← Back to log

Epilepsy vs Space Weather — Completed Correlation Analysis (2015–2026)

Posted 2026-07-17

Analysis Complete — First Results Published

This report presents the results of our first full correlation analysis between epilepsy-related Google Trends data and space-weather variables over 4,076 consecutive days (2015-03-18 → 2026-07-15).


Data: How We Built the Dataset

Google Trends Pipeline

The epilepsy search-interest series was constructed by pulling daily data from Google Trends in overlapping 90-day windows, then stitching the windows together using a normalisation procedure. Each query returns 90 days of relative interest values (0–100 scale), and consecutive windows were aligned at their overlap points to produce a continuous 4,076-day time series.

Space-Weather Variables

Matched to the same 4,076-day date range from public NASA/NOAA sources:

Variable Source Resolution
Kp NOAA SWPC 3-hourly → daily mean
Ap NOAA SWPC 3-hourly → daily
Dst WDC Kyoto hourly → daily mean
Solar wind speed NASA OMNI hourly → daily mean
IMF B (average) NASA OMNI hourly → daily mean
F10.7 solar flux NOAA daily
Sunspot number SILSO daily

Preprocessing

  • Both series were detrended using STL decomposition (365-day seasonal period) to remove annual cycles and long-term drift
  • Outliers were winsorised at ±3 standard deviations
  • Final dataset: 4,076 matched daily observations — no missing days, no gaps

Methodology

  1. Cross-correlation function computed at lags τ = −14 to +14 days (Pearson and Spearman)
  2. Benjamini–Hochberg FDR correction at α = 0.05 across all lags × variables
  3. Surrogate-data validation: 1,000 random permutations per variable to establish empirical significance
  4. All code in Python (pandas, SciPy, statsmodels)

Key Findings

Top 10 Significant Correlations (F10.7)

All top results cluster around F10.7 solar flux at positive lags of +5 to +14 days. This means elevated solar radio flux today is associated with slightly higher epilepsy-related search activity 5–14 days later.

The table below shows the strongest Pearson correlations that survived FDR correction:

Lag (days) Pearson r p-value FDR p-value Significant
+11 +0.1417 < 0.000001 < 0.000001 ✓
+13 +0.1413 < 0.000001 < 0.000001 ✓
+12 +0.1407 < 0.000001 < 0.000001 ✓
+10 +0.1403 < 0.000001 < 0.000001 ✓
+14 +0.1391 < 0.000001 < 0.000001 ✓
+9 +0.1377 < 0.000001 < 0.000001 ✓
+7 +0.1366 < 0.000001 < 0.000001 ✓
+8 +0.1362 < 0.000001 < 0.000001 ✓
+6 +0.1362 < 0.000001 < 0.000001 ✓
+5 +0.1348 < 0.000001 < 0.000001 ✓

Bar Chart — Top 10 F10.7 Correlations

Pearson r by Lag (F10.7 vs Epilepsy Searches)
0.13 0.135 0.138 0.140 0.142 .1391 +14 .1413 +13 .1407 +12 .1417 +11 .1403 +10 .1377 +9 .1362 +8 .1366 +7 .1362 +6 .1348 +5 Lag (days after solar event) Pearson r

Surrogate Test — All 7 Variables

The permutation test confirms that all 7 space-weather variables produce signals stronger than random shuffles would generate. However, the effect sizes are small in absolute terms.

Variable Real r Empirical p Significant
Kp −0.0385 0.0120 ✓
Ap −0.0330 0.0230 ✓
Dst +0.0331 0.0410 ✓
Solar wind speed −0.0419 0.0050 ✓
IMF B (avg) −0.0392 0.0080 ✓
F10.7 flux +0.1333 0.0000 ✓
Sunspot number +0.0741 0.0000 ✓

Timeline Chart

Data Coverage — 4,076 Days (2015-03-18 → 2026-07-15)
2015-03
2018
2021
2024
2026-07

90-day sliding windows stitched at overlap points · 4,076 matched daily observations · No gaps


Effect Size Table (All Variables at Optimal Lag)

Correlation Summary — Optimal Lag per Variable
VariableBest LagPearson rp-valueFDR SigEffect
F10.7 flux+11 days+0.1417< 0.000001✓Weak positive
Sunspot number+9 days+0.07410.0000✓Very weak positive
Dst−3 days+0.03310.0410✓Negligible
Solar wind speed−5 days−0.04190.0050✓Negligible negative
IMF B (avg)−2 days−0.03920.0080✓Negligible negative
Kp−1 day−0.03850.0120✓Negligible negative
Ap−1 day−0.03300.0230✓Negligible negative

Interpretation

The analysis reveals a weak but statistically significant positive association between F10.7 solar radio flux and epilepsy-related search interest, peaking at lag +11 days (r = +0.1417).

Key Points

  • Not null, not strong. The signal is real (confirmed by FDR + surrogate tests), but the effect size is modest (r ≈ 0.14)
  • Lag structure is smooth. Correlations at lags +5 through +14 are all positive and similarly sized — not a single spike, but a broad elevation
  • Other variables show only trace signals. Kp, Ap, Dst, solar wind, and IMF-B show statistically significant but practically negligible associations (|r| < 0.05)
  • Compare to literature. Earlier Chizhevsky-era work reported stronger effects in experimental settings; modern clinical EEG studies often find null results. Our finding sits between these extremes

Limitations

  1. Daily resolution — finer-grained (hourly) data may reveal stronger effects
  2. Single keyword — "epilepsy" is only one term; a broader health-term panel is needed
  3. No geographic breakdown — Google Trends aggregates globally; regional effects may be diluted
  4. Observational, not causal — correlation ≠ causation; multiple confounders possible

Next Steps (Planned)

  • Expand to a panel of 20+ health-related search terms
  • Acquire hourly-resolution space-weather data for fine-grained lag testing
  • Test regional sub-series (US, Europe, India) if data availability permits
  • Publish replication analysis on independent 2026–2027 data
  • Release methodology whitepaper + code on GitHub

Analysis completed: 2026-07-17. Data period: 2015-03-18 → 2026-07-15 (4,076 days). Status: Completed.