ANS Predictor
Wearable symptom forecasting.

N-of-1 longitudinal design

A single post-Lyme patient wearing a Polar Grit X2 every night. Raw RR intervals are processed into 13 HRV features across 3 time-domain and frequency-domain families. Each morning, symptom severity is scored across 5 dimensions using a standardized diary (DSQ-PEM adapted).

The autonomic pattern targeted by this predictor is consistent with dysautonomia documented in post-infectious syndromes. → Scientific basis

Feature selection: Forward selection per target — each symptom model independently selects its own best features from the 13 candidates across 3 lag windows (t0, t-1, t-2). No shared feature set imposed.

Validation: Leave-one-out cross-validation (LOO-CV) for primary AUC. 1,000× bootstrap for confidence intervals. No train/test split — every data point serves as both training and validation exactly once.

Models: Logistic regression, random forest, and gradient boosting trained on each target independently. Best model selected by LOO-CV AUC.

Multi-symptom prediction performance

Five independent models, each predicting a different symptom dimension from nocturnal HRV data.

Three algorithms, one winner

All models trained on the severity target with the same feature set. LOO-CV evaluation.

Temporal structure of autonomic signals

Spearman correlations between nocturnal HRV features and next-day symptom severity at different lag windows. Lag 0 = same night, Lag 2 = two nights prior.

Automated nightly collection

Every night at 06:00 UTC, a GitHub Actions workflow pulls fresh data from the Polar AccessLink API. Raw RR intervals are processed into 13 HRV features using NeuroKit2. The resulting JSON is committed to the repository and deployed automatically to this page.

When a new diary entry is pushed, a separate workflow triggers model retraining with the updated dataset. Leave-one-out cross-validation runs on the full history, and the predictor coefficients in polar_live.json are updated.

Stack: Python · scikit-learn · neurokit2 · Polar Grit X2 · GitHub Actions · CSV

How well does each model separate good days from bad days?

A ROC curve shows how well each model separates symptomatic days from asymptomatic ones across every possible decision threshold. The top-left corner is perfect; the diagonal is a coin flip. The dot marks the operating point chosen for clinical use (Youden's J: the best trade-off between catching true cases and avoiding false alarms).

Side-by-side accuracy across symptoms and models

Each bar shows the model's accuracy (AUC) at predicting one symptom, with its uncertainty range (95% confidence interval) drawn as a horizontal line. An AUC of 1.0 means perfect prediction; 0.5 is random chance. The bars let you compare models side by side for each symptom.

Where exactly does each model make mistakes?

A confusion matrix shows exactly where each model makes mistakes. For each symptom and model the table reports: how many truly bad days were correctly flagged (True Positives), how many were missed (False Negatives), how many good days were wrongly flagged (False Positives), and how many good days were correctly left alone (True Negatives).

What does each model actually use?

This section shows which physiological signals each model relies on most, and — for logistic regression — the direction of each signal's effect (positive bars push toward "symptom present", negative bars push toward "symptom absent"). For the tree-based models we show feature importance: how much each feature contributes to the model's splits, always positive.

Scientific basis

The ANS Predictor is grounded in three lines of published evidence: autonomic dysfunction as a measurable feature of post-infectious syndromes, HRV as a predictive signal for symptom burden, and the statistical case for individual-level modelling over population averages.