Structured abstract
Background: A strength-aware cycling forecast should demonstrate value beyond recognising a rider's usual response or the demands of the workout. Objective: To determine whether detailed strength execution and recent exposure history improve next-day prediction of whole-body perceived effort during a specified familiar cycling workload beyond a strong monitoring benchmark. Design: A proposed prospective longitudinal study, with measurement feasibility and model development separated from a frozen evaluation in new participants. Measurement: A candidate fixed-power submaximal segment within an existing ride, accompanied by a genuine timestamped effort rating. The exact procedure requires scientific review and feasibility assessment. Primary comparison: Athlete-weighted difference in absolute forecast error between a detailed model and a benchmark already containing personal history, cycling, basic strength workload and permitted recovery information. Safeguards: Prospective opportunity selection, common information cutoffs, explicit non-completion, field-level provenance, uncertainty assessment and restricted final-evaluation access. Status: No participant results are reported. Measurement tolerances, instrument rights, practical performance thresholds, final sample size and the inferential method require resolution before the relevant participant or confirmatory phase. Predictive success would not establish that changing training according to the forecast improves performance.
1. Question, scope and target population
The primary question is fixed:
Does detailed strength execution and recent strength history improve next-day prediction of whole-body perceived effort during familiar cycling work, beyond a strong, simpler monitoring model?
The initial population is adult cyclists undertaking ordinary cycling and strength training. A power-enabled measurement cohort is proposed because the primary comparison requires sufficiently comparable observed work. The first claim would apply to the recruited setting and tested task, not automatically to runners, all app users, professionals or other performance tasks.
Routine maintenance training must be represented alongside unfamiliar strength work and returns after interruptions. The study will not increase fatigue deliberately to make a stronger signal. All training authority remains with the existing athlete/coach arrangement; the experimental forecast has no prescription-write role.
The present design is observational with respect to training decisions, but includes a research measurement procedure. It must not be described as passive data collection that avoids participant-research responsibilities.
2. Rationale and prior work
Existing research reports differing responses across soreness, submaximal cycling measures and time-trial performance after lifting. That provides a rationale for defining the outcome precisely rather than treating all recovery measures as equivalent. [1]
Published submaximal cycling tests also support considering a repeatable cycling measurement. The Lamberts protocol used heart-rate-controlled stages; the fixed-power procedure proposed here cannot inherit its reliability. [2]
The contribution would be prospective incremental prediction in the stated setting. The study does not claim to invent repeated-bout adaptation, perceived exertion, personal baselines or submaximal testing.
3. Phases and progression gates
Measurement feasibility. Confirm understanding of the effort question, task matching, timing, safety, burden, floor/ceiling patterns and the amount of useful within-person variation. An existing operational proposal of 30-50 riders over approximately 12 weeks is a planning envelope, not a completed power calculation or recruitment authorisation. A smaller operational start may be appropriate; no cohort is committed by this manuscript.
Model development. Use development data to define predictors, transformations, model structure, tuning and any permitted personal updating. Preserve all candidate analyses and disclose changes to the original plan. Development evidence must not later be presented as independent final validation.
Frozen evaluation. Evaluate the selected candidate and benchmark in a separate participant cohort. Freeze the global model, preprocessing, measurement rules, personal updating policy, fallback, analysis and practical success criteria before access to the relevant final outcomes. New participants may contribute prespecified personal familiarisation data, but cannot silently become a new global training set.
Product review. Review predictive performance, coverage, uncertainty and applicability before any forecast is displayed as validated. A separate decision is needed for using it to change training automatically.
4. Participants, screening and generalisation
Proposed inclusion requires adulthood, normal participation in cycling and strength training, a compatible power-recording route, ability to undertake the approved familiar task, and the relevant informed participation and data permissions. Screening, exclusions and symptom handling must be approved by the responsible human investigator and safety role.
Recruitment reporting should describe cycling level, strength experience, age distribution, sex information where justified, device mix and the coaching setting. Low representation of any subgroup limits the claim rather than justifying an unsupported subgroup model.
A new-participant test within one coaching service can assess transfer to similar users in that service. Generalisation to other services, seasons or training cultures requires evidence from those settings. Relevant setting and period information should be retained under an approved minimisation policy. TRIPOD-Cluster informs transparent reporting of between-setting structure; it does not substitute for a longitudinal analysis of repeated observations. [3]
5. Candidate reference task and outcome
The present candidate is five minutes of ordinary easy warm-up followed by six minutes at a familiar fixed endurance workload. The primary rating refers to the final minute of that six-minute segment. It would replace a suitable part of an existing warm-up rather than add a hard session. At most two measurement opportunities per week is the initial burden proposal.
Target watts are selected from credible current training information by the approved human scientific/coach procedure. No universal FTP percentage or biological recovery threshold is established here. Target changes create a new task version and familiarisation record. Bike, power source, trainer mode, position, cadence range and cooling should remain comparable where practical; meaningful changes are recorded.
The primary outcome is whole-body perceived effort, not soreness, pain or the whole ride's session rating. The candidate instrument is an appropriately authorised unchanged Borg CR10 implementation, subject to scientific and licensing review. Official scale wording is not reproduced in this manuscript. Electronic and commercial-use permissions must be confirmed rather than assumed. [4]
A response within two minutes of the segment is the current candidate timing rule, not an established validity cutoff. Exact prompt language, timing tolerance and permitted response domain must be settled before the measurement phase. No default or model-derived rating is accepted as the primary outcome. Leg and breathing effort may be collected separately as secondary responses, not substituted retrospectively for the primary rating.
Measured power verifies the performed work. Heart rate, cadence and objective execution are secondary observations. Required coverage, gaps, target deviation and other comparability tolerances must be specified before using the relevant outcomes. Failure to sustain the task remains a scientifically relevant execution result even when it prevents inclusion in the comparable-task rating analysis.
6. Opportunity selection and forecast timing
Select opportunities prospectively under the approved scheduling rule, before their forecast cutoff. Preserve original opportunities after cancellation, rescheduling, missing data or refusal. Do not choose only post-lift rides that later produce clear ratings.
The nominal primary forecast is issued approximately 24 hours before the planned task. An actual lead-time window of 18-30 hours is the present proposal. Store the information cutoff, issue time, scheduled task time and actual task time separately. Both models use the same eligible cutoff. Late-issued or out-of-window forecasts remain in explicit reporting categories rather than disappearing.
At issue, the preceding strength session may already be completed, may be planned to occur before the target ride, or may be absent/unknown. Completed future work cannot enter an earlier forecast. These states should be reported as prespecified strata, not selected after observing performance.
Morning and warm-up updates are separate forecast classes. They may contain newer information, but are not replacements for the primary advance prediction. Future sleep, actual next-day nutrition and outcome-period environment may be retained for secondary explanation; they cannot leak into the original prediction.
7. Data and measurement provenance
Use the existing Good Legs canonical records rather than a second training ledger. Distinguish original prescription, accepted adaptation, delivered version, version executed and completed work. Link each research observation to its task, source and version.
Record field-level origins: measured, athlete-entered, explicitly confirmed, prescribed, default-filled, derived or unknown. A value named actual in a database is not sufficient evidence that a participant reported it. Strength dose also needs a load convention and exercise/equipment identity; unrelated machine loads should not be treated as identical biological doses.
Retain native activity timing and sampling where permitted, not only smoothed display traces. Separate one physical activity from its multiple recordings and avoid counting duplicate workload. Preserve source-specific sleep episodes and compatible HRV methods without treating device outputs as interchangeable.
Each forecast uses only eligible revisions available to the relevant system at its cutoff. Later correction may improve a later derived record, but cannot rewrite the originally issued forecast. Research use and replay remain subject to current permissions and lawful retention.
8. Candidate models and fair benchmarks
Scroll the table horizontally when needed.
Candidate models and their information| Model | Intended role and information |
|---|
| M0: personal-history baseline | Prior personal responses and the demands of the reference task; a transparent sanity check. |
|---|
| M1: strong monitoring benchmark | M0 plus eligible cycling history, basic gym duration/effort or equivalent coarse workload, and available recovery/context information. |
|---|
| M2: detailed strength-aware candidate | M1 plus actual exercise-specific execution, timing, load convention and recent exposure-history information. |
|---|
M1 must know that strength training occurred. M2 must not receive better historical coverage, later cutoffs or outcome-derived information unavailable to M1. Missing-data handling and tuning opportunity should be comparable.
Simple regularised and hierarchical models are sensible development candidates. Individualisation and a more complex recovery shape must demonstrate incremental value, not be installed as assumptions. Compare an appropriate population formulation and a personal formulation using the same eligible inputs.
For absolute-error scoring, the point forecast should target the conditional median. A chart reporting a mean instead must identify that choice and use a consistent analysis. The personal usual reference is estimated only from earlier eligible observations. Subtracting the same reference from outcome and both forecasts leaves their absolute-error comparison unchanged; estimating a retrospective full-period usual value as a predictor would create leakage.
9. Primary estimand and analysis
For athlete i and valid observed occasion t, define:
d_it = |y_it - prediction_M1_it| - |y_it - prediction_M2_it|.
Positive values favour M2. The primary estimate is:
Delta_hat = mean_i [ mean_t(d_it) ].
The inner average uses that athlete's eligible observed comparable tasks; the outer average gives each evaluable athlete equal total weight. It does not give prolific loggers more weight simply because they supply more rows.
Report absolute error for both models, the paired difference, uncertainty, forecast availability and measurement coverage. Pooled correlation is descriptive only. Report within-person comparisons and the performance distribution across athletes, including ordinary versus exposure-transition periods.
The final inferential method must reflect plausible athlete, coaching-setting and calendar-time dependence. The companion synthetic study illustrates failure modes but does not select a universal estimator. The method and its assumptions must be fixed before final comparative outcomes are examined.
10. Missing observations and operational performance
Publish a flow from scheduled opportunities through forecasts, attempts, completion, comparability and observed ratings. Report causes of missingness, withdrawal, modifications and model non-issuance.
The primary accuracy target is conditional on valid observed tasks. It must not be described as accuracy for all days or all intended rides. A model that abstains on difficult days could appear accurate while offering little useful coverage. Prespecify operational fallback and report both candidate-only and fallback-inclusive performance where appropriate.
For completed comparable tasks with two eligible forecasts but a missing rating, a separate sensitivity calculation can use:
-|b-c| <= |y-b| - |y-c| <= |b-c|.
This elementary bound limits the possible paired error difference without inventing a rating. Aggregate the observed differences and missing-response limits within athletes, then average equally across athletes. The target includes completed tasks with missing ratings, so it is not identical to the observed-task primary target. The bounds are not confidence intervals and do not identify responses for unattempted, aborted or noncomparable tasks.
Quality adjudication and label corrections should follow predefined rules, preferably without access to which model performs better. Preserve both original and accepted outcome versions.
11. Sample size, uncertainty and success
There is no justified final recruitment number yet. Development complexity, per-athlete observations, outcome variability, missingness, calibration precision and dependence all matter. External-evaluation guidance supports planning for explicit precision, but clinical formulas cannot simply be transferred to this athlete-weighted paired comparison. [5]
Use feasibility/development evidence and sensitivity simulations to justify the final design. Report assumptions and examine plausible alternatives. More repeated observations cannot necessarily replace more independent athletes or broader settings.
Success requires a meaningful improvement over M1 and acceptable absolute accuracy, useful coverage and honest individual uncertainty. Define the minimum useful improvement, acceptable error and burden, and interval-performance criteria before the confirmatory results are available. Statistical detectability alone is insufficient.
A next-day prediction interval, a confidence interval for average model advantage and simultaneous coverage of a multi-day curve are different objects. Longer forecast horizons and new tasks require their own evaluation.
12. Safety, governance and dissemination
A named human sponsor/investigator and safety contact must approve the participant-facing process. Participants may decline or stop without compensatory exercise or coaching penalties. Experimental scores remain hidden during the first evaluation, while ordinary safety information remains available. Record relevant exposure to existing product advice without assuming that displayed information was read.
Research participation, permitted use of source data, pooled modelling, future reuse and recontact must be handled under the approved consent/privacy arrangements. API access approval is accepted as an operational starting point, not a substitute for the study-specific governance process. Use restricted pseudonymous analysis and avoid sending personal records to unapproved tools.
Before the relevant confirmatory phase, resolve registration/classification, publish or register the frozen protocol and disclose any development data already examined. A proposed protocol on a website is not registration, ethics approval or a Registered Report. A Registered Report would require review and acceptance through a participating venue. [6]
Report negative and inconclusive findings as well as positive ones. Follow an appropriate adaptation of prediction-model reporting guidance and publish the tested model/instrument versions, scope and limitations. [7] A later intervention study is required to establish whether acting on the forecast improves training or taper outcomes.
Decisions that must close before the relevant phase
The measurement phase requires the accountable investigator/safety arrangement, screening and permissions, exact task and instrument rights, numerical measurement-validity rules and burden/safety procedures. Confirmatory evaluation additionally requires the selected frozen models, final sample justification, statistical method, practical success thresholds and holdout-access policy. These are explicit decisions, not completed approvals.
Declarations
This is a proposed protocol developed for Good Legs, a commercial product. No human dataset has been analysed for the proposed empirical comparison and no participant results are reported. This manuscript does not appoint an investigator, imply a university affiliation or authorise recruitment.