Instead of reducing a head-motion event to a peak acceleration or peak rotational velocity, IFS represents the entire waveform as a measurement vector — one coordinate per computed measure, drawn from six mechanical domains. The position of that point in the space is the description of the event.
A peak value is a one-dimensional summary — it answers a single question: how big was this? IFS instead treats the multidimensional space itself as the object of study. Each recorded event becomes a point x ∈ ℝⁿ, positioned by measures extracted from the resultant linear acceleration, angular velocity, and angular acceleration signals across the full event window.
Adding variables to a classifier is a common approach in sensor research. IFS is different in intent: the variable set isn't chosen to optimize a prediction target. It's chosen to describe the mechanics of the event — temporal structure, coupling, and frequency content — so the geometry of the resulting space can itself be examined, probed for its dimensionality, and mapped onto sport-specific or clinical questions.
Strictly speaking, not everything the current pipeline computes is IFS. Six domains — 285 variables — are IFS proper: continuous descriptions of the waveform itself, with no external labels and no assumed body constants. The rest is computed by the same pipeline as useful extras, described further down this page.
The temporal shape of the signal itself — secondary peaks, oscillations, rise and decay, how energy is distributed across the event window. A sharp, brief transient and a broad, sustained load can share a peak value and still be structurally distinct.
The frequency content of the waveform, and how that content changes over the course of the event — not just a static spectrum, but its evolution in time.
How rotation and translation relate to each other during the event — not just their relative size, but their timing and statistical dependence.
Which direction the motion is concentrated in, and how stable that direction is over the course of the event.
Rotational energy split into the three anatomical planes of head motion — a coordinate transform, not a new measurement, but one that ties the waveform directly to the axis-dependence findings discussed on The Story.
The smallest of the six domains — and no less real for it. Algorithmic-complexity measures of how patterned or how irregular the waveform is, independent of its amplitude or frequency content.
The same 76-millisecond window that produces the six IFS domains above also produces three further families of output. They're real, they're computed on every event — but they're either external criteria the pipeline happens to also calculate, or they depend on labels and context from outside the waveform itself. Worth having; not part of the IFS representation.
HIC15, HIC36, BrIC, DAMAGE, HIP, and RIC36 — the historical scalar criteria discussed on The Story — computed directly from the same recorded waveform, as seven extra coordinates rather than replaced by IFS.
Sample and permutation entropy, coherence, HF/LF ratio, timing-reversal and retrigger-cascade detection, plus two physical-plausibility checks against fixed sensor limits — the one place in the pipeline where a genuine threshold is applied, as a sensor-fault sanity check rather than an injury or severity classification.
Football-heading directional and contact-type geometry, ice-hockey on-ice/bench/off-play descriptors, and an "active vs. passive impact" signature — all conditional on an external context label (video, player-tracking, or manual tag), not derived from the waveform alone.
The published manuscript analyzes 111 IFS variables — the set behind every dimensionality, orthogonality, and classification result on the Evidence page. The four domains in that analysis are a subset of the six described above; development since publication has grown the pipeline to 417 variables in total. The commitment was never to a fixed inventory — only to representing the waveform as a structured, multidimensional object rather than a scalar.
The structural information IFS is built to capture is exactly what gets corrupted by loose sensor coupling. A helmet-mounted sensor records helmet motion, not head motion. In the demonstration dataset, the sensor was pressed firmly against the mastoid process with an elastic headband — not adhesive alone — to limit relative movement and hold a consistent anatomical measurement geometry. Other placements or retention methods require independent validation.
Dimensionality, orthogonality, and classification results from 4,402 recorded events across 19 elite women's ice-hockey matches.