https://arxiv.org/html/2610.31944#S4.T2)\)\. The interval, computed as the training set mean±1\.96\\pm 1\.96standard deviations, achieved96\.5%96\.5\\%empirical coverage overall but only69\.9%69\.9\\%for rare routes and44\.2%44\.2\\%for very rare routes\. This illustrates that satisfactory aggregate coverage does not necessarily imply comparable coverage across operationally relevant subgroups\.

Figure 3:Standard deviation of ambulance travel time by hour of day and day of week \(Darker cells indicate greater travel time variability\.\)Collectively, these observations motivate evaluating prediction interval calibration both overall and across operationally relevant contexts in the Skåne scenario\.

### 4\.4Experimental Setup and Evaluation Protocol

The following experiments instantiate and evaluate the proposed methodology under the Skåne scenario described above\. For the point prediction stage, we considered polynomial linear regression, random forest, CatBoost with native categorical handling, CatBoost with one\-hot encoding, and an artificial neural network \(ANN\)\. All experiments were implemented in Python on an Apple MacBook Pro 2022 \(1616\-core GPU,1616GB RAM\)\.

To evaluate*point prediction generalization*, we reserved15%15\\%of the data as an independent test set\. The remaining85%85\\%was used for five\-fold cross\-validation, with candidate model configurations compared using their mean validation performance across folds\. After selecting the best\-performing configuration, the corresponding model was retrained on the complete development set and evaluated once on the independent test set using mean absolute error \(MAE\), root mean square error \(RMSE\), mean absolute percentage error \(MAPE\), and the coefficient of determination \(R2R^\{2\}\)\.

For*uncertainty estimation*, the same independent15%15\\%test set was retained for final evaluation\. Of the full dataset,70%70\\%was used for training and15%15\\%for calibration\. The training set was used to fit the selected point prediction model and the quantile models, whereas the calibration set was used exclusively to compute the conformal nonconformity scores\. Thus, the point prediction evaluation used an 85/15 development–test split, whereas the uncertainty estimation evaluation used a 70/15/15 training–calibration–test split to preserve an independent calibration set\.

The conformal calibration strategies were evaluated primarily at a nominal coverage level of90%90\\%\(α=0\.10\\alpha=0\.10\)\. Calibration was additionally examined across nominal coverage levels ranging from50%50\\%to99%99\\%, and the stability of the results was assessed across repeated random training–calibration–test partitions\.

To assess the computational feasibility of the scenario study implementation for real\-time use, we measured the end\-to\-end inference latency, including feature preprocessing, point prediction, and calibrated prediction interval generation\. The implementation achieved an average single\-trip inference latency of approximately22ms on the stated hardware, indicating low computational overhead for real\-time EMS decision support\.

Table 3:Effect of domain\-informed data curation on point prediction performance \(↓\\downarrowlower is better;↑\\uparrowhigher is better\)\.Task formulationMAPE↓\\downarrow\(%\)MAE↓\\downarrow\(min\)RMSE↓\\downarrow\(min\)R2R^\{2\}↑\\uparrowOperationally naive \(baseline\)12\.474\.8310\.860\.62Spatially disambiguated12\.904\.6510\.650\.65Operationally valid10\.354\.429\.120\.74
### 4\.5Effect of Domain\-Informed Data Curation on Point Prediction Performance

We first evaluated how the domain\-informed data curation stage affected point prediction performance in the Skåne scenario\. For this purpose, we constructed three progressively curated prediction task formulations\. Theoperationally naiveformulation retained the data before applying the two domain\-informed curation criteria\. Thespatially disambiguatedformulation additionally removed trips affected by co\-located healthcare facilities\. Finally, theoperationally validformulation also removed trips with implied average speeds outside the a priori plausible range of55–150150km/h described in Section[4\.2](https://arxiv.org/html/2610.31944#S4.SS2)\. All three formulations were evaluated using the same ANN architecture, input features, preprocessing procedure, and evaluation protocol\. This controlled comparison isolates the effect of progressively applying the domain\-informed data curation criteria on point prediction performance\. Table[3](https://arxiv.org/html/2610.31944#S4.T3)shows that theoperationally validformulation achieved the best overall point prediction performance\. Relative to theoperationally naivebaseline, it reduced MAPE by17\.0%17\.0\\%, MAE by8\.5%8\.5\\%, and RMSE by16\.0%16\.0\\%, while increasingR2R^\{2\}from0\.620\.62to0\.740\.74\. The intermediatespatially disambiguatedformulation reduced MAE and RMSE and increasedR2R^\{2\}, although its MAPE increased slightly\. This may result from removing degenerate short trips, thereby changing the distribution over which percentage errors are calculated\. The subsequent removal of trips with physically implausible average speeds produced the lowest prediction errors among the three evaluated formulations\. Overall, the results show that domain\-informed data curation improved point prediction performance while removing only7\.3%7\.3\\%of the original records\. This indicates that a relatively small proportion of operationally problematic records had a substantial effect on the prediction task\. These results concern point prediction performance only and do not, by themselves, establish improved uncertainty calibration\. Prediction interval calibration is evaluated separately in Section[4\.7](https://arxiv.org/html/2610.31944#S4.SS7)\.

### 4\.6Point Prediction Performance

Table[4](https://arxiv.org/html/2610.31944#S4.T4)compares the candidate point prediction models\. The ANN achieved the best overall performance, with a MAPE of10\.35%10\.35\\%, MAE of4\.424\.42min, RMSE of9\.129\.12min, andR2R^\{2\}of0\.740\.74, although the CatBoost models performed similarly\. Polynomial linear regression produced substantially larger prediction errors\. Based on its performance across the evaluated metrics, the ANN was selected as the point prediction model for the subsequent uncertainty estimation experiments in this instantiation of the proposed methodology\.

Table 4:Comparison of candidate point prediction models \(↓\\downarrowlower is better;↑\\uparrowhigher is better\)\.ModelMAPE↓\\downarrow\(%\)MAE↓\\downarrow\(min\)RMSE↓\\downarrow\(min\)R2R^\{2\}↑\\uparrowPolynomial Linear Regression21\.118\.0112\.970\.48Random Forest13\.545\.3610\.380\.67CatBoost one\-hot10\.374\.459\.570\.72CatBoost native10\.364\.449\.420\.73ANN10\.354\.429\.120\.74
### 4\.7Evaluation of the Proposed Uncertainty\-Aware Methodology

We evaluated the uncertainty estimation stage of the proposed methodology in the Skåne scenario study by considering three evaluation questions\.

Q1\. Do the conformal calibration strategies achieve empirical coverage close to the desired nominal coverage?We compared the three conformal calibration strategies against two non\-conformal baselines at a nominal coverage level of90%90\\%\(Table[5](https://arxiv.org/html/2610.31944#S4.T5)\)\. The Gaussian baseline produced the widest intervals, with a mean width of24\.724\.7min, and achieved93\.2%93\.2\\%empirical coverage\. Although its empirical coverage exceeded the nominal target, the resulting intervals were wider than those produced by the conformal methods, indicating lower interval efficiency\. Among the conformal strategies, global split conformal prediction achieved89\.1%89\.1\\%empirical coverage with a constant calibration margin ofq^=8\.8\\hat\{q\}=8\.8min\. CQR achieved90\.5%90\.5\\%empirical coverage, while priority\-stratified Mondrian conformal prediction achieved89\.0%89\.0\\%\. Thus, all three conformal strategies produced empirical coverage close to the nominal90%90\\%target\. Under the exchangeability assumption, the conformal procedures additionally provide finite\-sample marginal coverage guarantees\.

Table 5:Comparison of uncertainty estimation methods at nominal90%90\\%coverage \(↓\\downarrowlower is better\)\.MethodCoverage\(%\)Mean width↓\\downarrow\(min\)Median width↓\\downarrow\(min\)BaselinesGaussian \(f⁡\(x\)±1\.645​σ^f\(x\)\\pm 1\.645\\,\\hat\{\\sigma\}\)93\.224\.724\.7Quantile regression \(uncalibrated\)89\.720\.419\.3ConformalGlobal split conformal89\.117\.617\.6CQR90\.520\.819\.8Mondrian \(priority\-stratified\)89\.017\.618\.0Q2\. How well calibrated are the prediction intervals across different nominal coverage levels?

Figure 4:Calibration of empirical versus nominal coverage\. Dashed line: perfect calibration; markers: nominal90%90\\%coverage\.Matching the nominal coverage at a single operating point does not necessarily imply good calibration across other coverage levels\. We therefore evaluated empirical coverage across nominal levels ranging from50%50\\%to99%99\\%\(see Fig\.[4](https://arxiv.org/html/2610.31944#S4.F4)\)\. All three conformal strategies closely followed the ideal calibration diagonal, indicating consistent calibration across the evaluated range rather than only at the90%90\\%target level\. This flexibility allows the nominal coverage level to be adjusted according to the operational requirements of the application setting\. Furthermore, repeating the evaluation across twenty random training–calibration–test partitions yielded mean empirical coverages between89\.8%89\.8\\%and89\.9%89\.9\\%at the90%90\\%nominal level, indicating that the deviations observed for the primary split were small across repeated partitions\.

Q3\. How stable is empirical coverage across operationally relevant contexts?The exploratory analysis in Section[4\.3](https://arxiv.org/html/2610.31944#S4.SS3)showed that the ambulance travel time variability differs significantly across operational contexts\. Consequently, satisfactory marginal coverage at the aggregate level does not necessarily imply comparable empirical coverage across different types of trips\. We therefore examined empirical coverage across dispatch priorities and trip distance quartiles for the three conformal strategies \(see Fig\.[5](https://arxiv.org/html/2610.31944#S4.F5)\)\. Across dispatch priorities, all three strategies achieved empirical coverage between approximately86%86\\%and92%92\\%, indicating relatively limited variation across the evaluated priority groups\. A different pattern emerged across trip distance quartiles\. Global split conformal prediction applied a single fixed calibration margin to all trips, whereas priority\-stratified Mondrian conformal prediction applied a priority\-specific margin that remained fixed within each priority group\. Because neither strategy adapts interval width directly to trip distance, both showed overcoverage for shorter trips and undercoverage for longer trips, with empirical coverage falling to approximately83%83\\%for the longest\-distance quartile\. This pattern is consistent with heteroscedastic travel time errors, where predictive uncertainty increases with trip distance while the calibration margin remains constant globally or within each priority group\. In contrast, CQR generated trip\-specific interval widths through conditional quantile estimation before conformal calibration\. It maintained empirical coverage between88\.9%88\.9\\%and91\.6%91\.6\\%across the four trip distance quartiles, while allowing interval widths to increase for trips associated with greater predictive uncertainty\.

These results show that satisfactory marginal coverage does not necessarily imply equally stable empirical coverage across operational subgroups\. Among the three conformal strategies evaluated in the Skåne scenario study, CQR produced the most stable empirical coverage across the trip distance quartiles while retaining conformal marginal coverage under the exchangeability assumption\.

Figure 5:Empirical coverage and interval width across operational contexts at nominal90%90\\%coverage\.

## 5Discussion

Implications for AI\-Enabled Dispatch Systems\.Traditional EMS dispatch software has relied on deterministic rules or static lookup tables\. As AI\-enabled decision support methods are increasingly introduced into EMS operations, predictive accuracy alone may be insufficient for evaluating their reliability\. By producing calibrated prediction intervals, the proposed methodology makes predictive uncertainty explicit and allows empirical coverage to be evaluated against a predefined target as new operational outcomes become available\. Accordingly, calibration guarantees and empirical coverage across relevant operational contexts could complement conventional accuracy metrics when evaluating AI\-enabled dispatch support methods for safety\-critical use\.

Operational Self\-Sufficiency\.The implementation of the proposed methodology in the Skåne scenario study supports self\-hosted operation for two reasons\. First, inference relies only on dispatch\-available features and precomputed routing attributes; no additional data need to be queried at prediction time\. Second, the point prediction model selected for the scenario study, the ANN described in Section[4\.4](https://arxiv.org/html/2610.31944#S4.SS4), is lightweight enough to run on commodity hardware, as indicated by the inference latency measurements in Section[4\.4](https://arxiv.org/html/2610.31944#S4.SS4)\. External routing services are used only during offline data preparation and are not required during inference, reducing dependence on externally hosted infrastructure\. However, incorporating real\-time traffic or other contextual inputs would reintroduce external dependencies and create a trade\-off between richer predictive context and operational self\-sufficiency\.

Limitations and Human\-in\-the\-Loop Governance\.The proposed methodology is intended to support, rather than replace, human dispatchers\. Emergency dispatch remains a human\-centered decision\-making process, and uncertainty\-aware travel time estimates can provide additional information that dispatchers may consider alongside contextual knowledge not represented in the model\. For example, an implementation of the methodology cannot account for rare or unprecedented events outside the conditions represented in the training data, such as major road closures or simultaneous casualty incidents\. Uncertainty estimates could therefore be used to identify predictions associated with greater uncertainty and support discretionary review rather than directly driving automated dispatch decisions\. This preserves human oversight while allowing the proposed methodology to support routine decision\-making\. It also aligns with the European Union’s emphasis on human oversight of high\-risk artificial intelligence systems\.

Moreover, the real\-world scenario study demonstrates that the proposed methodology can be instantiated and evaluated using operational EMS data from a real regional setting\. Although the evaluation was limited to the Skåne region and the 2020–2021 observation period, the methodology is designed to support adaptation to other operational contexts through context\-specific data curation, feature configurations, point prediction models, and conformal calibration strategies\. The present retrospective evaluation provides evidence of predictive accuracy and prediction interval calibration, including reliability across operationally relevant contexts\. A natural next step is, therefore, external validation and a prospective, human\-centered evaluation of how uncertainty\-aware estimates affect dispatcher decisions and patient outcomes\.

## 6Conclusions and Future Work

In this paper, we proposed an uncertainty\-aware methodology for ambulance travel time estimation\. We evaluated the proposed methodology through a retrospective real\-world scenario study using two years of ambulance trip data from the Skåne region in Sweden\. The evaluation examined both point prediction performance and prediction interval calibration, including empirical coverage across operationally relevant contexts\. From the evaluation results, we draw three conclusions\. First, domain\-informed data curation substantially improved point prediction performance, increasingR2R^\{2\}from0\.620\.62to0\.740\.74\. Second, aggregate coverage alone is insufficient to assess reliability across operational contexts, as fixed\-margin strategies undercovered the longest trips \(approximately83%83\\%coverage\)\. Third, among the evaluated strategies, CQR provided the most stable empirical coverage across trip distance quartiles, maintaining88\.988\.9–91\.6%91\.6\\%coverage at the cost of moderately wider intervals\. These findings demonstrate the importance of evaluating both point prediction accuracy and context\-specific uncertainty when developing trustworthy ambulance travel time estimation methods for EMS decision support\.

Future work will extend the methodology by incorporating richer contextual information, such as real\-time traffic, weather, and incident\-specific characteristics\.

#### Disclosure of Interests\.

The authors have no competing interests to declare that are relevant to the content of this article\.

## References

- \[1\]Morar, A\. M\., Efleih\-Hassan, L\., & Lungeanu, D\. \(2020\)\. Apparent patterns in ambulance response time in Timisoara\. Applied Medical Informatics, 42\(1\), 28\-35\.
- \[2\]Zhao, Q\., & Vanberkel, P\. \(2025, May\)\. A Comparison of Ambulance Travel Time Approximation: Using Google Maps and Machine Learning\. In Proceedings of the International ISCRAM Conference \(Vol\. 22\)\.
- \[3\]Abid, M\. A\., Lorig, F\., Holmgren, J\., & Petersson, J\. \(2024\)\. Ambulance travel time estimation using spatiotemporal data\. Procedia Computer Science, 238, 265\-272\.
- \[4\]Fleischman, R\. J\., Lundquist, M\., Jui, J\., Newgard, C\. D\., & Warden, C\. \(2013\)\. Predicting ambulance time of arrival to the emergency department using global positioning system and Google maps\. Prehospital Emergency Care, 17\(4\), 458\-465\.
- \[5\]Westgate, B\. S\., Woodard, D\. B\., Matteson, D\. S\., & Henderson, S\. G\. \(2013\)\. Travel time estimation for ambulances using Bayesian data augmentation\. The Annals of Applied Statistics, 1139\-1161\.
- \[6\]Mahdiraji, S\. A\., Abid, M\. A\., & Holmgren, J\. \(2024\)\. Integrating machine learning\-based ambulance travel time estimation into an emergency medical services simulation modeling framework\. Procedia Computer Science, 251, 479\-486\.
- \[7\]Li, L\., Ran, B\., Zhu, J\., & Du, B\. \(2020\)\. Coupled application of deep learning model and quantile regression for travel time and its interval estimation using data in different dimensions\. Applied Soft Computing, 93, 106387\.
- \[8\]Angelopoulos, A\. N\., & Bates, S\. \(2023\)\. Conformal prediction: A gentle introduction\. Foundations and Trends in Machine Learning, 16\(4\), 494\-591\.
- \[9\]Patil, M\., Ahmed, Q\., & Midlam\-Mohler, S\. \(2024, September\)\. Urban traffic forecasting with integrated travel time and data availability in a conformal graph neural network framework\. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems \(ITSC\) \(pp\. 2482\-2487\)\. IEEE\.
- \[10\]Romano, Y\., Patterson, E\., & Candes, E\. \(2019\)\. Conformalized quantile regression\. Advances in neural information processing systems, 32\.
- \[11\]Boström, H\., Johansson, U\., & Löfström, T\. \(2021, September\)\. Mondrian conformal predictive distributions\. In Conformal and Probabilistic Prediction and Applications \(pp\. 24\-38\)\. PMLR\.