Player Load Spikes as False Positives: How to Fix the Unit Shake Artifact

A winger closes out a session at 612 arbitrary units of Player Load. Nothing in the video looks unusual - no heavy collision, no unplanned sprint volume, a fairly ordinary small-sided game and a cool-down jog. But the number is 14% above her rolling four-week average for a session of that duration, and if the staff meeting the next morning treats that figure as real external load, she gets flagged for extra monitoring or, worse, a load-management day she does not actually need.

Pull the raw accelerometer trace and the story changes. Three short bursts, each under two seconds, show a rapid tri-axial oscillation with no matching change in GPS speed or heart rate. They cluster right after two hard changes of direction and one landing off a header - moments when a slightly loose vest lets the unit itself move independently of the athlete's trunk. The device recorded something real: it just was not the athlete's movement. It was the unit shaking against her back.

Player Load is calculated from the instantaneous rate of change in acceleration across three orthogonal axes, summed across the sampling window and typically scaled by a constant of 100. That formula has no way to know whether the acceleration it is reading came from the athlete's trunk or from the sensor housing wobbling inside a harness that is not snug against the body. Three mounting conditions produce the artifact, and each has a slightly different signature.

Harness slack allowing free unit movement. If the pocket holding the unit has more than a small amount of give, hard changes of direction, landings, and deceleration create a secondary, independent oscillation of the unit itself - sometimes called vest slap or harness bounce - layered on top of the athlete's genuine acceleration. Because Player Load sums the magnitude of acceleration change rather than net displacement, this added high-frequency wobble adds directly to the score rather than canceling out.

Off-position mounting. Manufacturers specify a mounting location - typically upper back, between the scapulae, oriented with a fixed axis alignment - because the calculation assumes a consistent relationship between unit orientation and trunk movement. A unit that has slid laterally, sits too low, or is rotated even slightly reads a different acceleration profile for an identical movement, and one consequence is a harness with more surface area to catch air resistance or garment drag during running, which increases the chance of independent oscillation.

Cold-start settling. In the first few minutes after a harness is put on, particularly over a compression layer that has not yet warmed and settled against the skin, straps that felt snug standing still can loosen fractionally once the athlete is moving, sprinting, and changing direction. The unit has more room to move for roughly the first five to ten minutes of a session than it will for the rest of it, which is why early-session data sometimes carries a disproportionate share of a session's flagged spikes.

Not every high accelerometer reading is noise, and a coach who filters out every outlier will eventually erase a genuine collision or a hard, load-bearing landing that the program actually needs to see. Four checks separate a real external-load event from a harness artifact.

• Cross-reference against GPS speed and deceleration. A genuine hard cut, sprint, or collision almost always shows a corresponding change in the GPS speed trace - a deceleration spike, a sharp direction change in the position data, or a dead stop consistent with contact. A unit-shake artifact typically shows no matching kinematic event at all; the GPS trace is smooth while the accelerometer trace spikes.
• Check the oscillation pattern, not just the peak. A true collision or landing produces one sharp acceleration event followed by a return to baseline. Harness slap produces a short burst of repeated, decaying oscillation - two, three, sometimes four rapid direction reversals in well under a second - because the unit is physically bouncing against the body rather than registering a single kinematic impact.
• Look for repetition tied to specific movements, not specific plays. If the flagged spikes cluster around every hard change of direction a given athlete performs, regardless of what is happening on the field, that pattern points to fit rather than to game events. Real load-bearing incidents are tied to what actually happened in the session, not to a movement type in isolation.
• Confirm with video for anything above your normal threshold. For any single spike large enough to meaningfully move a session total, a 10-15 second video check costs little and settles the question immediately - either you see contact or a hard landing, or you see an athlete running smoothly while the number says otherwise.

Not every way of reporting Player Load is equally exposed to this artifact. Boyd, Ball and Aughey (2011, International Journal of Sports Physiology and Performance) examined the reliability of accelerometer-derived Player Load using a standardized Australian football movement circuit and reported a between-day coefficient of variation for Player Load in roughly the mid-single-digit percentage range, with reliability further reduced when comparing shorter, high-intensity segments rather than whole-session totals. Their finding matters here because a whole-session Player Load total averages the artifact's brief spikes across many minutes of otherwise clean data, diluting its visible effect on the aggregate number - while an instantaneous or short-epoch Player Load view (the kind used to flag individual events for review) shows the same artifact undiluted, as a sharp, isolated peak.

Nicolella, Torres-Ronda, Saylor and Schelling (2018, PLOS ONE) took this further by benchmarking accelerometer-based tracking units directly against a criterion measurement system and comparing multiple units of the same model under matched conditions. They reported inter-unit variability for accelerometer-derived load metrics that, in some testing conditions, exceeded 10% between units of identical make and model, and specifically flagged secure, consistent unit fit as a factor practitioners could control that the device firmware could not correct for after the fact. Their study was conducted under controlled testing conditions rather than live competition, which is a real limitation - field sessions introduce contact, uneven surfaces, and garment variability that a lab protocol does not fully replicate - but the core point holds across settings: a device cannot distinguish a well-mounted unit's true signal from a poorly-mounted unit's mechanical noise, because both arrive at the accelerometer as acceleration.

Filtering after the fact helps, but the more reliable fix happens before the session starts. Run this fitting check with every athlete, every session, not just the ones who have shown spikes before.

1. Position the unit between the scapulae, centered on the upper back, oriented exactly as the manufacturer specifies. A unit rotated even 10-15 degrees off its intended axis changes how it reads the same physical movement.
2. Apply the two-finger snugness test. You should be able to slide two fingers flat under the strap at the chest or side closure, but not a full hand. Looser than that and the pocket has room for the unit to move independently during hard efforts; tighter than that starts to restrict breathing and normal trunk motion.
3. Have the athlete perform 20-30 seconds of movement before recording starts - a jog, a few change-of-direction cuts, a vertical jump. Straps that felt snug standing still often reveal slack once the body is actually moving, and this settling period lets you re-tighten before the data that matters begins.
4. Recheck fit at the half or after any kit change, including a jersey swap or added layer for weather. A garment added over the vest, or a vest repositioned during a substitution, resets the fit conditions that were checked at the start.

Run this once per athlete whenever a session total looks disproportionate to session RPE, distance, or your own eyeball read of the video.

1. Pull the last 3 sessions of instantaneous or per-epoch Player Load data for the athlete in question, alongside the synchronized GPS speed trace.
2. Identify every window where instantaneous Player Load exceeds roughly 2.5 standard deviations above that session's own rolling mean.
3. For each flagged window, check the GPS speed trace for the same timestamp. No matching deceleration, direction change, or dead-stop event is the artifact signature; a matching event confirms a genuine load-bearing incident.
4. Count artifact-flagged windows as a share of total flagged windows across the three sessions. Below roughly 10% is typical background noise from ordinary movement variability; above 25% points to a fit issue worth addressing directly with that athlete's harness rather than filtering around session after session.
5. If the share is high, re-run the fit-and-settle protocol with that athlete specifically and re-audit after the next session to confirm the artifact rate has dropped.

The limitation worth naming here is one both cited studies share: reliability and validity testing for accelerometer-based Player Load has mostly happened in standardized circuits or lab settings rather than live competition, and match-day contact, weather, and kit changes introduce variability that a clean protocol does not fully capture. That is a reason to keep auditing across real sessions rather than trusting a single lab-derived reliability figure as a permanent guarantee - fit can degrade mid-session in ways a controlled test never sees.

A senior academy midfielder logged a 68-minute small-sided and possession-based training session. Raw instantaneous Player Load flagged three windows against the synchronized GPS trace below.

The raw session total came in at 604 AU. The two flagged artifact windows contributed roughly 29 AU combined with no corresponding GPS event to justify them, both occurring during low-speed movement where a genuine spike of that magnitude would be physiologically implausible. Excluding them dropped the adjusted session total to 575 AU - a 4.8% correction that, on its own, would not have changed a load-management decision this particular week, but that compounds across a training block if the same athlete's harness keeps producing the same artifact session after session. The genuine spike at 31:07 stayed in the data untouched, because it had exactly what the artifacts lacked: a matching kinematic event in the GPS trace.