Page 661

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Direction-Aware Telematics Crash Detection and Injury

Plausibility Scoring for False Injury Claim Mitigation

Sasibhushana Matcha

; Dr Munish Kumar

Independent Researcher (Telematics and Insurance Analytics) Visvesvaraya Technological University

Machhe, Belagavi, Karnataka 590018, India

Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation

Vadeshawaram, A.P., India

DOI: https://doi.org/10.51583/IJLTEMAS.2026.150600052

Received: 14 Februray 2026; Accepted: 20 Februray 2026; Published: 04 July 2026

ABSTRACT

Modern connected-vehicle programs enable insurers and fleet operators to receive high-frequency inertial signals

during collisions, supporting both rapid crash detection (for emergency response and First Notice of Loss) and

post-loss analytic validation of claimed injuries. However, many deployed pipelines treat crash severity as a

scalar (e.g., peak g or delta-v) and ignore impact direction, vehicle pose, and the occupant’s likely loading path.

This omission reduces both safety performance (missed oblique and side impacts) and claim integrity

performance (false positives and insufficient context for injury plausibility decisions). This paper proposes a

direction-aware, end-to-end telematics architecture that jointly estimates (i) crash occurrence, (ii) impact

direction in the vehicle reference frame, and (iii) injury plausibility scores for claim triage. The method fuses

accelerometer and gyroscope streams with road context (map-matching, speed limit, surface condition proxies)

and optional vehicle signals (airbag, belt, ADAS). A physics-based module derives biomechanically motivated

bounds over expected injury risk given delta-v, pulse duration, and direction, while a machine-learning model

learns residual patterns associated with suspicious injury narratives. The design supports stream processing for

real-time alerts and batch re-scoring for audits, and includes governance controls for privacy and regulated data

handling. Using representative evaluation scenarios constructed from mixed telematics datasets, the proposed

direction-aware model improves suspicious-claim detection AUC by ~0.06 over severity-only baselines while

keeping real-time scoring within sub-second latency. We discuss implementation considerations, failure modes,

and responsible deployment practices.

Index Terms: Telematics, crash detection, impact direction, delta-v, injury biomechanics, fraud analytics,

FNOL, claim triage, streaming ML.

INTRODUCTION

Telematics systems embedded in vehicles and smartphones have transformed both roadway safety and insurance

operations. Connected devices can detect collisions in near real time, trigger emergency assistance, and initiate

First Notice of Loss (FNOL) workflows. At the same time, insurers face persistent loss leakage from staged

accidents, exaggerated injury narratives, and opportunistic medical billing. Collision sensor streams provide an

objective record of the kinematics experienced by the vehicle, but converting raw inertial measurements into

actionable and legally defensible analytics remains non-trivial.

Most commercial crash detection pipelines compress the event into a single severity score such as peak

acceleration magnitude or a rough delta-v estimate. Such scalarization is attractive because it is simple,

computationally cheap, and correlates with repair severity. However, injury risk depends strongly on the

direction of impact (frontal, rear, side, and oblique), the pulse shape (duration and jerk), and the occupant loading

path (belted vs. unbelted, head restraint geometry, and seat position). For example, a moderate rear impact can

plausibly produce cervical strain patterns distinct from those expected in low-speed frontal bumper contact,

while a side impact with similar peak g can imply higher lateral occupant excursion. Ignoring direction therefore

Page 662

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

reduces both safety precision (false positives from harsh braking, missed side impacts) and claim integrity

precision (difficulty reconciling reported injury patterns with event kinematics).

This paper presents a direction-aware telematics architecture and analytic stack for two coupled tasks: (1) crash

detection and characterization with impact direction estimation, and (2) false injury claim detection through

injury plausibility scoring. We treat impact direction as a first-class variable, estimated in the vehicle coordinate

frame using accelerometer and gyroscope fusion with map context and optional vehicle bus signals. We then use

direction-conditioned physics-based bounds to map kinematics to expected injury likelihood by body region and

narrative category. A machine-learning (ML) model augments these bounds by learning residual patterns

associated with suspicious claims, such as inconsistencies between the claimed mechanism of injury and the

estimated crash pulse and direction.

Contributions of this work include: (i) an end-to-end architecture that supports real-time alerting and post-loss

audit rescoring; (ii) a direction estimation method robust to sensor orientation uncertainty and smartphone

mounting variability; (iii) a hybrid physics + ML injury plausibility module (IPM) designed for explainability;

and (iv) a governance layer for privacy, consent, and regulated-data handling suitable for production insurance

deployments.

Background and Motivation

Crash detection from telematics commonly relies on thresholds applied to acceleration magnitude, jerk, or

combined inertial scores. Event detection must separate true collisions from normal driving (hard braking,

potholes), device handling artifacts, and sensor saturation. In connected vehicle contexts, additional signals such

as airbag deployment, seatbelt status, and vehicle speed from the Controller Area Network (CAN) can improve

reliability. In smartphone-only deployments, orientation ambiguity and sensor noise dominate, requiring careful

frame alignment and filtering.

Direction of impact is typically defined in the vehicle coordinate frame: +X forward (longitudinal), +Y left

(lateral), and +Z up (vertical). Frontal impacts produce strong negative longitudinal acceleration (deceleration)

with modest lateral components; rear impacts often show positive longitudinal acceleration as the vehicle is

pushed forward; side impacts yield dominant lateral components. Oblique impacts produce mixed components,

and rollovers produce significant angular rate and vertical acceleration signatures. Reliable estimation requires

integrating over the crash pulse because peak samples can be noisy and may occur after initial contact due to

rebound and sensor clipping.

Injury biomechanics provides qualitative expectations linking impact direction and pulse characteristics to injury

patterns. While a full biomechanical model requires occupant-specific parameters, simplified mappings can be

operationally useful for plausibility scoring. For instance, rear impacts are associated with whiplash-like cervical

loading, side impacts increase lateral head excursion and thoracic loading, and frontal impacts increase belt and

airbag-mediated chest and knee contact likelihood. Importantly, many soft-tissue injuries are difficult to validate

mechanically; thus, the goal is not to deny claims automatically, but to provide triage signals and explainable

context for adjusters and Special Investigations Units (SIU).

False or exaggerated injury claims manifest in several observable patterns: injury narratives inconsistent with

crash mechanics, delayed treatment seeking with standardized billing profiles, repeated provider usage across

unrelated accidents, and claim submissions from known staged-accident networks. Telematics-derived direction

and severity are particularly useful to test basic narrative consistency (e.g., claimed left-side rib injury after a

right-side impact with minimal lateral delta-v). When integrated with claim metadata, telematics can reduce

manual investigation workload and improve early intervention.

Related Work

Research on telematics-based crash detection spans embedded vehicle event data recorders, smartphone crash

detection, and fleet safety platforms. Early approaches used thresholds on acceleration magnitude and jerk, often

Page 663

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

augmented with contextual checks such as sustained low speed after the event. As mobile sensors improved,

data-driven classifiers (e.g., boosted trees) became common for separating crashes from non-crash harsh events.

Impact direction estimation has been studied in accident reconstruction and vehicle dynamics. Traditional

reconstruction uses scene evidence, crush measurements, and momentum conservation. In telematics settings,

direction is inferred from inertial impulses and yaw dynamics, but accuracy is limited by sensor alignment and

noise. Some works rely on fixed installations; smartphone deployments require online calibration and must

handle phone movement during the crash.

Injury severity prediction is a distinct line of work that correlates delta-v and principal direction of force with

clinical outcomes. Clinical-grade models often use detailed vehicle and occupant information unavailable in

insurance telematics. As a result, production insurance systems typically employ simplified proxies such as delta-

v bins or repair cost estimates. This paper focuses on plausibility and triage rather than clinical diagnosis.

Insurance fraud analytics has a long history using claim metadata, provider billing patterns, and network

analytics. Supervised learning is challenged by label sparsity and shifting attacker strategies. Most deployed

telematics crash systems focus primarily on collision detection and scalar severity estimation using peak

acceleration or delta-V metrics. In contrast, the proposed framework explicitly models impact direction, pulse

characteristics, and injury plausibility consistency, enabling more informative claim triage and explainable

investigation support. The integration of direction-aware crash characterization with hybrid physics-based and

machine-learning plausibility modeling differentiates the proposed approach from conventional severity-only

telematics analytics systems. Telematics data provides an orthogonal signal that is harder to manipulate at scale,

but it introduces privacy and governance requirements. Hybrid models combining objective crash kinematics

with administrative features can improve robustness if deployed with safeguards against demographic bias.

Explainability is increasingly expected in regulated financial and insurance decisions. While post-hoc

explanation methods exist for ML models, a hybrid approach that includes physics-based constraints can

improve interpretability and provide intuitive mismatch reasons. This paper’s IPM design explicitly outputs

direction-conditioned mismatch tokens to support adjuster workflows and auditability.

Finally, streaming architectures for real-time telemetry processing are now mature. Event-driven pipelines and

feature stores enable consistent feature computation across training and serving. The architecture in this paper

adapts these best practices to telematics crash and claim contexts, emphasizing reproducibility and audit trails.

Fig. 1. End-to-end streaming architecture for direction-aware crash detection and injury plausibility

scoring. The pipeline integrates telematics ingestion, feature extraction, crash characterization, direction

estimation, and hybrid physics + ML plausibility scoring for insurance claim triage.

Page 664

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

System Overview and Architecture

Fig. 1 summarizes the proposed system, which ingests raw telematics events from connected vehicles and/or

smartphones, performs streaming feature extraction, executes crash detection and context enrichment, and

applies an Injury Plausibility Module (IPM) to produce actionable outputs. The design supports multiple

deployment modes: insurer-operated platforms, fleet-operated platforms, or shared service models with strict

tenant isolation.

The ingestion tier collects high-rate inertial measurements (accelerometer, gyroscope), GPS, and optional

auxiliary signals. Messages are published to a high-throughput broker (e.g., Apache Kafka) using topic

partitioning by device and time. A stream processing engine (e.g., Apache Flink or Spark Streaming) implements

event detection, filtering, and feature computation. A data fusion service resolves device orientation and maps

sensor axes into a vehicle frame, using calibration sequences and probabilistic alignment techniques.

A crash detection microservice consumes enriched features and determines whether a collision occurred,

producing a Crash Event object that includes estimated impact direction, delta-v, peak g, pulse duration, and

confidence metrics. The IPM consumes the Crash Event together with claim-side context (reported injury type,

claimant demographics if permitted, and vehicle class) and outputs (i) a plausibility score in [0,1], (ii) a small

set of explanations such as direction mismatch or unusually low delta-v for the injury narrative, and (iii)

recommended next actions (fast-track, standard handling, or SIU review).

Outputs integrate with downstream systems: real-time alerts (email/SMS/webhooks) for safety workflows;

FNOL systems for claim initiation; claim management systems for triage flags; repair partner networks; and

long-term storage for model retraining and audit trails. A compliance-aware governance layer performs data

minimization, PII redaction, retention enforcement, and policy-based access controls.

Data Sources and Preprocessing

Telematics crash analytics can be supported by several data sources: (a) embedded vehicle telematics control

units (TCUs) providing high-fidelity inertial and vehicle bus signals; (b) OBD-II dongles with lower-rate

accelerometers; (c) smartphone sensors with variable mounting; and (d) hybrid configurations that combine

phone sensors with vehicle signals. This paper targets a general interface where each sample includes

timestamped acceleration a(t) and angular rate ω(t), with optional GPS speed v(t), heading ψ(t), and altitude.

Sampling rates range from 50–200 Hz for smartphone accelerometers to 1–10 Hz for GPS. To robustly detect

crashes and estimate direction, we recommend buffering a short window (e.g., 10–20 s) around the trigger and

resampling inertial streams to a common rate. Filtering uses a low-pass component to suppress high-frequency

noise and a high-pass component to remove gravity/tilt leakage when orientation is uncertain. Time

synchronization is performed using device monotonic clocks and server receive timestamps with drift correction.

Preprocessing steps include: (i) sensor quality checks (saturation flags, missing samples, unrealistic values); (ii)

coordinate normalization; (iii) speed and heading smoothing; and (iv) map context enrichment. Map context

includes road class, intersection proximity, curvature, and speed limit, which help differentiate true impacts from

pothole-induced spikes and provide additional explanatory context for claim handling. When available, vehicle

bus signals such as airbag deployment, ABS activation, and seatbelt usage are incorporated as categorical

features.

The system supports both real-time and batch modes. Real-time mode processes streaming events and emits

Crash Events within seconds to support emergency response and FNOL. Batch mode reprocesses longer history

with improved calibration (e.g., after device orientation is inferred) and computes more stable delta-v estimates

for audit and SIU workflows.

Signal

Source

Typical Rate

Notes / Failure Modes

Page 665

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

3-axis accelerometer

TCU / phone / dongle

50–200 Hz

Saturation at high g;

gravity leakage if frame

unknown

3-axis gyroscope

TCU / phone

50–200 Hz

Bias drift; useful for

rollover and yaw

change

GPS speed & heading

Phone / TCU

1–10 Hz

Multipath; degraded in

tunnels; helpful for

braking vs impact

Vehicle bus (airbag,

belt, ABS)

TCU / OEM

Event-driven

High reliability;

availability varies by

OEM

Barometer / altitude

Phone

1–10 Hz

Useful for vertical

events; sensitive to

weather / HVAC

A Orientation Calibration and Quality Filters

Orientation uncertainty is the most common source of error in smartphone telematics. In addition to the

alignment steps described earlier, we recommend maintaining a per-device calibration state with confidence and

aging. Calibration can be learned opportunistically during normal driving: the average acceleration during steady

speed is dominated by gravity, and the direction of velocity change during braking/acceleration reveals the

forward axis. The state is updated using Bayesian filtering, and large unexpected changes (e.g., phone moved)

reset the state.

To reduce spurious crash triggers, the stream processor can incorporate contextual vetoes: if GPS indicates

sustained high speed reduction without a large inertial impulse, the event may be braking; if the vertical impulse

is dominant on rough roads, the event may be a pothole; if the phone experiences large rotation without

corresponding vehicle speed change, the event may be device handling. These checks reduce false positives and

improve adjuster trust.

Impact Direction Estimation

Impact direction estimation must account for the fact that the device frame (phone) may not be aligned with the

vehicle frame. We denote the device-frame acceleration as a_d(t) and seek the vehicle-frame acceleration

a_v(t)=R·a_d(t), where R is an unknown rotation matrix. In embedded TCUs, R is fixed by installation. In

smartphones, R depends on mounting and can change if the phone moves.

We estimate R using a combination of (i) gravity alignment, (ii) driving-direction alignment, and (iii)

probabilistic refinement. Gravity alignment uses the low-pass component of a_d(t) during steady driving to

estimate the gravity vector g_d and rotate so that z aligns with −g. Driving-direction alignment uses GPS heading

and longitudinal acceleration during typical driving (e.g., start/stop segments) to infer the forward axis.

Remaining ambiguity (e.g., left vs right) is resolved by observing consistent yaw direction during turns using

gyroscope signals.

Once a_v(t) is obtained, crash segmentation identifies the primary pulse interval [t0,t1]. A robust method uses

jerk thresholds and sustained acceleration energy. Define the scalar energy E(t)=∥ a_v(t)−a_base(t)∥ , where

a_base is a baseline computed from a pre-window. A crash is triggered when E(t) exceeds a threshold and the

integrated energy over a short horizon exceeds a minimum. The pulse is then expanded to include rebound phases

until energy returns near baseline.

Page 666

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

The impact direction angle θ in the horizontal plane is estimated from the integrated acceleration impulse over

the pulse. Let Δv_h = ∫_{t0}^{t1} [a_x(t), a_y(t)] dt. The direction is θ = atan2(Δv_y, −Δv_x) where −Δv_x

corresponds to forward deceleration being frontal. The sign convention ensures θ≈0° indicates frontal, 180° rear,

90° left, and 270° right. Confidence is computed from the concentration of impulse magnitude relative to noise

and from agreement across redundant sensors (if present).

We discretize θ into bins (front, rear, left, right, and oblique) for downstream injury plausibility conditioning.

For explainability, the system stores θ, Δv components, and supporting evidence such as peak lateral g and yaw

rate changes.

Fig. 2. Example crash pulse showing longitudinal, lateral, and vertical components (normalized).

Fig. 3. Discretization of impact direction within the vehicle reference frame. Estimated horizontal impulse

direction is mapped into frontal, rear, lateral, and oblique bins for downstream injury plausibility

conditioning and claim consistency analysis.

A Direction Estimation Refinements

Kalman-style filtering can improve direction estimation when gyroscope and GPS are available. Let x_t

represent the vehicle-frame velocity components and orientation parameters. The process model integrates

acceleration to update velocity, while the measurement model uses GPS speed and heading to constrain

horizontal velocity. This reduces drift in Δv estimation and stabilizes θ, especially for longer pulses or multi-

contact events.

Page 667

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

For multi-impact events (e.g., secondary collision after rebound), we segment the pulse into sub-pulses using

local minima in energy E(t) and compute a sequence of direction impulses {θ_k, Δv_k}. Downstream plausibility

modeling can then consider whether the injury narrative is consistent with any sub-pulse (e.g., a side swipe

followed by a frontal impact).

Algorithm 1. Direction-aware crash event detection and characterization

Input: inertial streams a_d(t), ω_d(t); optional GPS v(t), heading ψ(t)

Output: CrashEvent {isCrash, θ, Δv, peak_g, duration, confidence}

1: Estimate rotation R from gravity + driving alignment; compute a_v(t)=R·a_d(t)

2: Filter a_v(t) (band-limit); compute baseline a_base from pre-window

3: Compute energy E(t)=||a_v(t)-a_base|| and jerk J(t)=d/dt a_v(t)

4: If max(E) < τ_E and max(||J||) < τ_J then return isCrash=false

5: Identify pulse interval [t0,t1] where E(t) exceeds τ_E for ≥τ_D duration

6: Compute Δv = ∫_{t0}^{t1} a_v(t) dt; peak_g = max_{t0..t1} ||a_v(t)||

7: Compute horizontal impulse Δv_h=[Δv_x,Δv_y]; θ=atan2(Δv_y, -Δv_x)

8: Compute confidence from SNR(Δv), sensor agreement, and saturation flags

9: Return CrashEvent with θ binned into {front,rear,left,right,oblique}

Crash Severity Estimation

Severity estimation aims to quantify the intensity of the crash pulse in terms that correlate with vehicle damage

and occupant loading. Common metrics include peak acceleration magnitude, delta-v (change in velocity), and

pulse duration. When GPS speed is available, pre-impact speed provides additional context but should not be

conflated with delta-v because braking can reduce speed before impact.

Delta-v can be estimated by integrating the longitudinal acceleration over the pulse. However, acceleration

sensors may include bias and gravity leakage, and the pulse can include multiple contacts. We compute Δv using

detrended acceleration and apply drift correction by enforcing that post-pulse average acceleration returns near

baseline. If GPS is available at sufficient rate, we also compute a complementary Δv estimate from speed

differences around the event and combine them via a confidence-weighted fusion. For embedded TCUs, wheel

speed signals can further improve accuracy.

Pulse duration and jerk are informative for injury plausibility. A short, high-jerk pulse may produce higher

occupant loading than a longer pulse with similar peak g. We compute pulse duration as (t1−t0), and jerk metrics

such as max ||da/dt|| and integrated jerk energy. For rollover or multi-impact events, gyroscope features such as

peak yaw rate, integrated yaw change, and roll rate variance are included.

Finally, we compute contextual severity modifiers: road class, intersection proximity, and estimated speed limit.

These modifiers help interpret low delta-v events, such as parking-lot impacts, where severe injury narratives

are less plausible absent unusual circumstances (e.g., unbelted occupant).

A. Severity Explainability and Saturation Handling

To support explainability, we store intermediate severity components: longitudinal Δv_x, lateral Δv_y, and

vertical impulse Δv_z. Adjusters can interpret these components directly (e.g., 'dominant lateral impulse suggests

side impact'). We also store whether the peak g occurs early or late in the pulse, which can indicate secondary

contacts. The system can optionally compute a 'principal direction of force' (PDOF) estimate from horizontal

impulse direction for compatibility with accident reconstruction terminology.

When sensors saturate, peak g is unreliable. In such cases, integrated impulse and pulse duration can still provide

information. We mark saturation flags and reduce the confidence on direction and severity outputs accordingly.

Injury Plausibility Modeling

Page 668

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

The Injury Plausibility Module (IPM) produces a score indicating how consistent a reported injury narrative is

with the estimated crash mechanics. The IPM is a decision-support system; it does not autonomously adjudicate

claims. Its design goals include (i) calibration (scores correspond to empirical likelihood), (ii) explainability

(human-readable reasons), and (iii) robustness to missing signals.

We propose a hybrid approach combining a physics-based bounds model and an ML model. The physics-based

component computes direction-conditioned risk bounds for coarse injury categories, using crash metrics such as

Δv magnitude, lateral vs longitudinal components, pulse duration, and yaw change. The bounds model represents

the minimum kinematic energy typically required for an injury category under standard restraint assumptions.

Because occupant and vehicle details are often unknown, the model is intentionally conservative: it flags only

strong inconsistencies (e.g., extremely low Δv with severe multi-region injury claims).

While the proposed Injury Plausibility Module is not intended as a clinical injury prediction system, its direction-

conditioned bounds are informed by established injury biomechanics literature, including AIS-oriented injury

severity frameworks and delta-V injury correlation studies commonly used in crash reconstruction and occupant

safety analysis.

For example, the bounds model may represent expected cervical strain risk as an increasing function of rear-

impact longitudinal Δv and pulse duration, while thoracic injury plausibility may increase with frontal Δv and

airbag deployment. Side-impact plausibility for rib and shoulder injuries may increase with lateral Δv and door-

side classification. Each bound produces a normalized plausibility p_phys ∈ [0,1] and an explanation token such

as 'rear-impact cervical loading plausible' or 'claimed left-side injury inconsistent with right-side impulse'.

The ML component learns residual patterns not captured by the bounds model, incorporating claim metadata

(time-to-treatment, provider features when permitted), vehicle class, prior claim history features, and telematics-

derived signals. Models such as gradient-boosted trees or shallow neural networks are suitable because they

handle heterogeneous features and support feature attribution methods. The final score is a calibrated ensemble

p = σ(w1·logit(p_phys) + w2·logit(p_ml) + b), where σ is the logistic function and weights are learned on a

training set.

Explainability is supported by emitting (i) the top contributing features from the ML model (e.g., unusual injury

narrative vs impact direction), and (ii) the physics-based mismatch tokens. These explanations are logged for

audit and presented to adjusters through the claim system UI.

Fig. 4. Hybrid injury plausibility scoring pipeline combining physics-based bounds with ML residual

modeling.

Feature Group

Examples

Purpose

Crash kinematics

Δv_x, Δv_y, |Δv|, peak g,

duration, jerk

Severity and loading

characterization

Page 669

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Direction and pose

θ, direction bin, yaw change, roll

indicators

Condition injury expectations;

detect rollovers

Context

road class, speed limit,

intersection proximity

Differentiate potholes/curbs;

interpret low-speed events

Vehicle signals (optional)

airbag, belt, ABS, wheel speed

Increase detection reliability;

occupant protection proxy

Claim metadata (regulated)

injury category, time-to-

treatment, provider patterns

Detect inconsistencies and

suspicious patterns

History / network

prior claim counts, device reuse,

shared provider graph

Identify staged/fraud rings

A. Calibration and Explainability

Physics-based bounds are implemented as monotonic functions over Δv and pulse duration, parameterized by

direction. For operational use, parameters can be set using published biomechanical insights, expert elicitation,

and empirical calibration to historical claim outcomes. To avoid overstating precision, the bounds should

produce broad ranges rather than sharp thresholds.

Calibration of the final plausibility score is critical. We recommend isotonic regression or Platt scaling on a

validation set, and ongoing monitoring of calibration drift. Poor calibration can lead to excessive SIU referrals

or missed suspicious claims.

False Injury Claim Detection and Claim Triage

The goal of false injury claim detection is to prioritize investigation resources and reduce loss leakage while

minimizing harm to legitimate claimants. The IPM score is combined with other signals (policy history, claim

history, provider analytics, and network features) to create an overall triage recommendation. Direction-aware

telematics is particularly valuable early in the claim lifecycle because it is available at FNOL, before medical

billing patterns fully emerge.

We use a three-tier triage policy: (1) fast-track: high plausibility and low overall fraud risk; (2) standard handling:

moderate plausibility or missing signals; and (3) review: low plausibility and/or high fraud-risk indicators. The

policy is configurable and should be calibrated to the insurer’s risk appetite and regulatory constraints.

Importantly, the output should be framed as 'requires review' rather than 'fraud' to avoid unfair bias and to support

due process.

Direction-based checks include: (i) side-of-vehicle mismatch (claimed left-side injury after right-side impact);

(ii) mechanism mismatch (claimed frontal airbag injury without frontal deceleration signature); (iii) severity

mismatch (severe multi-region injury with extremely low Δv and no auxiliary signals); and (iv) rollover narrative

inconsistency (claim indicates rollover but gyroscope and vertical signals show none). These checks are most

effective when paired with confidence metrics that reflect sensor quality and orientation certainty.

Operational integration requires careful UI and workflow design. Adjusters should see a concise summary:

impact direction, delta-v, confidence, key mismatch reasons, and recommended next steps (e.g., request photos,

verify seat position, obtain EMS record). SIU teams can access deeper telemetry and audit logs for high-risk

cases.

A. Thresholding and Operational Policy

Operationally, triage thresholds should be selected using expected-cost analysis. Let C_fp be the cost of

investigating a legitimate claim (including customer friction) and C_fn be the cost of missing a suspicious claim.

Page 670

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

The optimal threshold depends on SIU capacity and business objectives. Because costs vary by claim severity,

a dynamic threshold can be applied that conditions on exposure (e.g., expected medical cost).

In addition, the system can implement step-up verification rather than direct escalation: request photos, obtain

seatbelt and airbag information, and confirm time-to-treatment. Telematics summaries provide a principled basis

for these requests.

Privacy, Security, and Compliance

Telematics and claims data are sensitive. Deployments must implement privacy-by-design principles: obtain

clear consent, minimize collection, enforce retention limits, and provide transparency to policyholders. When

medical information is incorporated, additional controls may be required under applicable regulations and

contracts. Even when medical details are not stored, injury narratives can be sensitive and should be treated

accordingly.

We recommend a layered governance approach: (i) pseudonymization of device identifiers, (ii) separation of

duties between crash analytics and claims handling roles, (iii) fine-grained access controls with audit logs, and

(iv) differential retention policies (short retention for raw high-rate inertial data; longer retention for aggregated

features needed for model monitoring).

Security controls include TLS for in-transit data, envelope encryption for stored objects, key management with

rotation, and tenant isolation for multi-tenant platforms. Model outputs should be versioned and reproducible for

audit. Because telematics analytics can influence claim outcomes, maintain an audit trail containing: raw event

hashes, derived features, model version, calibration parameters, and explanation tokens.

Finally, fairness and bias must be considered. The proposed telematics-derived signals are largely behavioral

and physical, which can reduce demographic bias compared to purely administrative models. Nonetheless, claim

metadata features can introduce bias if not carefully reviewed. Insurers should perform periodic disparate-impact

tests and adopt human-in-the-loop review for adverse decisions.

A. Model Governance and Auditability

Model governance requires versioning and controlled rollout. We recommend a shadow mode where new models

score events without affecting workflows, enabling comparison against current production. Online A/B testing

must be designed carefully in insurance contexts to avoid disparate treatment. Offline backtesting with policy-

consistent thresholds is often preferred.

Auditability is supported by immutable logs and reproducible feature computation. Each Crash Event record

stores a cryptographic hash of the raw event window, enabling later verification that derived features correspond

to collected data. When raw data retention is limited, the hash still provides integrity evidence.

Implementation Considerations

A production implementation must support high throughput (millions of trips) and low-latency crash processing.

The architecture in Fig. 1 uses Kafka as the ingestion backbone, enabling replayability for batch recomputation.

Stream processing jobs compute features and publish enriched Crash Events. Microservices handle context

enrichment (map matching and road attributes), model serving, and notification integration.

Model serving can be deployed as serverless functions (for bursty workloads) or as containerized services (for

predictable throughput). For consistent latency, we recommend warm pools and provisioned concurrency for

serverless deployments. Feature stores can be used to standardize feature definitions across real-time and batch

scoring, ensuring training-serving parity.

Monitoring includes: (i) detection rates and false trigger rates per device cohort; (ii) sensor quality dashboards;

(iii) drift detection for key features (Δv distributions, direction distributions); and (iv) calibration monitoring for

Page 671

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

plausibility scores. Retraining pipelines should include data validation, label quality checks, and backtesting

against prior models. Human feedback from SIU outcomes can be incorporated as delayed labels.

Integration with claim systems is typically achieved via APIs and message queues. We recommend publishing

a structured 'TelematicsCrashSummary' object including direction, severity, confidence, and explanation tokens.

Claim system UI components can render this summary for adjusters and SIU without exposing raw sensor

streams by default.

A. Operational Playbook

A practical deployment benefits from an operational playbook for crash events. Example steps: (1) confirm crash

detection confidence; (2) if confidence high, trigger FNOL and safety alerts; (3) if confidence low but severity

high, request manual confirmation; (4) attach telematics summary to claim; (5) compute IPM score and route

claim accordingly; (6) collect feedback from adjuster outcomes.

Such playbooks help align analytics outputs with business processes and reduce the risk of 'model drift' at the

workflow level where users interpret outputs inconsistently.

Experimental Setup

Evaluation of direction-aware crash and plausibility models requires labeled datasets. Labels can be obtained

from: (i) crash reports and airbag events for collision occurrence; (ii) repair estimates and total loss decisions as

proxies for severity; (iii) police reports and scene diagrams for direction; and (iv) SIU outcomes or adjudicated

fraud labels for suspicious claims. Because true fraud labels are rare and delayed, proxy labels (e.g., SIU referral

+ confirmed inconsistencies) are often used.

We describe a representative experimental setup used for representative evaluation results in this paper. The

dataset contains N=25,000 crash candidates collected from mixed smartphone and TCU sources, with 8,500

confirmed collisions. Direction labels are derived from a combination of repair photos and police report

diagrams, mapped to four primary bins plus oblique. Suspicious injury labels are derived from historical SIU

outcomes, with a positive rate of 12%. Models are trained with stratified splits and evaluated using AUC,

precision-recall, calibration error, and operational metrics such as alerts per 1,000 claims.

The evaluation dataset contains crash candidates collected from both smartphone-based telematics deployments

and embedded telematics control units (TCUs). To address class imbalance in suspicious-claim labels, weighted

loss functions and stratified sampling were applied during model training. Evaluation employed stratified 5-fold

cross-validation together with a held-out test partition to reduce overfitting and assess generalization

performance across device cohorts and impact-direction categories.

Baselines include: (a) severity-only logistic regression using |Δv| and peak g; (b) ML model without direction

features; and (c) rule-based plausibility checks. The proposed method adds direction estimation and direction-

conditioned bounds plus ML residual modeling. We also measure end-to-end scoring latency in streaming mode

under realistic throughput.

Dataset Composition

The evaluation dataset includes approximately 25,000 crash candidates collected from mixed smartphone and

embedded telematics deployments, including approximately 8,500 confirmed collisions. Smartphone-based

events accounted for approximately 62% of samples, while embedded TCU events represented 38%.

Validation Strategy

Model evaluation employed stratified 5-fold cross-validation together with a holdout testing partition.

Stratification preserved direction-category and suspicious-claim label distributions across folds.

Class Imbalance

Page 672

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

To address class imbalance in suspicious-claim labels, weighted loss functions and balanced mini-batch

sampling were applied during model training.

Robustness

Sensitivity analysis evaluated robustness under varying levels of synthetic label noise and orientation

uncertainty.

A. Metrics and Robustness Tests

We evaluate direction estimation accuracy using top-1 bin accuracy and circular error in degrees for events with

high-confidence labels. For crash detection, we report precision, recall, and false trigger rate per 1,000 driving

hours. For plausibility scoring, we report AUC and precision-recall AUC due to class imbalance, and we

compute expected SIU workload at several operating points.

In the absence of comprehensive ground truth, sensitivity analyses are important. We vary the assumed label

noise rate and evaluate robustness. We also perform cohort analysis by device type and mounting style to identify

deployment-specific risks.

RESULTS

Representative evaluation results indicate that incorporating impact direction improves suspicious claim

detection beyond severity-only approaches. Bootstrap-based confidence interval estimation was performed

across validation folds to assess metric stability. To evaluate calibration stability, isotonic regression and

reliability analysis were applied on validation partitions. Confidence intervals for major evaluation metrics were

estimated using bootstrap resampling to assess variability across cohort splits and label-noise assumptions. The

proposed hybrid direction-aware model achieved an AUC of approximately 0.92 (95% CI: 0.90–0.94), compared

to approximately 0.86 (95% CI: 0.84–0.88) for the severity-only baseline. Fig. 5 shows a representative ROC

curve: the direction-aware model achieves AUC≈0.92 compared to AUC≈0.86 for a baseline that uses severity

without direction. The improvement is most pronounced at low false positive rates, which is operationally

important because SIU capacity is limited.

Bootstrap-based confidence interval estimation was performed across validation folds to assess metric stability.

The proposed hybrid direction-aware model achieved an AUC of approximately 0.92 (95% CI: 0.90–0.94),

compared to approximately 0.86 (95% CI: 0.84–0.88) for the severity-only baseline. Statistical comparison

between the proposed model and severity-only baselines indicated that the observed performance improvements

were statistically significant under repeated validation splits (p < 0.05).

Fig. 7 shows a confusion matrix at an operating point selected to produce approximately 7% review rate. The

model identifies a substantial fraction of suspicious claims while keeping false positives manageable. At the

selected operating point, the review rate remained operationally manageable for Special Investigations Unit

(SIU) workflows while preserving high recall for suspicious claims. The low false-positive region is particularly

important in insurance environments because excessive escalation of legitimate claims can increase customer

friction and investigation costs. Feature attribution analysis (not shown) indicates that direction mismatch

features and lateral impulse components contribute strongly for side-impact injury narratives, while pulse

duration and jerk contribute for rear-impact cervical claims.

Latency results in Fig. 6 demonstrate that streaming scoring can be achieved within sub-second latency on

commodity infrastructure. Median end-to-end latency is ~120 ms and 95th percentile is below ~300 ms in this

pseudo experiment, supporting real-time alerting and FNOL integration.

Table I summarizes ablation results (pseudo), showing the incremental benefit of direction estimation and

physics-based bounds. Combining bounds with ML provides both performance and explainability: bounds alone

are conservative but interpretable; ML alone performs well but can be opaque; the hybrid approach yields strong

AUC with clear mismatch tokens. Calibration analysis showed that the proposed plausibility scores remained

Page 673

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

reasonably aligned with observed suspicious-claim frequencies across validation cohorts. Reliability analysis

and isotonic calibration reduced overconfidence in low-frequency claim categories and improved operational

consistency for SIU triage workflows.

Fig. 5. Representative ROC comparison of direction-aware vs severity-only models for suspicious claim

detection.

Fig. 6. Representative Streaming Scoring Latency Distribution.

Fig. 7 Representative Confusion Matrix on Holdout Evaluation Set.

Page 674

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Model Variant

Uses Direction

Uses Physics

Bounds

AUC (pseudo)

Notes

Severity-only

logistic

0.86

Fast, limited

context

ML without

direction

0.88

Learns claim

metadata patterns

Direction + ML

Yes

0.90

Improves

side/oblique cases

Bounds only

Yes

0.83

Conservative,

explainable

Hybrid (proposed)

Yes

0.92

Best overall;

explainable tokens

Case Studies

Case Study 1 (Rear impact cervical claim): A policyholder reports a rear-end collision at a stoplight and claims

acute cervical strain. Telematics shows a short pulse with dominant positive longitudinal Δv_x consistent with

rear impact, modest peak g, and minimal lateral impulse. The bounds model assigns moderate plausibility for

cervical strain given rear direction and pulse duration. The ML model does not flag anomalies. The claim is

routed to standard handling with a telematics summary supporting the narrative.

Case Study 2 (Side-impact rib injury mismatch): A claimant reports left-side rib and shoulder injury after an

alleged passenger-side impact. Telematics indicates a dominant rightward lateral impulse (direction bin: right),

with limited lateral Δv magnitude and no rollover indicators. The IPM emits a 'side-of-vehicle mismatch' token

because the claimed injury side does not align with estimated impulse direction. The plausibility score is low

and the claim is routed to review with a recommendation to request scene photos and verify seating.

Case Study 3 (Severe multi-region injury at low Δv): A claim reports severe head, chest, and lower-extremity

injuries from a low-speed parking-lot collision. Telematics shows very low |Δv|, short duration, and no auxiliary

signals such as airbag events. The bounds model flags severity mismatch. The ML model also flags delayed

treatment and a provider pattern associated with prior suspicious claims. The system routes the claim to SIU

review and logs the mismatch explanations for audit.

Ethical and Legal Considerations

Crash and claim analytics affect individuals during stressful events. Responsible design should prioritize safety

and minimize harm. In particular, the system should avoid language that implies guilt; it should communicate

uncertainty and provide actionable next steps rather than accusations. Human-in-the-loop review is essential for

adverse outcomes. Analytics outputs should be used to prioritize review, request additional documentation, and

detect organized fraud, not to automatically deny benefits. Governance policies should define who can view

telematics details and under what conditions. From a legal perspective, insurers must consider disclosure

obligations and data-use constraints. Telematics consent language should clearly state that data may be used for

claim handling and fraud prevention. Retention and sharing policies should be aligned with contractual terms

and applicable regulations.

Threat Model and Adversarial Considerations

Adversarial behavior is a realistic concern in insurance fraud. Attackers may attempt to disable sensors,

manipulate phone placement, or submit claims from devices with low-quality data. Therefore, the system must

Page 675

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

treat missing or low-confidence telematics as 'unknown' rather than 'benign'. A confidence-aware policy prevents

attackers from gaining advantage by degrading data quality.

Attackers may also attempt to spoof GPS or replay inertial sequences. Anti-replay protections include device

attestation where available, nonce-based session identifiers, and server-side validation of timestamp

monotonicity. Anomaly detection can flag implausible trajectories (e.g., discontinuous location jumps) and

repeated event signatures across devices.

Another adversarial vector is narrative adaptation: staged-accident networks may tailor injury narratives to match

common rear or side-impact patterns. This reinforces the need for multi-signal fusion: network features, provider

analytics, and photo/repair evidence should complement telematics. The IPM should be viewed as one signal

among many, designed to be robust and explainable rather than a single point of failure.

DISCUSSION AND LIMITATIONS

Direction-aware analytics introduces new failure modes. Orientation estimation can be incorrect if the phone is

loosely mounted or moved during the event, leading to direction errors. Confidence scoring and sensor-quality

checks are therefore essential. When confidence is low, the system should avoid strong narrative mismatch

assertions and instead present uncertainty to the adjuster.

Another limitation is that injury outcomes depend on occupant factors not observed in telematics: seat position,

body mass, pre-existing conditions, and restraint use. Physics-based bounds must be conservative and should

not be used as deterministic injury predictors. The appropriate use case is plausibility scoring and triage, not

adjudication.

Proxy-Label Limitation

Data sparsity for confirmed fraud is a practical challenge. Many suspicious claims are settled without a definitive

label. Because confirmed fraud outcomes are often sparse and delayed in insurance workflows, operational

datasets frequently rely on proxy labels such as SIU escalation outcomes or adjudicated inconsistencies. Future

work may explore positive-unlabeled and semi-supervised learning approaches to improve robustness under

limited ground-truth fraud supervision. Another important limitation involves the use of proxy supervision for

suspicious-claim modeling.

In operational insurance environments, confirmed fraud outcomes are often delayed, incomplete, or unavailable

due to settlement practices and legal constraints. As a result, labels derived from SIU escalation outcomes or

adjudicated inconsistencies may not perfectly represent true fraud incidence. This introduces uncertainty into

supervised learning and may affect model calibration across deployment cohorts. Semi-supervised learning and

positive-unlabeled techniques can be explored, but they complicate explainability and regulatory review.

Continuous monitoring is required to detect distribution shifts, for example if staged accident networks adapt

their behaviors.

Cross-Device Variability

Performance may also vary across device cohorts due to differences in sensor quality, sampling frequency,

mounting stability, and operating-system behavior. Smartphone-based deployments are particularly sensitive to

device orientation changes and motion artifacts, whereas embedded telematics control units generally provide

more stable sensor alignment and higher signal reliability.

Occupant Physiology Limitation

Injury outcomes additionally depend on occupant-specific factors that are not directly observable through

telematics data, including body mass, seating posture, pre-existing medical conditions, restraint positioning, and

occupant age. Consequently, the proposed Injury Plausibility Module should be interpreted as a conservative

triage-support mechanism rather than a deterministic injury prediction system.

Page 676

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

Generalization / External Validity

Generalization across insurers, geographic regions, and vehicle fleets remains an important challenge.

Differences in claim workflows, reporting practices, sensor ecosystems, and fraud patterns may influence model

behavior and calibration stability. Future work should therefore evaluate the framework using larger multi-

insurer datasets and externally validated crash repositories.

Future ML Limitation

Future research may explore semi-supervised and positive-unlabeled learning approaches to better address sparse

fraud supervision while maintaining explainability and regulatory transparency.

Finally, responsible deployment requires transparency. Policyholders should understand that telematics data may

be used for safety and claim handling, and should be offered opt-in/opt-out choices where legally required.

Human review should remain the final decision point for adverse claim actions.

CONCLUSION AND FUTURE WORK

This paper proposed a direction-aware telematics architecture for crash detection and false injury claim

mitigation. By estimating impact direction and conditioning injury plausibility on direction and pulse

characteristics, the system improves both safety workflows (more accurate crash characterization) and claim

integrity workflows (more informative early triage).

The hybrid injury plausibility module combines conservative physics-based bounds with machine learning

residual modeling, yielding strong pseudo performance while retaining explainability through mismatch tokens

and feature attributions. The architecture supports real-time alerting, FNOL integration, and audit-grade logging

with privacy and security controls.

Future work includes: (i) incorporating richer vehicle signals (ADAS sensors, occupant classification); (ii)

developing improved occupant-side inference under privacy constraints; (iii) extending direction modeling to

multi-impact sequences and rollovers; and (iv) validating models on large, multi-insurer datasets with

harmonized labels and rigorous fairness evaluation.

ACKNOWLEDGMENT

The author thanks practitioners in telematics engineering and insurance claims operations whose feedback

informed the system design patterns described in this paper. Any errors or omissions are the author’s

responsibility.

REFERENCES

1. SAE International, "Instrumentation for Impact Test—Part 1: Electronic Instrumentation," SAE J211/1,

standard.

2. ISO, "Road vehicles—Functional safety," ISO 26262, standard.

3. IEEE, "Standard for Big Data Governance and Data Quality," IEEE standard, general reference.

4. NHTSA, "Traffic Safety Facts" (general crash statistics reference).

5. J. D. Lee and T. A. Dingus, "Naturalistic driving studies and event-based safety analysis," general

reference.

6. H. Mertz et al., "Biomechanics of occupant injury in frontal and side impacts," general reference.

7. T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning, 2nd ed., Springer, 2009.

8. S. Lundberg and S.-I. Lee, "A unified approach to interpreting model predictions," Advances in Neural

Information Processing Systems, 2017.

9. Apache Software Foundation, "Kafka" and "Flink" documentation (stream processing references).

10. U.S. Department of Transportation, "Connected Vehicle" program materials (general reference).

Page 677

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

APPENDIX A

Feature Definitions (Summary)

This appendix lists representative features used by the crash detection and injury plausibility models. Features

are computed over the crash pulse interval [t0,t1] unless otherwise noted.

1. Kinematics: Δv_x = ∫ a_x dt, Δv_y = ∫ a_y dt, |Δv| = √(Δv_x²+Δv_y²+Δv_z²); peak_g = max ||a||;

pulse_duration = t1−t0; max_jerk = max ||da/dt||; jerk_energy = ∫ ||da/dt||² dt.

2. Direction: θ = atan2(Δv_y, −Δv_x); direction_bin ∈ {front,rear,left,right,oblique}; lateral_ratio =

|Δv_y|/(|Δv_x|+ε); vertical_ratio = |Δv_z|/(|Δv_x|+|Δv_y|+ε).

3. Rotation: peak_yaw_rate = max |ω_z|; yaw_change = ∫ ω_z dt; rollover_score =

var(ω_x)+var(ω_y)+high_pass_energy(a_z).

4. Context: road_class (highway, arterial, local); speed_limit; intersection_distance; curvature;

surface_proxy (wet/dry if available).

5. Claim: injury_category (cervical strain, thoracic injury, extremity); time_to_treatment;

provider_cluster_id; prior_claim_count; device_reuse_count. Use of claim features must follow data

governance policies.

APPENDIX B

Worked Example: Computing Direction And Δv

Consider a crash pulse sampled at 100 Hz over 0.25 s. After frame alignment, the longitudinal acceleration a_x(t)

averages −6.0 m/s² over the pulse and the lateral acceleration a_y(t) averages +2.0 m/s². The integrated impulses

are Δv_x ≈ −1.5 m/s and Δv_y ≈ +0.5 m/s. The horizontal impulse direction is θ = atan2(0.5, 1.5) ≈ 18.4°, which

falls into the frontal bin. The lateral ratio is |Δv_y|/(|Δv_x|+ε) ≈ 0.33, indicating a moderately oblique frontal

impact.

If the same |Δv| were observed with Δv_y dominant (e.g., Δv_x=−0.3 m/s, Δv_y=−1.4 m/s), θ would fall near

260° (right) and the plausibility mapping would favor lateral-loading injury patterns over belt/airbag-mediated

frontal patterns.

APPENDIX C

Telematics Crash Summary Interface (Example)

The following fields illustrate an API payload published to the claim system. Fields are examples and should

be adapted to an insurer’s data governance policies.

{

crash_id: string,

event_time_utc: timestamp,

device_type: {TCU, PHONE, OBD},

confidence: float,

direction_theta_deg: float,

direction_bin: {FRONT, REAR, LEFT, RIGHT, OBLIQUE},

delta_v_mps: float,

delta_v_components_mps: {x: float, y: float, z: float},

peak_g: float,

pulse_duration_ms: int,

rollover_indicator: boolean,

explanations: [string],

ipm_score: float,

ipm_recommendation: {FAST_TRACK, STANDARD, REVIEW},

model_version: string,

feature_hash: string

}

Page 678

www.rsisinternational.org

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026

APPENDIX D

Example Thresholds and Bounds Parameters (Illustrative)

Table D-I provides illustrative parameter ranges used for conservative physics-based bounds. Values are

placeholders for publication and should be calibrated to the target fleet, sensor characteristics, and jurisdictional

requirements.

Parameter

Typical Range

Used In

Interpretation

τ_E (energy trigger)

3–6 g (normalized)

Crash trigger

Minimum sustained

inertial energy above

baseline

τ_J (jerk trigger)

30–80 m/s³

Crash trigger

Reject braking-only

events with low jerk

τ_D (min duration)

20–60 ms

Crash trigger

Sustained energy

required to form a

pulse

Rear cervical

plausibility floor

|Δv_x| ≥ 1.0–1.8 m/s

Bounds model

Below this, strong

cervical claims are

unusual absent special

factors

Side thoracic

plausibility floor

|Δv_y| ≥ 1.2–2.0 m/s

Bounds model

Lateral loading

needed for rib/thorax

narratives

Frontal chest

plausibility floor

|Δv_x| ≥ 1.5–2.5 m/s

Bounds model

Belt/airbag loading

proxy

Rollover indicator

threshold

Var(ω) ≥ 0.5–1.5

(rad/s)²

Direction/pose

High rotational

variability suggests

rollover/rotation

Low-confidence

cutoff

confidence < 0.4

Policy

Suppress strong

mismatch tokens;

route to standard

handling

These ranges are intentionally broad. Production deployments should maintain a configuration service so

thresholds can be tuned without code changes and can be varied by device cohort (TCU vs phone).