Making oscillation detection more robust

Background:

Neural oscillations are important for understanding cognitive functions. To quantify them, certain methods, including Better OSCillation detection (BOSC), distinguish oscillatory activity from non-oscillatory 1/f background activity and derive detection thresholds in order to disregard most background signal. When successful, this produces detection criteria that are fairly calibrated across frequencies. However, if the background estimate is misaligned, this can backfire and potentially introduce a frequency bias.

New method:

The optimized BOSC method incorporates several improvements after testing each independently and as combinations before comparing them all together with the standard BOSC method. The improvements in question are: removing high-power values across frequencies, using median rather than mean power values, and robust regression.

Results:

The new BOSC method showed enhanced performance when using shorter time windows and when substantial power existed at one end of the measured spectrum. Synthetic signals were used to demonstrate further versatility and the limitations of the new method.

Comparison with existing methods:

The standard BOSC method fared reasonably well aside from some extreme edge cases. Outcomes suggested that at very short time windows, or when artifacts or lopsided power spectra are a concern, the optimized BOSC method could result in a more selective fit that shows greater alignment with the colored-noise background signal.

Conclusion:

The standard BOSC method performs well in many typical scenarios, but the optimized version is ideal for less conventional scenarios and addresses many of the shortcomings of the standard method in these cases.