Prediction Algorithm¶

Strimma predicts glucose trajectory using dampened velocity extrapolation — a weighted least-squares velocity estimate with exponential dampening.

Overview¶

The prediction algorithm answers: "Given recent glucose behavior, where is glucose heading in the next 15–30 minutes?"

It uses three stages:

1. Weighted velocity estimation — compute how fast glucose is changing, emphasizing recent data
2. Dampened projection — extend the trajectory forward, assuming velocity naturally decreases over time
3. Threshold crossing detection — find when (if ever) the projection crosses low or high thresholds

Stage 1: Weighted Velocity Estimation¶

Data Window¶

The algorithm uses readings from the last 12 minutes (LOOKBACK_MS = 720,000 ms). At least 2 readings are needed.

Exponential Time Weighting¶

Recent readings carry more weight than older ones:

weight(t) = e^(DECAY × t)

where t is minutes before now (negative, so recent readings get higher weights) and DECAY = 0.35.

Effect: The last 4 minutes hold approximately 84% of the total weight. The full 12-minute window is preserved for better slope estimation, but the velocity is dominated by recent behavior.

Weighted Linear Regression¶

A weighted least-squares regression computes the slope (velocity) of glucose over time:

v₀ = (Σw × Σw·t·y - Σw·t × Σw·y) / (Σw × Σw·t² - (Σw·t)²)

where:

• w = weight for each reading
• t = time (minutes)
• y = glucose (mg/dL)

The result v₀ is the estimated velocity in mg/dL per minute.

Velocity Sanity Gate¶

If the computed velocity exceeds ±9.0 mg/dL/min (MAX_VELOCITY), the prediction is rejected as a sensor artifact. Normal physiological glucose change rarely exceeds this rate.

Stage 2: Dampened Projection¶

Instead of projecting glucose as a straight line (which over-predicts), Strimma uses a dampened model where velocity naturally decreases:

x(t) = x₀ + (v₀ / DAMP) × (1 - e^(-DAMP × t))

where:

• x₀ = current glucose (mg/dL)
• v₀ = current velocity (mg/dL per minute)
• DAMP = 0.05 — dampening coefficient
• t = minutes into the future

Properties¶

• At t = 0: x(0) = x₀ — prediction starts at current glucose (natural anchor)
• Initial slope equals v₀ — matches current rate of change
• Velocity halves at approximately 14 minutes
• As t → ∞: x → x₀ + v₀/DAMP — glucose approaches a maximum excursion

Why Dampening?¶

Straight-line projection causes false alarms. If glucose is rising at 2 mg/dL/min, a straight line predicts it will rise 60 mg/dL in 30 minutes — but in reality, glucose trends usually slow down as the body responds.

The dampened model was specifically chosen to handle V-recovery scenarios: when glucose drops sharply (treating a low) and then bounces back, a straight-line or polynomial model predicts a false high from the upward bounce. The dampened model correctly predicts the bounce will decelerate.

Stage 3: Threshold Crossing Detection¶

If the current glucose is in range (between bgLow and bgHigh):

1. Walk through the prediction minute by minute
2. Check if the predicted glucose crosses bgLow (going low) or bgHigh (going high)
3. Return the first crossing with:
   • Type: LOW or HIGH
   • Minutes until crossing
   • Predicted glucose at the crossing point

Predicted values are clamped to 18–540 mg/dL to stay within physiologically plausible bounds.

Parameters¶

Sensor Interval Compatibility¶

The algorithm works with any CGM reading interval:

• Libre 3: 1-minute readings — the 12-minute window contains ~12 data points, excellent velocity estimation
• Dexcom G6/G7: 5-minute readings — the 12-minute window contains 2–3 data points, sufficient for velocity estimation
• Irregular intervals: The weighted regression handles non-uniform spacing naturally

Visual Output¶

On the Graph¶

The prediction appears as a dashed line extending from the most recent reading into the future. Each predicted point is computed at 1-minute intervals up to the configured prediction horizon (15 or 30 minutes).

In the BG Header¶

When a threshold crossing is detected, a prediction pill appears:

• Low in X min — red-tinted background, red text
• High in X min — amber-tinted background, amber text

In the Notification¶

The notification subtitle includes a compact warning (e.g., "Low 10m") when applicable.