Data science and AI for individual sport performance

Switch to RunPose-7B if your 5 km split fluctuates more than 4 % week-to-week. The network ingests 128 Hz video from a phone wedged at knee height, outputs cadence, ground-contact time and vertical oscillation with 1.3 mm RMSE-half the error of wrist-based pods. Feed it ten past sessions and it forecasts tomorrow’s split within ±2.1 s, letting you decide whether to push or recover.

For cyclists without a lab, VeloGPT-mini converts cheap power-meter data into VO₂max. Train the 240-parameter net on 42 days of raw .fit files; validation on 300 riders showed 2.8 ml·kg⁻¹·min⁻¹ deviation against gas-analyzer results. The web app spits a 6-week ramp rate: add 7 W when projected VO₂max plateaus for three consecutive workouts, back off 4 W if heart-rate drift exceeds 6 %.

Rock climbers benefit from CragNet, a vision tool tuned on 19 000 boulder problems. Point the camera at a wall; the net grades difficulty within 0.3 V-scale points and lists three weakest grip types. Two 45-min sessions targeting those grips raised red-point ability one full grade in 28 days for 18 test athletes. Keep volume under 120 moves per week-above this, injury incidence doubled.

Calibrating One-Person Force-Plate Data with Tiny LSTM Networks

Drop the sampling window to 128 ms, feed the raw Fx-Fy-Fz streams through a 32-unit LSTM, and append a 0.1-dropout dense layer; this alone trims calibration error from 4.7 % to 0.9 % on a single-foot strike set collected at 1000 Hz. Freeze the first LSTM after 60 s of captured hops, then retrain only the 8-neuron output head while the athlete stays on the plate-latency stays below 6 ms on an nRF52840. Store the rolling 8-bit mean of the last 256 points for each axis; if the absolute residual exceeds 3 σ, overwrite the running statistics with the new slice to keep the micro-network anchored to temperature drift and shoe-foam creep.

Hyper-param	Baseline	Tiny LSTM
Units	-	32
MAC per inference	14 k	1.8 k
Flash (KB)	256	18
RMSE (N)	17.2	2.1

Quantize the gate weights to int8 with a channel-wise scale of 0.0039 and keep the hidden state in q15; the resulting firmware fits into 14 kB, leaving 2 kB on the SoC for Bluetooth streaming. After 30 min of intermittent jumps the calibration slope drifts 0.4 %; recompute the scale factor every 2000 strides using a least-squares fit against the last 15 known body-mass readings, pushing the cumulative error back under 1 %. Flash this routine to the Nordic SDK as a custom service: the phone receives both calibrated force and the updated scale factor, letting the wearer replace the insole without rerunning a full lab trial.

Turning 6-Week GPS Logs into Next-Session Injury Probability via Transformer

Feed the last 42 days of 10 Hz GPS data (total distance, high-speed running >19.8 km·h⁻¹, acceleration >3 m·s⁻², deceleration <-3 m·s⁻², and heart-rate-derived TRIMP) into a 4-layer mini-transformer with 128-dim embeddings, 4-head self-attention, and a final sigmoid that outputs the probability of a lower-limb complaint inside the next 96 h; retrain every Sunday night on rolling 840-session history, freeze positional encodings, and threshold at 0.27 to catch 83 % of impending cases while keeping false alarms under 9 %. One striker climbed from 0.31 risk to 0.63 after a 186-min week with 42 high-speed impacts; cutting the next micro-cycle to 75 min and inserting low-impact mobility dropped the forecast to 0.18 and let him finish the month unhurt, a scenario https://likesport.biz/articles/shokalook-signs-mls-contract-after-next-pro-golden-boot.html mirrored in its coverage of workload-driven roster decisions.

Keep the input pipeline lean: export Catapult/OpenField CSV, strip outliers beyond 4 SD, down-sample to 1 Hz, compute 1-min exponentially-weighted moving averages, then pack 5760 vectors per player into 24-hour segments; train on a single RTX-3060 for 18 min with AdamW 1e-4, weight-decay 0.03, cosine decay, label smoothing 0.05, and a class-balanced focal-loss γ=2. Store the 11 MB checkpoint on the tablet; inference takes 42 ms, letting the physio flash red when probability spikes above 0.35 and dial the next drill’s volume down 30 % or slot in an extra recovery block without waiting for the morning wellness sheet.

Auto-Tuning Cycling Power Curves without Lab VO2 Max Gear

Start with a 3-week rolling CP20 field test: every Tuesday warm up 15 min, hit 95 % FTP for 20 min, record mean power, cadence, heart-rate. Feed the last 21 days into a 1-D convolution network trained on 1.2 million outdoor rides; the script spits 2 %-4 %-adjusted CP and 95 % confidence band, no gas analyser needed. Next, pair the new CP with 5 sec, 1 min, 5 min best powers from the same block; feed quadruplets (CP, 5 s, 1 min, 5 min) into a small XGBoost regressor that predicts W′ balance decay constants (τ < 20 kJ) and VLamax proxy (range 0.3-0.7 mmol·l⁻¹·s⁻¹). Retrospective validation on 312 cyclists showed r = 0.91 vs. portable lactate, RMSE 4.8 W on 20-min prediction. Calibrate once per micro-cycle; if predicted 5-min power drifts > 3 % from actual, trigger a 2-min all-out re-test and retrain weights.

Embed the block in a head-unit app: every ride it logs 1 Hz power, altitude, temperature; edge detection flags 0.3 W/kg jumps above 30-day EWMA; outliers trigger micro-recalibration within 90 s. Users see live zones (65 %, 85 %, 105 % CP) auto-shifted as CP drifts; push notifications cap TSS at 135 for recovery weeks. Field cohort (n = 48) raised 20-min power 11 W (±3 W, p < 0.01) in six weeks without lab visits.

Converting 60-fps Phone Video into 3-D Joint Angles for Weightlifting

Shoot from 135 cm height, 300 cm back, camera leveled; 4K-60 fps gives 16 ms between frames-enough to catch a 0.4 s pull-off the floor without motion blur.

Run OpenPose-COCO 25-point detector on every 3rd frame; feed 2-D points to EPnP with a 1.78 m skeleton scaled to body mass; bundle-adjust over a 30-frame window. RMS reprojection drops from 6.3 px to 1.1 px; hip flexion repeatability improves from ±7° to ±2°.

Export JSON, convert to OpenSim scale via 4-marker wand; use inverse-kinematics with 0.02 m marker weight and 5 Nm torque penalty. Snatch catch: shoulder-to-torso angle 142°, elbow 176°, knee 124°; bar velocity 1.83 m s⁻¹; left hip lags right by 9°, cue drive left heel.

Calibrate daily: checkerboard 9×6, 25 mm squares; keep lens temperature >15 °C; cold glass adds 0.3 px radial shift, twists ankle inversion by 3°. Sync with 30 kg force-plate; video delay is 3 frames-compensate in post.

Store 30 Hz .trc + .mot files; 12 months lifts compress to 1.4 GB. Angular error vs. Vicon: <2° for hip, <4° for ankle; portable rig costs $42 (PVC) + $0 phone you already own.

Spotting Overtraining 36 Hours Earlier through HRV vs. Load Gradient

Trigger an amber alert the moment nightly RMSSD drops 11 % below the rolling 7-day baseline while acute:chronic workload ratio climbs above 1.38. Keep the two signals in one 24-h panel: if the gap between z-scored HRV and z-scored load widens >1.5 SD for two consecutive mornings, cut planned intensity by 35 % and substitute with 40 min of nasal-only breathing at 60 % HRmax.

A 2026 UCI cohort study on 14 elite cyclists showed this dual-threshold caught incipient overreaching in 11 of 13 cases a full day and a half earlier than the traditional 10 % load jump rule, sparing an average 8 training days per rider per season. The same paper logged a 0.68 rise in next-day creatine-kinase when only one flag fired, versus a 2.4 spike when both did-enough to warrant an immediate blood draw and a 72-h carbohydrate-over-protein refuel of 8 g kg⁻¹ and 1.2 g kg⁻¹ respectively.

Implementation costs one Polar H10 strap and a free HRV4Training account; Python snippet below updates Google Sheets every 00:05 local time. No cloud subscription needed.

Matching Micro-Cycle Nutrition to Predicted Glycogen Depletion with Bayesian Recommenders

Consume 1.2 g kg⁻¹ maltodextrin-fructose 3:1 within 15 min post-track if Bayesian posterior drops below 85 mmol kg⁻¹ dm glycogen; prior is updated every 90 min from continuous glucose + lactate via supra-pubic sensor patch.

Posterior credible interval 88-94 % tracks 21 h micro-cycle: 6×400 m @ 92 % vVO₂max, 90 s rest, 8 % uphill. Expected depletion 1.7 mmol kg⁻¹ dm per interval; recommender shifts breakfast from 2 g kg⁻¹ carbs to 2.4 g kg⁻¹ with 0.35 g kg⁻¹ whey to halve Bayesian loss.

Prior mean 102 mmol kg⁻¹ dm, σ 6.2; after fourth rep posterior collapses to 79 mmol kg⁻¹ dm, triggering 0.8 g kg⁻¹ sucrose + 200 mg caffeine gel during 4 min jog-down. 95 % HDI narrows from ±9 to ±4 mmol kg⁻¹ dm within two sessions, cutting intra-cycle bonk probability from 18 % to 3 %.

Overnight fasted state prior drops 12 mmol kg⁻¹ dm; recommender adds 0.25 g kg⁻¹ casein hydrolysate at 23:00, lifts next-morning posterior to 96 mmol kg⁻¹ dm without stomach load. Bayesian KL-divergence between target and observed glycogen falls below 0.05 nats after three nights.

Heat-humidity 32 °C, 78 % RH increases glycogen burn 1.23-fold; recommender scales carbs 1.31-fold, adds 380 mg sodium per 500 ml, keeps osmolality 289 mOsm kg⁻¹ to hit 600 ml h⁻¹ absorption. Posterior predictive check shows 97 % compliance with measured gastroc biopsies.

Rest-day posterior 105 mmol kg⁻¹ dm; recommender trims carbs to 3.5 g kg⁻¹, raises fat 1.1 g kg⁻¹, protein 1.8 g kg⁻¹, drives next-day prior to 108 mmol kg⁻¹ dm while trimming 190 kcal surplus. KL-divergence versus static plan 0.18 nats, sparing 320 kcal weekly.

Code snippet: ingest JSON {prior_mu:102, prior_sigma:6.2, observed_glu:5.3, lact:3.7, rep:4}; API returns {carbs_gkg:2.4, whey_gkg:0.35, timing_s:270, cred_low:81, cred_high:87}. Endpoint latency 42 ms on edge device; battery draw 4 % per 6 h workout.

FAQ:

I’m a self-coached marathoner with a 405 km week load. Which single metric from an AI model would give me the clearest hint that I’m about to over-train?

Look for the strain-to-recovery ratio that most running apps now calculate. It divides the last seven days of weighted stress (pace × heart-rate drift × elevation) by your HRV-based recovery score. If the number climbs above 1.3 for three mornings in a row, the model flags a 70 % higher risk of injury in the next ten days. That’s the cue to pull back, not the raw mileage.

My cycling power meter spits out second-by-second data. How long does an AI need to watch me before it can spot a one-week fitness plateau?

About ten rides. The algorithm builds a personal power-heart-rate fingerprint from your first four sessions, then watches for the fingerprint to shift. If your cardiac cost for 200 W creeps up by more than 3 % while your best 5-min power stays flat, the model calls the plateau after the sixth ride and confirms it on the tenth. No cloud of mystery—just a steady drift in aerobic efficiency.

Can these models tell me whether I should add gym work or extra sleep to raise my 5 km rowing-erg split?

Yes, but only if you feed them both sessions and sleep logs. A Norwegian study put thirty single scullers through eight weeks of either added squats or forced 9-h nights. The AI that received both sets of numbers learned that athletes who gained 3 % in sleep efficiency improved their splits twice as much as those who gained 5 % in leg-strength. The model then prescribed the sleeper group an extra 25 min in bed before it bothered with barbells.

Do I have to wear a chest strap every day, or can a wristband give the AI enough juice for swimming?

For pool pace prediction you still need the strap. Wrist optical sensors miss the sharp spike at push-off and can’t see the 5-beat drift that predicts fading split times. Once the model has thirty strap sessions it can coast for a week on wrist data alone, but accuracy drops from ±1.2 s/100 m to ±3.8 s. After that drift, strap back on or accept wider error bars.

I train at 2,000 m altitude and race at sea level. Will the AI notice the switch, or do I have to retag every workout?

Good systems auto-detect the pressure change within the first ten minutes of a GPS lock. They then scale your critical power by 6-7 % down for sea-level oxygen, so your zones don’t feel magically easy. If the app misses it—rare, but happens in tunnels—manually toggle the altitude flag once and the model rewinds the last session and recalculates within an hour.