Data-driven computational framework for personalized training plans in cycling

Technical Summary: Data-driven computational framework for personalized training plans in cycling

Problem Statement
Despite the proliferation of wearable technology in cycling generating massive longitudinal datasets, the transition from descriptive monitoring to data-driven prescriptive decision-making remains unrealized. Current training plans are predominantly designed through manual expert interpretation rather than mathematical optimization. Furthermore, existing methods for assessing performance rely on invasive, infrequent laboratory tests (e.g., lactate thresholds, $VO_2max$) or maximal field efforts (e.g., FTP tests) that disrupt training cycles. There is a distinct lack of systems capable of mathematically determining optimal training interventions to achieve specific performance goals, leaving a gap between predictive analytics and prescriptive sports science.

Methodology
The authors propose a modular computational framework designed to automate the generation of personalized training plans. The framework consists of four sequential, interchangeable steps:

1. Continuous Performance Metric Definition:
   Instead of relying on discrete lab tests, the study introduces a continuous metric, PO@HRtarget (Power Output at Target Heart Rate). This metric is derived from submaximal longitudinal sensor data (power output and heart rate).

   • Calculation: It filters efforts within a personalized heart rate range centered on the 60th percentile HR. It discards the first 5 minutes (to account for HR lag) and data after 60 minutes (to avoid cardiovascular drift).
   • Normalization: Power values are linearly interpolated to a target heart rate ($HR_{target}$) using resting heart rate and effort heart rate as bounds.
   • Aggregation: For a given date, the metric aggregates corrected power values from the preceding four weeks, applying a trimmed mean (removing the top and bottom 20%) and clipping values against historical ranges to ensure robustness.

2. Feature Engineering:
   Features are extracted from workout data to quantify training load and type, specifically focusing on variables that can be prescribed. These include:

   • Load Metrics: Chronic Training Load (CTL), Acute Training Load (ATL), Chronic/Acute Duration Load (CDL/ADL), and Chronic Intensity Load (CIL), calculated using exponentially decaying averages with specific time constants (e.g., $\tau=56$ for CTL).
   • Intensity Distribution: The Polarization Index (PI), derived from the proportion of time spent in three heart rate zones (below LT1, between LT1/LT2, and above LT2).

3. Predictive Modeling:
   The framework trains regression models to predict the PO@HRtarget metric based on the extracted features. Four architectures are evaluated to capture varying dose-response relationships:

   • Lasso: A linear model with L1 regularization.
   • Random Forest (RF): A non-linear ensemble method.
   • XGBoost: A gradient boosting framework.
   • RaFFLE (Random Forest Featuring Linear Extensions): A novel model combining random forests with linear model trees.
   • Robustness: To handle noise in field data, the models prioritize robustness over standard least-squares minimization, utilizing Mean Absolute Error (MAE) as the optimization objective and clipping predictions based on Median Absolute Deviation (MAD).

4. Prescriptive Optimization:
   A Differential Evolution algorithm is employed to invert the predictive models. The algorithm searches for a set of training plan parameters that maximize the predicted future performance (PO@HRtarget) or Training Stress Balance (TSB).

   • Parameters: The optimization varies four weekly parameters: Load increase fraction, Duration standard deviation, Amount of rest days, and Polarization index.
   • Translation: These parameters are translated into daily training plans (duration, intensity, and zone distribution) using constraints such as the inverse relationship between intensity and duration ($IF \leq 0.9d^{-1}$).

Key Results

• Metric Validation: Validation against laboratory lactate threshold tests on four professional cyclists showed that PO@HRtarget effectively tracks physiological trends, though it exhibited instability following rest periods due to low data volume.
• Model Performance: Across 18 athletes, the models achieved an average out-of-sample Mean Absolute Error (MAE) between 5 and 10 Watts. This error magnitude is comparable to the intrinsic measurement uncertainty of commercial power meters ($\pm 1.5\%$).
• Inter-Subject Variability: No single model architecture performed best for all athletes. Linear models (Lasso) sufficed for some, while non-linear ensembles (RF, RaFFLE) were required for others, validating the necessity of a flexible, modular architecture.
• Feature Importance: SHAP analysis and feature importance plots consistently identified chronic features (e.g., CTL) as the dominant predictors of performance, with intensity distribution (PI) and intensity load (CIL) providing secondary value.
• Optimization Outcomes: The prescriptive optimization demonstrated a potential average performance increase of 5.5% when targeting PO@HRtarget, with a corresponding average TSB change of +1.21. When targeting TSB directly, the performance change was negligible (-0.1%), but TSB increased significantly (+13.20).

Significance and Claims
The paper claims to bridge the gap between predictive analytics and prescriptive decision-making in sports science. Its primary contributions are:

1. A Modular Framework: Unlike rigid solutions, the framework allows for the independent selection of performance metrics, predictive models, and optimization strategies, accommodating the diverse physiological characteristics of individual athletes.
2. Continuous Monitoring: By deriving performance metrics from daily submaximal training data, the framework eliminates the logistical burden and training disruption associated with frequent maximal laboratory or field testing.
3. Automated Prescription: It demonstrates a pathway to move from retrospective analysis to forward-looking, data-driven training prescription, generating optimized plans that maximize specific performance goals.
4. Scalability: While validated on cycling data, the authors assert the methodology is sport-agnostic and scalable to other endurance disciplines, offering a mechanism to leverage big data for individualized performance optimization.

The authors maintain a modest tone regarding limitations, noting that the current framework generates static plans based on historical data without real-time feedback loops for daily recovery (e.g., HRV, sleep) and that the precision of prescriptions depends on the quality of athlete metadata (e.g., accurate FTP and threshold estimates).