AI Physical Fitness Monitoring via Wearables

A coach notices: an athlete complains of fatigue, HRV has dropped 30%, and tests aren't improving. Yet the training log shows overload. With unstable Wi-Fi, data may be lost, and HRV intervals contain artifacts — without proper filtering and interpolation, the AI model will err. Without continuous monitoring and AI analysis, it's hard to distinguish overtraining from the onset of illness. We build systems that turn raw wearable data into actionable insights: recovery assessment, performance prediction, and early failure warnings.

Wearable devices — Whoop, Oura Ring, Garmin, Apple Watch — continuously collect biometrics. An AI system aggregates these streams, computes a Recovery Score, a personal physiological baseline, and long-term trends. The result is an objective picture of readiness for load. A personalized baseline is 2 times more sensitive than population norms for detecting deviations. Stack: Python, PyTorch, PostgreSQL with TimescaleDB for time-series storage, MLflow for experiment tracking.

How We Do It

How AI Analyzes Wearable Sensor Data

Cardiovascular metrics: HRV, resting heart rate, SpO2. Activity: steps, GPS, gyroscope. Sleep: REM, deep, light. Skin temperature: deviation from baseline — indicator of illness.

Recovery Score Model

def calculate_recovery_score(hrv_today, hrv_baseline,
                              sleep_quality, sleep_duration,
                              resting_hr, resting_hr_baseline):
    hrv_score = min(1.0, hrv_today / hrv_baseline)
    sleep_score = (sleep_quality * 0.5 + min(1.0, sleep_duration / 8.0) * 0.5)
    hr_score = max(0, 1.0 - (resting_hr - resting_hr_baseline) / resting_hr_baseline)
    recovery = hrv_score * 0.5 + sleep_score * 0.35 + hr_score * 0.15
    return recovery * 100

Recovery < 33% — red, 34-66% — yellow, 67%+ — green.

Why Personal Physiological Baseline Matters

The key principle is comparison with one's own baseline, not population "norms":

class PersonalBaseline:
    def __init__(self, lookback_days=30, percentile=50):
        self.lookback = lookback_days
        self.percentile = percentile

    def fit(self, history):
        self.hrv_baseline = np.percentile(history['hrv'], self.percentile)
        self.hr_baseline = np.percentile(history['resting_hr'], self.percentile)
        self.sleep_baseline = np.percentile(history['sleep_hours'], self.percentile)
        return self

    def deviation(self, today):
        return {
            'hrv_dev': (today['hrv'] - self.hrv_baseline) / self.hrv_baseline,
            'hr_dev': (today['resting_hr'] - self.hr_baseline) / self.hr_baseline,
            'sleep_dev': (today['sleep_hours'] - self.sleep_baseline) / self.sleep_baseline
        }

Sports Performance Prediction

Fitness-Fatigue model (Banister):

Performance(t) = Fitness(t) - Fatigue(t)
Fitness(t) = Σ TSS(i) × exp(-(t-i)/τ_fitness), τ=45 days
Fatigue(t) = Σ TSS(i) × exp(-(t-i)/τ_fatigue), τ=15 days

Personal τ are estimated via nonlinear optimization (scipy.optimize). The model designs tapering for competition. Saves up to 50% of coach's time on manual data analysis.

Early Illness Detection

def illness_risk_score(temp_deviation, hrv_drop, hr_elevation, symptom_report):
    if temp_deviation > 0.5 and hrv_drop < -0.2 and hr_elevation > 5:
        return 0.8
    return 0.1

Research Stanford COVID study shows: wearables detected COVID 0-2 days before symptoms in 63% of participants. Reduces medical consultation costs through early detection.

Long-term Progress

VO2max estimation via Firstbeat methodology (error ±3-5 ml/(kg·min)). Analysis of load dynamics over 12-52 weeks, adaptation through resting HR and HRV trends.

What We Provide

• Documentation on API and data model
• Dashboard with Recovery Score, prediction, and trends
• REST API for integration into your ecosystem
• Team training (3 sessions)
• Technical support for 3 months
• Source code of models (upon agreement)

Process of Work

1. Analytics: requirements gathering, selection of wearable APIs.
2. Design: data ingestion architecture, Recovery Score model.
3. Implementation: API integration, ML models (fitness-fatigue, illness detection).
4. Testing: validation on real data, A/B test.
5. Deployment: dashboard + REST API, team training.

Signal Processing: Artifact Filtering

Raw data from wearables contains outliers and gaps. For HRV intervals, we apply a Berthou filter: remove RR intervals deviating >20% from median of neighbors. Gaps are filled via cubic interpolation for gaps ≤5 minutes and forward extrapolation with confidence degradation for gaps >5 minutes. For accelerometer and gyroscope — median filter with window of 5 points removes impact artifacts during wear. Proper preprocessing reduces Recovery Score RMSE by 12–18% compared to raw data. Interpolation quality is verified on control gaps intentionally inserted into the test dataset.

Estimated Timelines

Get a consultation from our engineers: we will analyze your data and find the right architecture. Contact us to evaluate your project. Request demo access to a working prototype. Our experience: 5 years in the AI/ML solutions market, over 20 successful projects in sports analytics. We guarantee quality and timely delivery.

When does a time series forecasting model fail in production?

The CFO requests a quarterly sales forecast. An analyst builds SARIMA on three years of data, achieves MAPE 8.3% on the test set, and deploys. Two months later, the metric in production jumps to 23%. The root cause: the model was trained on pre‑COVID data, tested on a stable period, but production hit a promotion and supply chain disruption. Data leakage plus distribution shift—perfect notebook numbers, a broken forecast in reality. We have seen this pattern dozens of times across retail, fintech, and IoT. Our team has delivered more than 50 forecasting projects over 5+ years.

Incorrect cross-validation. Standard train_test_split for time series creates data leakage: the model sees future values during training. The correct approach is TimeSeriesSplit or walk‑forward validation with an expanding window.

Multiple seasonality. Hourly electricity consumption has three seasonalities: daily (24h), weekly (168h), yearly (8760h). SARIMA handles only one. Prophet can handle multiple but scales poorly to thousands of series.

Missing values and anomalies. A missing sensor reading is information (the sensor turned off), not NaN. Linear interpolation destroys this signal. Proper handling depends on the missingness mechanism.

Cold start. A new SKU in a 50,000‑item assortment has no history, yet a forecast is needed. Standard approaches fail; cross‑learning or feature‑based methods are required.

Why is model selection critical for your data?

Prophet (Meta) – a solid start for business data with clear seasonality and holidays. Fast setup, interpretable, built‑in outlier detection. Fails on irregular patterns and does not scale beyond ~10k series without parallelization.

Gradient boosting on features (LightGBM, XGBoost) – often underestimated. Engineer lags (t‑1, t‑7, t‑28), rolling means, day‑of‑week, holidays. The model trains on all series simultaneously, solving cold start via transfer learning. MAPE in retail often beats neural nets with proper feature engineering.

TFT (Temporal Fusion Transformer) – a transformer designed for interpretable forecasting with covariates. Built‑in variable selection, temporal attention, quantile outputs. Available in pytorch‑forecasting. Requires ~10,000+ records per series for stable training.

PatchTST – splits the series into patches (like ViT for images), capturing local patterns better than classic transformers. Excellent for long‑horizon forecasting (96–720 steps ahead).

N‑HiTS, N‑BEATS – attention‑free neural architectures, faster than TFT, competitive accuracy. N‑BEATS won the M4/M5 benchmarks for tasks without covariates.

How do we deploy TFT in production?

A typical pipeline via pytorch‑forecasting:

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="sales",
    group_ids=["store", "sku"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,  # 120 days
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,  # 28 days
    static_categoricals=["store_type", "category"],
    time_varying_known_reals=["price", "promo_flag"],
    time_varying_unknown_reals=["sales"],
    target_normalizer=GroupNormalizer(groups=["store", "sku"], transformation="softplus"),
)

A common mistake: the default target_normalizer (StandardScaler) breaks predictions for series with zero values (no sales on weekends). GroupNormalizer with transformation="softplus" is the correct choice for count data.

Case study: retail demand forecasting

A chain of 120 stores, 8,000 SKUs, 28‑day forecast horizon. The original system: SARIMA per series, MAPE 18.4%, retraining cycle – 6 hours. We replaced it with TFT on PyTorch + pytorch‑forecasting: a single model for all series, MAPE 11.2%, retraining – 40 minutes on an A10G. Feature importance via variable selection revealed that day_before_holiday influences more than the holiday date itself. Annual savings on inference alone exceeded $50,000.

Step‑by‑step configuration

1. Data collection and preparation. Handle missing values (mark NaN, interpolate only for technical failures), aggregate to required frequency, engineer covariates (holidays, promotions, prices).
2. Create TimeSeriesDataSet. Set group_ids (store + SKU), time index, forecast horizon. Choose target_normalizer based on target distribution.
3. Train a baseline. Prophet or LightGBM first – to understand complexity.
4. Train TFT. Use TemporalFusionTransformer with loss=QuantileLoss(), tune learning rate and hidden layer sizes.
5. Validate and interpret. Walk‑forward test, analyze variable selection, build attention heatmaps.

How to properly evaluate forecast quality?

RMSE alone is misleading – it over‑penalizes large values. Our standard set:

• MAPE – interpretable, unstable near zero.
• sMAPE – symmetric, avoids division by small numbers.
• MASE (Mean Absolute Scaled Error) – normalized relative to a naive seasonal forecast, ideal for comparing series of different scales.
• Pinball loss – for probabilistic forecasting, inventory management.

We guarantee a model card with these metrics on the validation set and walk‑forward results on at least 6 months of history.

What deliverables do you receive?

• Documentation of chosen architecture and hyperparameter rationale.
• Reproducible training and inference pipeline (Docker + CI/CD + Airflow/Prefect).
• Committed code with unit tests for key components.
• Team training: retraining, output interpretation, deployment of new versions.
• 3 months of post‑delivery support (consultations, bug fixes, fine‑tuning).

The model is deployed via FastAPI or Triton Inference Server. Retraining is scheduled (e.g., weekly) via Airflow with drift validation and automatic rollback if metrics deteriorate.

Process and timeline

We start with EDA: visualization, ADF test, STL decomposition, analysis of missing values and outliers. This takes 2–3 days but often reveals systemic data issues that block forecasting. Then we build a baseline (naive seasonal, Prophet), engineer features for LightGBM, and select a neural architecture if needed. Walk‑forward validation with a realistic horizon. Deployment via API with automatic retraining scheduled via Airflow or Prefect.

Timeline: MVP forecast on one data type – 3–6 weeks. Hierarchical forecasting system with automation – 2–5 months. Cost is calculated individually based on data volume, number of series, and required accuracy.

Our team consists of certified ML engineers (AWS ML Specialty, GCP Professional ML Engineer) with 5+ years on the market and over 50 completed forecasting projects. Contact us for a free analysis of your data – we will assess the task and provide initial recommendations within 1–2 days. Request a consultation to ensure your forecasts work in production, not just in a notebook.