Seizure Detection Algorithms

OpenSeizureDetector can use one or more different methods (‘Algorithms’) to detect seizures. Some aim to detect the shaking movement associated with a tonic-clonic seizure, and others detect more general conditions such as abnormal heart rate or a fall.

This section explains what each algorithm does and which settings you can tune.

How decisions are made

In normal operation, the app processes incoming data in analysis cycles and then combines algorithm outputs.

flowchart TD
    A[Incoming watch or data source samples] --> B[Run enabled algorithms]
    B --> C1[OSD result]
    B --> C2[ML result]
    B --> C3[Heart Rate result]
    B --> C4[Fall result]
    C1 --> D[Voting strategy and timing rules]
    C2 --> D
    C3 --> D
    C4 --> E[Fall immediate alarm override]
    E --> F[Final app state: OK, WARNING, or ALARM]
    D --> F

Where to configure algorithms in the app

Open the main screen menu (three dots).
Tap Settings.
Open Seizure Detector.
Use Seizure Detection Algorithms Selection to enable or disable algorithms.
Open Algorithm Settings to tune thresholds and timing.

Quick guidance

Start with defaults unless you already know you need different sensitivity.
Change one setting at a time, then monitor behavior for several days.
If false alarms increase after a change, revert that change before trying another.

High-level comparison

Use this as a practical guide when choosing which algorithms to enable.

Algorithm	Main signal used	Best at detecting	Typical strengths	Typical limitations	Key settings to tune
Original OSD	Wrist accelerometer frequency pattern	Rhythmic tonic-clonic type movement	Proven and transparent logic, works well for sustained rhythmic motion	Can false-alarm on repetitive non-seizure movement (for example brushing teeth or washing dishes)	AlarmFreqMin, AlarmFreqMax, AlarmThresh, AlarmRatioThresh
Machine Learning (ML)	Accelerometer waveform features learned by model(s)	Complex movement patterns captured by trained models	Often better balance of sensitivity vs false alarms, supports multiple models	Depends on model quality and thresholds, may need periodic tuning	MlSeizureProbabilityThresholdPct, MlAccelStdThresholdPct, model selection
Heart Rate Alarms	Heart rate trend and thresholds	Abnormal heart-rate events, including rapid changes or sustained abnormal rates	Adds physiological signal independent of movement pattern	Needs reliable continuous HR source, can alarm from non-seizure causes of HR change	HRThreshMin/Max, adaptive and average window/threshold settings
Fall Detection	Short-window acceleration min/max event pattern	Acute fall-like events (free-fall plus impact sequence)	Fast response to sudden fall-like events	Crude by design, can false-alarm on abrupt non-fall movements	FallThreshMin, FallThreshMax, FallWindow

Choosing a starting combination

PineTime users: start with Original OSD + ML.
Garmin users: start with Original OSD + ML + Heart Rate.
Fall risk concern: enable Fall Detection in addition to your normal seizure-detection algorithms.

flowchart TD
    A[Choose data source] --> B{Garmin available?}
    B -->|Yes| C[Enable OSD + ML + Heart Rate]
    B -->|No| D[Enable OSD + ML]
    C --> E{Fall risk concern?}
    D --> E
    E -->|Yes| F[Also enable Fall Detection]
    E -->|No| G[Keep current set]