Original OSD Algorithm

The Original OSD algorithm looks for seizure-like rhythmic motion in accelerometer data. It uses a frequency analysis approach and has been part of OpenSeizureDetector since the early versions.

How it works

At a high level:

1. Collect a time window of acceleration samples.
2. Convert the signal into the frequency domain (FFT).
3. Measure power in a seizure-focused frequency band.
4. Compare that power and its ratio against thresholds.
5. Report an algorithm alarm when both threshold checks pass.

flowchart LR
    A[Acceleration samples] --> B[FFT]
    B --> C[ROI power between AlarmFreqMin and AlarmFreqMax]
    B --> D[Total spectral power up to cutoff]
    C --> E[Check 1: ROI power > AlarmThresh]
    C --> F[Check 2: 10 x ROI power / spectrum power > AlarmRatioThresh]
    D --> F
    E --> G{Both checks true?}
    F --> G
    G -->|Yes| H[OSD algorithm alarm]
    G -->|No| I[OSD algorithm OK]

User settings

These are the key settings you can tune in current Android app versions:

Setting	What it changes
OsdAlarmActive	Enables or disables the Original OSD algorithm.
AlarmFreqMin	Lower edge of the target frequency band. Lower values make it more sensitive to slower rhythmic movement.
AlarmFreqMax	Upper edge of the target frequency band. Higher values include faster rhythmic movement.
AlarmThresh	Minimum movement power needed before OSD can trigger. Increasing it usually reduces sensitivity and false alarms.
AlarmRatioThresh	How concentrated movement must be in the target band. Higher values require a cleaner seizure-like frequency signature.

Practical tuning effect

flowchart TD
    A[Increase AlarmThresh] --> B[Harder to trigger]
    C[Decrease AlarmThresh] --> D[Easier to trigger]
    E[Increase AlarmRatioThresh] --> F[Needs clearer rhythmic pattern]
    G[Widen AlarmFreqMin or AlarmFreqMax range] --> H[Detects broader movement patterns]

Typical when to adjust

• Too many alarms during repetitive daily movements: increase AlarmThresh or AlarmRatioThresh slightly.
• Missed events with slower rhythmic movement: lower AlarmFreqMin slightly.
• Missed events with faster rhythmic movement: raise AlarmFreqMax slightly.

For context and history of the frequency method, see the project notes at: https://www.openseizuredetector.org.uk/?page_id=455

Technical details

This section provides the detailed technical description of the OSD algorithm and design rationale.

Algorithm sequence

The original OSD acceleration algorithm is described as:

1. Collect accelerometer data at 25 Hz for 5 seconds.
2. Run an FFT to convert the acceleration signal into frequency-domain bins.
3. Calculate average power in the Region Of Interest (ROI), typically 3-8 Hz.
4. Calculate average power in the overall spectrum (up to configured cutoff).
5. Calculate the ratio of ROI power to spectrum power.
6. Mark seizure-like movement when both checks pass:
   • ROI power above threshold
   • ROI/spectrum ratio above threshold
7. Raise warning/alarm states when seizure-like movement persists over consecutive analysis windows.

Region Of Interest (ROI) diagram

Detection diagram:

Conceptually, the ROI check is:

flowchart LR
    A[FFT spectrum 0 to Nyquist] --> B[ROI band AlarmFreqMin to AlarmFreqMax]
    A --> C[Reference spectrum power]
    B --> D[ROI power]
    C --> E[Spectrum power]
    D --> F[Power threshold check]
    D --> G[Ratio check with spectrum power]
    E --> G

Why 25 Hz sampling was chosen

The sampling rationale compares watch sampling choices 10, 25, and 100 Hz.

• 10 Hz gives Nyquist 5 Hz, which is marginal for seizure movement expected around 5-10 Hz.
• 25 Hz gives Nyquist 12.5 Hz, covering the target movement band with moderate compute cost.
• 100 Hz gives Nyquist 50 Hz, but significantly increases data volume and processing cost.

With a 5 second analysis window:

• Frequency resolution is Delta f = 1/T = 1/5 = 0.2 Hz.
• Maximum detectable frequency is f_max = f_s/2 = 25/2 = 12.5 Hz.

This is why the design summary selected 25 Hz sampling with 5 second windows.

Additional notes

• The algorithm documentation also describes an experimental multiple-ROI variant to improve detection of lower-frequency seizure movement (for example around 4 Hz).
• The documentation also notes expected FFT behavior, including strong low-frequency/DC components caused by gravity in accelerometer signals.

What the design is trying to detect

The design targets sustained rhythmic movement typical of tonic-clonic seizure phases, with energy often concentrated in approximately the 5-10 Hz region.

Important assumptions:

• Motion axis is not known in advance, so acceleration channels are combined into a single motion signal for analysis.
• Detection is based on rhythmic structure over several seconds, not a single instantaneous spike.
• A secondary low-frequency objective (for example breathing-related motion) was discussed but not part of the standard seizure detector path.

Sampling system and FFT trade-offs

The practical trade-off is between sampling frequency, maximum detectable frequency, and compute/data load.

Sampling frequency	Nyquist (f_s/2)	Interpretation
10 Hz	5 Hz	Considered borderline/too low for seizure motion around 5-10 Hz
25 Hz	12.5 Hz	Chosen compromise: covers target band with manageable processing
100 Hz	50 Hz	Plenty of bandwidth, but much higher data and compute load

For a 5 second analysis window:

• Samples per window at 25 Hz: N = 25 x 5 = 125.
• Frequency resolution: Delta f = 1/T = 0.2 Hz.

This means the detector updates in discrete analysis windows (every 5 seconds), each with 0.2 Hz spectral bin spacing.

Warning and alarm timing interpretation

Timing can be interpreted as:

• One analysis decision per 5 second window.
• Warning after persistent seizure-like windows over about 10 seconds.
• Alarm after additional persistent seizure-like windows.

In current app architecture, final app-level WARNING/ALARM timing is additionally controlled by Algorithm Voting timing settings (WarnTime and AlarmTime), so effective user-visible timing depends on both algorithm output persistence and global state-machine settings.

flowchart LR
    A[5 s analysis window] --> B[OSD alarm/no alarm result]
    B --> C[Consecutive positive windows accumulate]
    C --> D[Global warning and alarm timers]
    D --> E[User-visible OK/WARNING/ALARM state]

Expected FFT behavior

Expected FFT-derived behavior:

Observation	Why it happens	Practical implication
Strong DC or very low-frequency component	Accelerometer includes gravity offset	Low-frequency bins can dominate unless normalized/ratio-based checks are used
ROI band rises during rhythmic shaking	Seizure-like rhythmic motion concentrates energy in ROI	ROI power threshold is a primary detector gate
Ratio check improves specificity	Compares ROI concentration to broader spectrum	Helps reject broad-spectrum movement that is not seizure-like

Alternative multiple-ROI concept

The documentation also describes an alternative detector variant using more than one Region Of Interest (for example standard 3-8 Hz plus an additional lower-frequency band) to improve sensitivity for some partial or lower-frequency patterns.

This is presented as an experimental extension rather than the baseline default path.

flowchart TD
    A[FFT spectrum] --> B[ROI 1: standard seizure band]
    A --> C[ROI 2: lower-frequency auxiliary band]
    B --> D[Compute band powers and ratios]
    C --> D
    D --> E[Combine criteria for detection decision]

Historical context

The algorithm write-up was created in 2015 and updated in 2024 to better match Garmin/PineTime implementation while retaining the original sampling and FFT rationale.