Skip to main content

Medical Research Overview: Migraine Navigator

Focus: Longitudinal Tracking, N-of-1 Methodology, and Machine Learning Validity

1. Research Objective​

Migraine Navigator serves as a platform for high-resolution longitudinal tracking. The project explores how individualized (N-of-1) machine learning models can identify latent environmental and behavioral triggers that are often obscured in large-scale population studies due to physiological heterogeneity.

2. Statistical Methodology​

The platform addresses the classic "small data" challenge in personalized health through a phased architecture involving a User-Configurable Heuristic (Phase 1) and Machine Learning (Phase 2):

2.1 Gradient Boosting Decision Trees (GBDT) & The Hurdle Model​

Migation tracking data is inherently "zero-inflated"—patients have many more healthy days than sick days. Standard regression models often "average out" these zeros, leading to under-prediction of severe events. We utilize a Two-Stage Hurdle Model (Scikit-Learn) to address this:

  1. Binary Classification Stage: Estimates the probability of $Pain > 0$.
  2. Regression Stage: Estimates the log-severity of pain, conditional on $Pain > 0$.

This approach allows us to:

  • Handle Variance: Accurately model both the occurrence and the severity independently.
  • Capture Non-Linear Interactions: Identify synergistic risks (e.g., combined barometric pressure drops and sleep deprivation) that linear models (Logistic Regression) may underestimate.

2.2 Feature Engineering & Encoding​

  • Cyclical Temporal Encoding: Days of the week and months are transformed using sine/cosine transforms to preserve the mathematical proximity of cyclical boundaries (e.g., ensuring Monday is as "close" to Sunday as it is to Tuesday).
  • Meteorological Resolution: Integrating hourly historical and forecast data (Open-Meteo) to capture volatility markers, such as the stability index (24-hour barometric delta).

3. The N-of-1 Paradigm​

By focusing on the individual as their own control, the model avoids the "mean-field" error where population-level averages fail to capture a specific patient's idiosyncratic triggers. This methodology is particularly relevant for diseases with high symptomatic variance like migraines.

4. Potential Research Applications​

  • Trigger Identification: Quantifying the lag-time between weather events and symptom onset.
  • Medication Response Variability: Analyzing the "pain decay" curve of different acute interventions in a real-world setting.
  • Prodromal Analysis: Using autoregressive features (lagged pain states) to identify potential digital markers of the prodromal phase.
  • Explainable AI (XAI): Future integration of SHAP (SHapley Additive exPlanations) to decompose daily risk into marginalized feature contributions (e.g. visualizing the specific weight of "Sleep Debt" vs "Barometric Pressure" for a given prediction).

5. Data Privacy & Ethics​

To facilitate trust and ethical research:

  • Edge Computing: All model training and inference occur locally on the user's machine.
  • Zero-Server Persistence: This architecture demonstrates a path forward for "Privacy-Preserving ML" in highly sensitive medical domains.