Using machine learning and data analytics in cycling coaching

Good coaching has always been about pattern recognition. Machine learning gives us faster, clearer patterns from messy ride files, heart rate, and recovery data. Used well, AI can predict performance trends and tailor your plan so every session has a purpose.

What data matters and how AI reads it

The goal is to turn raw numbers into useful signals. A modern coaching stack typically ingests:

• Power and cadence: watts by second, power curves, time-in-zone, cadence distribution.
• Heart rate and HRV: response to workload, cardiac drift/decoupling, day-to-day recovery readiness.
• Training load and context: TSS or session RPE x duration, training monotony/strain, sleep duration/quality, nutrition notes.
• Environment: temperature, altitude, terrain, surface, wind; indoor vs outdoor differences.

From this, AI extracts features coaches care about:

• Critical power and FTP trends, W′ balance, durability (how long you hold a % of FTP before power or HR drifts).
• Recovery kinetics: heart rate lag and post-interval recovery speed.
• Load-response curves: how your FTP, 5-minute power, or repeatability change with different doses of endurance, tempo, threshold, VO2 max, and anaerobic work.
• Compliance and execution: how closely you hit targets, and which sessions you typically cut short.

From raw numbers to predictions that help you train

AI models don’t replace judgment. They produce probabilities and forecasts that inform your next decision. Common, useful outputs include:

• Performance trend forecasting: expected FTP or critical power change over 2–6 weeks; likelihood of a personal best in specific durations (1, 5, 20, 60 minutes).
• Readiness and recovery: daily readiness score combining HRV, sleep, and prior load; risk of not completing a hard session as prescribed.
• Session outcome prediction: probability of completing 5×5 minutes at 108–112% of FTP based on recent execution and recovery metrics.
• Load risk alerts: flags for rising training monotony or rapid weekly load spikes that increase illness/injury risk.

Personalization: finding your dose–response

Personalization is about discovering which stimuli move your fitness most efficiently, then delivering the right dose at the right time.

How AI does it:

• Clustering: groups your rides and responses to find patterns (e.g., you progress best when weekly tempo time hits 60–90 minutes, while threshold beyond 40 minutes adds fatigue without extra gain).
• Nonlinear dose–response models: estimate how changes in weekly time-in-zone, intensity, and density relate to changes in FTP or specific power durations.
• Adaptive planning: updates the next block based on your recent response, not a one-size-fits-all plan.

A simple 8-week example:

1. Weeks 1–2 (probe): Alternate two stimuli to test response.
   • Week 1: Tempo focus (2×20–30 min at 80–88% FTP), endurance rides at 60–70% FTP.
   • Week 2: Threshold focus (3×10–12 min at 95–100% FTP), endurance as above.
2. Weeks 3–6 (exploit): If your model shows stronger gains from tempo, bias the block toward tempo with a light touch of VO2 max (e.g., 6×2–3 min at 115–120% FTP every 10–14 days).
3. Week 7 (deload): Reduce volume by 30–40%, keep 1–2 short efforts to maintain feel.
4. Week 8 (retest and re-learn): 20–40 minute sustained effort or formal test; update CP/FTP and zones.

Throughout, the plan flexes day-to-day with readiness and compliance probabilities rather than forcing hero workouts on low-recovery days.

Guardrails and good habits when using AI

• Data quality first: zero-offset your power meter, use the same heart rate strap, and keep indoor vs outdoor context consistent when comparing.
• Label your sessions: add session RPE and a one-line note. These are powerful features for prediction and help separate “hard but fine” from “too hard.”
• Trust trends over single days: a red readiness flag once isn’t a crisis; three in a row after big load is a signal.
• Mind intensity distribution: avoid stacking threshold and VO2 max on consecutive days unless you’re very fresh.
• Respect privacy and consent: know what’s shared and why. Ask for model explanations when possible.
• Coach override stays: algorithms propose; you and your coach decide.

A sample weekly workflow with AI

Here is how to use predictions without overcomplicating your week.

• Monday (rest or 45–60 min easy at 55–65% FTP): Check readiness; if HRV is suppressed, keep it truly easy. Log sleep and nutrition.
• Tuesday (VO2 max): Plan 5×3–4 min at 110–120% FTP. If the model shows a high non-completion risk, shift to 6–8×2 min or move to Wednesday.
• Wednesday (endurance): 90 minutes at 60–70% FTP. Watch heart rate decoupling; if drift exceeds ~5–7%, cut short and prioritize recovery.
• Thursday (threshold): 3×10–12 min at 95–100% FTP with 1:1 recoveries. If the model predicts better gains from tempo, swap to 2×25 min at 80–88% FTP.
• Friday (recovery): 45 minutes easy or full rest. Confirm sleep and HRV have normalized.
• Saturday (long endurance with tempo): 2.5–3 hours mostly 60–70% FTP, include 2×20–30 min at 80–88% FTP if readiness is green.
• Sunday (endurance): 1.5–2 hours at 60–70% FTP. If the week’s predicted monotony is high, add short spins or cadence drills instead of more load.

Rule of thumb: let AI set the odds, then choose the safest path to the same goal. Adjust the how, not always the what.

Key takeaways

• Use AI to forecast trends (FTP, CP, durability) and to personalize dose–response, not to micromanage every pedal stroke.
• Keep data clean and context-rich (RPE, sleep, notes) so predictions stay trustworthy.
• Make small, frequent adjustments based on readiness and compliance probabilities; reassess every 2–4 weeks.

Think of machine learning as a skilled second set of eyes on your training. It spots the patterns, you make the call.