Modified-Emergency Index (MEI)

• MEI is a criticality metric that quantifies evasive effort in lateral driving conflicts by combining Interaction Depth with Time for Evasive Maneuver.
• It is computed frame-by-frame from trajectory data in urban scenarios, leveraging methods like 2D collision time computation under a constant-velocity model.
• MEI outperforms traditional metrics such as TTC and PET by accurately capturing risk evolution and supporting real-time conflict classification.

Searching arXiv for papers on "Modified-Emergency Index" and closely related index-design/triage work. Calling arXiv search. The Modified-Emergency Index (MEI) is a criticality metric introduced for autonomous driving to quantify evasive effort in lateral conflicts such as intersection crossings, turns, merges, and pedestrian interactions, where the primary risk is sideways rather than purely longitudinal (Cheng et al., 31 Oct 2025). Formally, it combines a spatial term, Interaction Depth (InDepth), with a temporal term, Time for Evasive Maneuver (TEM), so that larger values indicate larger required motion change within less remaining time (Cheng et al., 31 Oct 2025). In adjacent literature on emergency-decision support and index construction, the term also functions as a useful organizing concept for data-driven prioritization: emergency-department triage work motivates model-derived resource indices that reduce subjectivity in nurse assessment (Gligorijevic et al., 2018), while measurement-oriented work on emergency management emphasizes that any such index is a low-complexity model whose outputs can be highly sensitive to variable choice, weighting, thresholds, and spatial scale (Gosciak et al., 17 May 2026).

1. Origins and problem setting

MEI was proposed because existing surrogate safety measures primarily target longitudinal conflicts, whereas urban driving contains many lateral or multi-angle interactions, including intersection crossing, left/right turns across traffic, lane changing and merging, and pedestrian and cyclist crossings (Cheng et al., 31 Oct 2025). The underlying notion of criticality is “the composite risk involving all actors as the traffic situation evolves,” and the paper links this notion to two coupled quantities: how much change to the current motion is needed to avoid collision, and how much time remains to make that change (Cheng et al., 31 Oct 2025).

The original Emergency Index (EI), discussed as MEI’s precursor, already used Interaction Depth as a spatial measure of how deeply two vehicles will intrude into each other’s safety regions in the future if neither performs evasive maneuvers (Cheng et al., 31 Oct 2025). Its limitation lies in time estimation: EI uses a centroid-based approximation for the remaining time until collision, and this approximation can be inaccurate in critical near-miss lateral scenarios because the actual first-contact point pair between two rectangular vehicles may not correspond to the line joining the vehicle centers (Cheng et al., 31 Oct 2025). MEI preserves the spatial foundation of EI but replaces that temporal approximation with a more exact 2D collision time computation under a constant-velocity assumption (Cheng et al., 31 Oct 2025).

This emphasis on replacing subjective or approximate judgment with learned or formalized estimates has a parallel in emergency-department triage. There, the Emergency Severity Index (ESI) is described as the current standard ED triage framework, but it “greatly depends on nurse’s subjective judgment and is thus prone to human errors” (Gligorijevic et al., 2018). The source material explicitly identifies the mapping from resource categories to acuity levels as central for translating ESI into a numeric resource-based target for modeling, “a key idea for any Modified-Emergency Index” (Gligorijevic et al., 2018). In both domains, the motivating problem is similar: a coarse or heuristic priority scheme is replaced or augmented by a quantitatively defined index.

2. Mathematical definition in autonomous driving

For each vehicle $i$, the state at time $t$ is

$$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$

where $x_i(t), y_i(t)$ are center coordinates, $v_i(t)$ is speed magnitude, $\theta_i(t)$ is heading angle, and $l_i, w_i$ are the vehicle’s length and width (Cheng et al., 31 Oct 2025). From this, the position, velocity, and heading-direction vectors are defined as

$$\begin{aligned} \boldsymbol{P}_i(t) &= \begin{bmatrix} x_i(t) \ y_i(t) \end{bmatrix}, \ \boldsymbol{v}_i(t) &= \begin{bmatrix} v_i(t)\cos\theta_i(t) \ v_i(t)\sin\theta_i(t) \end{bmatrix}, \ \boldsymbol{\theta}_i(t) &= \begin{bmatrix} \cos\theta_i(t) \ \sin\theta_i(t) \end{bmatrix} = \frac{\boldsymbol{v}_i(t)}{\|\boldsymbol{v}_i(t)\|}. \end{aligned}$$

Vehicles are modeled as oriented rectangles (Cheng et al., 31 Oct 2025).

For two vehicles $A$ and $B$, the relative position and relative velocity are

$t$0

with direction of relative motion

$t$1

These quantities define the geometric frame in which the future overlap is evaluated (Cheng et al., 31 Oct 2025).

Interaction Depth is constructed from the tangential distance between centers along the direction orthogonal to relative motion,

$t$2

together with vehicle-specific projection radii derived from the four corners of each oriented rectangle (Cheng et al., 31 Oct 2025). For vehicle $t$3,

$t$4

and similarly for vehicle $t$5 (Cheng et al., 31 Oct 2025). The maximal orthogonal projections are

$t$6

With safety buffer $t$7, the paper defines

$t$8

For all reported experiments, $t$9, so the safety region coincides with the exact vehicle body (Cheng et al., 31 Oct 2025).

MEI’s temporal term is the Time for Evasive Maneuver,

$$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$0

where TTC2D computes the exact collision time between two moving oriented rectangles under the constant velocity model by identifying the actual pair of points that will collide first in the future, rather than the closest pair at the current time as in ACT (Cheng et al., 31 Oct 2025). The Modified-Emergency Index is then

$$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$1

Its units are meters per second, and a larger MEI indicates larger spatial intrusion depth within a smaller remaining time window (Cheng et al., 31 Oct 2025).

3. Interpretation, conflict taxonomy, and temporal behavior

The interpretation of MEI is explicitly tied to evasive difficulty. A high MEI corresponds to large positive InDepth and small TEM, meaning that the vehicles’ safety regions are set to overlap deeply unless a rapid and substantial change in motion occurs (Cheng et al., 31 Oct 2025). A low, near-zero, or negative MEI corresponds either to $$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$2 or to a large TEM, indicating low criticality or a manageable conflict (Cheng et al., 31 Oct 2025).

The metric is embedded in a broader conflict-classification framework using a Conflict Detection Model (CDM) and a threshold $$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$3, with the paper using $$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$4 (Cheng et al., 31 Oct 2025). The categories are:

State	Conditions	MEI status
Non-Conflict	$$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$5	No TEM/InDepth/MEI defined
Potential Conflict	$$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$6, but $$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$7 is larger or $$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$8	MEI not constrained
Critical Conflict	$$\boldsymbol{S}_i(t) = [x_i(t),\ y_i(t),\ v_i(t),\ \theta_i(t),\ l_i,\ w_i]^T,$$9, $x_i(t), y_i(t)$0, $x_i(t), y_i(t)$1	$x_i(t), y_i(t)$2
Crash	$x_i(t), y_i(t)$3, $x_i(t), y_i(t)$4, $x_i(t), y_i(t)$5	$x_i(t), y_i(t)$6

MEI is computed frame-by-frame from trajectory data. As a conflict unfolds, the paper describes a typical rise-and-fall pattern: early in the interaction, vehicles may be far apart so that InDepth is negative or small and TEM is large; as they approach a conflict zone, InDepth may become positive while TEM decreases, causing MEI to rise; if evasive maneuvers then occur, InDepth can drop or TEM can increase, causing MEI to fall (Cheng et al., 31 Oct 2025). The scenario-level criticality indicator is the maximum over time,

$x_i(t), y_i(t)$7

which the paper uses to summarize the severity of an 11-second conflict episode (Cheng et al., 31 Oct 2025).

This temporal interpretation distinguishes MEI from purely post-hoc or purely temporal measures. TTC and TTC2D are purely temporal, while PET reduces the interaction to time separation at a conflict zone; MEI retains TTC2D as the temporal component but scales it with InDepth, thereby adding a geometry-aware severity factor (Cheng et al., 31 Oct 2025).

4. Computation, data, and empirical evaluation

The reported validation uses a public lateral conflict dataset based on Argoverse-2. The base dataset contains 250,000 scenarios, each 11 seconds long at 10 Hz, plus HD maps (Cheng et al., 31 Oct 2025). Li et al. constructed a Lateral Conflict Resolution Dataset focused on high-quality lateral conflict scenes, and the MEI study further used the Separating Axis Theorem (SAT) to filter out false “collisions” due to sensor noise or annotation errors (Cheng et al., 31 Oct 2025). After filtering, the retained sample contains 1,548 conflict instances with maximum MEI $x_i(t), y_i(t)$8 involving the AV, of which 501 (32.4%) are labeled critical conflicts and 1,047 (67.6%) are labeled potential conflicts (Cheng et al., 31 Oct 2025).

The computation at each time step follows a fixed pipeline: read the two vehicle states, compute relative motion, compute InDepth from tangential distance and corner projections, compute TEM via TTC2D under the constant-velocity assumption, compute $x_i(t), y_i(t)$9 when the CDM indicates conflict, and then classify the frame using TEM and InDepth (Cheng et al., 31 Oct 2025). Sampling is 10 Hz, $v_i(t)$0, and the geometric representation is always oriented rectangles with per-vehicle length and width (Cheng et al., 31 Oct 2025).

For evaluation, the paper compares $v_i(t)$1 with $v_i(t)$2 and PET over the 1,548 conflict samples (Cheng et al., 31 Oct 2025). The authors analyze metric distributions and set percentile-based thresholds for risk levels; for example, top 1% risk according to MEI corresponds to $v_i(t)$3 (Cheng et al., 31 Oct 2025). The empirical conclusion is qualitative rather than ROC-based: MEI “consistently outperforms” ACT and PET in accurately quantifying criticality and capturing risk evolution, although the paper does not provide explicit correlation coefficients or AUC/ROC numbers (Cheng et al., 31 Oct 2025).

The case studies are central to that conclusion. In a high-risk near-miss involving an AV turning left versus a straight human-driven vehicle, PET is $v_i(t)$4, $v_i(t)$5 at $v_i(t)$6, and $v_i(t)$7 at $v_i(t)$8; MEI rises as the human-driven vehicle accelerates into the intersection and falls after it decelerates to yield, whereas ACT decreases monotonically and cannot reflect risk relief (Cheng et al., 31 Oct 2025). In a right-turn interaction with a pedestrian, ACT drops to $v_i(t)$9 while $\theta_i(t)$0; at the critical frame, $\theta_i(t)$1 but $\theta_i(t)$2, so $\theta_i(t)$3, which the paper interprets as negligible (Cheng et al., 31 Oct 2025). Additional examples show PET missing critical interactions when the second actor enters late and producing false alarms when the second actor enters only after the AV has passed (Cheng et al., 31 Oct 2025).

5. Relation to other emergency indices and triage frameworks

Although MEI is introduced in autonomous driving, the supplied literature places it within a broader family of index-based decision tools. In emergency-department triage, the ESI is a 5-level acuity scale in which levels 1–2 are highly urgent and levels 4–5 are non-urgent; for ESI levels 3–5, the nurse assigns acuity by assessing both acuity and predicted resource needs (Gligorijevic et al., 2018). The source material reconstructs the mapping between resource use and ESI approximation as 0 resources $\theta_i(t)$4 Level 5, 1 resource $\theta_i(t)$5 Level 4, 2 or 3 resources $\theta_i(t)$6 Level 3, and 4 or 5 resources $\theta_i(t)$7 Level 2, and explicitly states that this mapping is central for translating ESI into a numeric resource-based target for modeling, “a key idea for any Modified-Emergency Index” (Gligorijevic et al., 2018).

That ED work frames the problem as a prediction of resource intensity from routinely collected structured and unstructured triage data, including chief complaint, past medical history, medication list, and free-text initial nursing assessment (Gligorijevic et al., 2018). Using 338,500 ED visits over a three-year period, the proposed deep attention model achieves an AUC of $\theta_i(t)$8 for identifying resource-intensive patients and an accuracy of $\theta_i(t)$9 for predicting exact category of number of resources, with an estimated 16% lift over nurses’ performance in grouped resource prediction (Gligorijevic et al., 2018). A plausible implication is that, in healthcare settings, an MEI-like construct would function not as a conflict metric but as a model-derived prioritization score grounded in predicted resource use and attention-based interpretability (Gligorijevic et al., 2018).

The emergency-management literature introduces a different but complementary perspective: indices are “low-complexity models” that summarize many observed characteristics of spatial units into one standardized score or ranking for geographic targeting and prioritization (Gosciak et al., 17 May 2026). In the NYC heat case study, additive composite indicators, theory-based formulas, and hierarchical additive indices are all examined, and the paper emphasizes that such tools encode abstract principles and overarching priorities while remaining sensitive to reasonable design choices (Gosciak et al., 17 May 2026). This suggests that an MEI, regardless of domain, is not merely a numerical convenience; it is a measurement model whose validity depends on construct definition, variable choice, transformations, and the link between the score and downstream action (Gosciak et al., 17 May 2026).

6. Validity, sensitivity, and limitations

The autonomous-driving MEI paper identifies several scope conditions. Validation is mainly on urban lateral conflict scenarios derived from Argoverse-2, and future work is needed to correlate MEI directly with crash data and to validate cross-scenario applicability, including highways (Cheng et al., 31 Oct 2025). The constant-velocity assumption underlies TTC2D-based TEM, and MEI is evaluated only when the Conflict Detection Model indicates a conflict configuration (Cheng et al., 31 Oct 2025). The paper positions MEI as a complementary metric within a broader safety assessment framework rather than a replacement for all other criteria such as efficiency and comfort (Cheng et al., 31 Oct 2025).

The emergency-management case study sharpens these concerns in measurement-theoretic terms. It maps empirical findings to construct reliability, construct validity, and predictive validity, and shows that different reasonable choices of input variables or spatial scale can result in substantive differences to index risk scores (Gosciak et al., 17 May 2026). In the NYC heat setting, only 12% of neighborhoods keep their original HVI quintile across all four alternative specifications considered, and even when correlations are high, risk tiers can change substantially (Gosciak et al., 17 May 2026). The paper also shows weak to moderate correlations between existing heat indices and concrete outcomes such as power outages, heat-related EMS calls, and hydrant complaints, indicating limited predictive validity for some operational tasks (Gosciak et al., 17 May 2026).

Applied to MEI conceptually, these findings imply that an index should not be treated as ground truth merely because it is interpretable or low-complexity (Gosciak et al., 17 May 2026). In autonomous driving, MEI’s construct is relatively narrow—evasive effort in lateral conflicts—and its definition is explicit in geometry and time-to-collision terms (Cheng et al., 31 Oct 2025). In triage or emergency management, however, any MEI-like score would require more explicit normative choices about whether the index is intended to measure health burden, social vulnerability, exposure, adaptive capacity, or some trade-off among them (Gligorijevic et al., 2018, Gosciak et al., 17 May 2026).

A related limitation concerns thresholding and categorization. In the heat-index study, the conversion of a continuous score to quintiles is shown to collapse information and make classifications fragile near cut points (Gosciak et al., 17 May 2026). A plausible implication is that MEI thresholds used for “critical” versus “potential” conflict, or for downstream prioritization, should be interpreted as design decisions rather than immutable natural boundaries, even when they are operationally necessary (Cheng et al., 31 Oct 2025, Gosciak et al., 17 May 2026).

7. Uses and broader significance

In autonomous driving, MEI is proposed as a building block for safety assessment and benchmarking, scenario mining and test-case selection, and possibly online risk monitoring because it responds quickly to changes in trajectories and can trigger warnings or emergency interventions when it exceeds a threshold (Cheng et al., 31 Oct 2025). The paper also highlights an open-source implementation at https://github.com/AutoChengh/MEI, together with preprocessing and conflict filtering based on the Lateral Conflict Resolution Dataset (Cheng et al., 31 Oct 2025).

More broadly, the literature suggests two distinct but connected roles for an MEI. First, it can operate as a geometry-aware, time-sensitive metric tightly coupled to physical interaction, as in lateral conflict analysis (Cheng et al., 31 Oct 2025). Second, it can designate a formalized prioritization score that replaces subjective or ad hoc ranking in emergency operations, provided the construct, inputs, and validation procedures are explicit (Gligorijevic et al., 2018, Gosciak et al., 17 May 2026).

The triage literature shows how structured and unstructured data can be fused through a deep attention architecture to predict resource-intensive patients and provide token-level interpretability through attention weights, thereby supporting prioritization and disposition decisions in the waiting room (Gligorijevic et al., 2018). The emergency-management literature shows that indices are attractive because they are interpretable and can encode explicit value commitments, but that they are fragile enough to require systematic sensitivity analysis, fairness analysis, and careful separation between long-term planning and narrowly outcome-oriented prediction tasks (Gosciak et al., 17 May 2026). Taken together, these strands indicate that MEI is best understood not only as a specific ratio $l_i, w_i$0 for lateral conflicts, but also as a general design pattern for operational risk scoring: a concise index that mediates between raw multidimensional data and consequential decisions, while remaining accountable to formal definition, validation, and the limits of the construct it claims to measure (Cheng et al., 31 Oct 2025, Gligorijevic et al., 2018, Gosciak et al., 17 May 2026).