2. Injury biomechanics fundamentals

Injury biomechanics provides the quantitative bridge between the physical response of an occupant (accelerations, forces, deflections) and the likelihood of clinically meaningful injury. For uncertainty quantification, this bridge matters twice: the criteria themselves are deterministic formulas applied to noisy signals, and the tolerance limits they are compared against are statistical constructs derived from limited, heterogeneous data. Understanding both layers is a prerequisite for building defensible UQ models, because any uncertainty you propagate through a simulation must eventually terminate in an injury metric whose own variability is understood.

Head Injury Criterion (HIC)

HIC remains the most widely used head injury metric in automotive crash assessment. It is computed from the resultant translational acceleration at the head center of gravity, measured in units of g:

HIC = max over t1,t2 of [ (t2 - t1) * ( (1/(t2-t1)) * ∫ a(t) dt )^2.5 ]
subject to t2 - t1 ≤ 15 ms

The 2.5 exponent heavily penalizes high average acceleration over the window, and the 15 ms cap reflects the empirical observation that injury risk depends on both magnitude and duration in a nonlinear way. The regulatory threshold in FMVSS 208 is HIC ≤ 700 for the 50th-percentile male dummy; Euro NCAP and other protocols apply similar or slightly different limits depending on body region and test configuration. A practical subtlety: HIC is sensitive to the acceleration signal's high-frequency content, so the sampling rate and the CFC filter applied before computation materially change the result. Two analysts computing HIC from the same raw channel with different filter classes can disagree by tens of points.

Brain Injury Criteria: BrIC and Beyond

Rotational head motion drives diffuse axonal injury and subdural hematoma far better than linear acceleration alone. The Brain Injury Criterion (BrIC) addresses this by combining peak angular velocities about the three anatomical axes, each normalized by a critical value:

BrIC = max( |ωx|/ωxC , |ωy|/ωyC , |ωz|/ωzC )
with ωxC = 66.25 rad/s, ωyC = 56.45 rad/s, ωzC = 42.87 rad/s

A BrIC of 1.0 corresponds approximately to a 50% risk of abbreviated injury scale (AIS) 4+ brain injury in the underlying dataset. Note that BrIC uses the maximum normalized component rather than a Euclidean norm — a design choice that makes it conservative for single-axis-dominant loading but less sensitive to combined loading. Newer criteria such as DAMAGE and CSDM attempt to capture strain-based tissue response and directional dependence more explicitly, and they are increasingly used in AI-assisted injury prediction because they correlate better with finite-element brain model outputs.

Neck and Chest Criteria

Neck injury assessment typically combines axial force, bending moment, and the Neck Injury Criterion (Nij), which normalizes compression and tension by their respective tolerance values and takes the sum of the normalized force and moment terms. Chest injury is characterized by the Chest Compression Criterion (maximum deflection normalized by half-chest depth) and the Viscous Criterion (VC), the product of compression velocity and instantaneous compression, which captures soft-tissue rate effects. Each of these criteria has its own tolerance limit and its own sensitivity to signal processing choices.

Channel Frequency Class (CFC) Filtering

CFC filtering is the standardized low-pass preprocessing applied to transducer channels before injury metric computation. The CFC number defines the −3 dB cutoff frequency: CFC 60 corresponds to 100 Hz, CFC 180 to 300 Hz, CFC 600 to 1000 Hz, and CFC 1000 to 1650 Hz. The filter is a fourth-order Butterworth applied in both forward and reverse directions (zero-phase) to avoid phase distortion. This matters for UQ because the filter class is a modeling choice, not a physical property of the occupant. If your simulation outputs are unfiltered while your reference data were CFC-filtered, the comparison is invalid, and the resulting uncertainty estimate will conflate filtering differences with genuine biomechanical variability.

Tolerance Limits as Statistical Constructs

A tolerance limit is not a property of human tissue; it is a summary of a probability-of-injury curve fitted to a dataset. The widely cited head injury risk curve (e.g., the logistic relationship between HIC and AIS 3+ risk) was derived from a specific mix of cadaver, animal, and volunteer data, with substantial censoring and limited sample sizes at high severity. The curve's shape and its confidence bands are themselves uncertain. When you treat HIC = 700 as a hard threshold, you are implicitly accepting a particular risk level (roughly 20–30% for AIS 3+ depending on the curve used) and ignoring the uncertainty in the curve itself.

This is where UQ becomes essential rather than optional. A rigorous approach distinguishes three variability sources: inter-individual variability (anthropometry, material properties, pre-existing conditions), intra-individual variability (posture, muscle state, boundary conditions), and epistemic uncertainty in the criteria and tolerance data themselves. The first two are aleatory and irreducible by more data collection; the third is reducible and should be represented explicitly, for example as a distribution over the logistic regression coefficients of the risk curve.

Working Through a Concrete Example

Consider a frontal sled simulation of a mid-size male occupant. The head CG acceleration channel is sampled at 10 kHz and contains visible high-frequency ringing from the restraint system. Suppose the unfiltered peak resultant acceleration is 82 g with a 12 ms pulse above 60 g, yielding an unfiltered HIC of approximately 780. Applying CFC 1000 (1650 Hz cutoff) removes the ringing and reduces the peak to 74 g, giving HIC ≈ 690. Applying CFC 180 (300 Hz) instead reduces it further to HIC ≈ 640. The same physical event produces three HIC values spanning a 140-point range purely from preprocessing choices — a spread comparable to the difference between passing and failing a regulatory threshold.

This example is illustrative, not a measured result, but it captures the practical reality: before attributing variability to occupant characteristics, you must lock down the signal processing pipeline. In a UQ study, treat the CFC class as a discrete epistemic variable with a small set of plausible values, run the metric computation under each, and report the resulting spread separately from the aleatory occupant variability. This decomposition prevents the common error of inflating inter-individual variance to absorb filtering uncertainty.

From Deterministic Thresholds to Probabilistic Risk

The mature way to use injury criteria in UQ is to abandon the pass/fail framing and work directly with risk. For each simulated occupant, compute the injury metric, then evaluate it against a probabilistic tolerance model — for example, a logistic curve P(AIS 3+) = 1/(1 + exp(−(β₀ + β₁·HIC))). Propagate the uncertainty in β₀ and β₁ alongside the occupant and simulation uncertainties. The output is not a single number but a distribution of injury probability, from which you can extract percentiles, exceedance probabilities, and sensitivity indices identifying which input contributed most to the spread.

This reframing also clarifies what AI models should learn. Rather than regressing directly onto a binary injury label — which discards the severity gradient and amplifies label noise — a well-designed surrogate can predict the continuous injury metric and its associated uncertainty, leaving the tolerance model as an explicit, auditable layer. That separation keeps the biomechanical assumptions visible and makes the UQ results interpretable to both engineers and regulators.