.png){kind=logo}
Minds AI
Research Methods and Results
Comparative Results of Minds AI Filter
versus State-of-The-Art Filtering Techniques
in emotion classification
Minds AI Filter (patent pending) techniques are based on the observation that neuronal oscillations exhibit patterns called synchrony (Wolf 2017). This synchrony emerges from coupled neurons communicating with each other locally and globally across the brain (Wang 1995). Any interference that does not fit in with the global synchrony, as a true neural signal would, is suppressed or removed.
Comparisons are made in terms of artificial intelligence classification accuracy of positive or negative emotional valence using the DEAP (Koelstra 2011) benchmark data set. Similar results have been seen applied to alternative classifications. Combining multiple filters with MAI has only shown to improve downstream signal quality, never reducing it, if any affect is seen. Comparison results of artifact and noise removal performance to be released shortly. A real-time application is available here for testing artifact and noise injection.
Bandpass Filter
Reference
Filters: When, Why, and How (Not) to Use Them - PubMed
MAI Filter Comparison
The DEAP dataset comes pre-filtered with a bandpass at 3 and 47 Hz. However, the addition of the Minds AI Filter increases predictive accuracy by an average of 6 percent for all trials and 17% for trials of which there was an effect, if any. No trials saw a reduction in accuracy with the addition of the Minds AI Filter.
​
Conclusion
Band Pass and Notch Filters (de Cheveigné, 2019) target specific frequency ranges but do not take into account spatial artifacts, structural noise, or signal quality across channels. They can be applied before, as this dataset has done, but we recommend applying alternate filters after the MAI Filter, which improves overall signal structure before frequency selection.
Principal Component Analysis (PCA)
Reference
Principal Component Analysis | SpringerLink
MAI Filter Comparison
On average, MAI filter performed 9% better in balanced classification accuracy but individually up to 25%. The median gain was 19%.
​
Conclusion
PCA assumes linearity and ranks components by variance instead of signal relevance. Removing low-variance components can discard meaningful neural activity. PCA Requires extensive hyperparameter tuning to decide how many components to retain.
The MAI Filter requires only a single hyperparameter and makes no assumptions about variance and dimensionality, operating on the full signal and removing the need for component-based reconstruction.
Independent Component Analysis (ICA)
Reference
Independent Component Analysis of Electroencephalographic Data
MAI Filter Comparison
The MAI Filter performed better than the ICA when dynamic or unexpected artifacts are introduced in the data. ICA performs better when the artifact components of the signal stay constant for the individual or trial, however when applying a configured filter to alternative data, the Minds AI filter outperforms ICA.
​
Conclusion
ICA offers no clear indication of how much variance each component contributes. It requires selecting among multiple decomposition algorithms with inconsistent results. ICA is not viable for real-time because it requires the user to identify which component corresponds to noise since ICA does not rank components by the amount of variance it explains as PCA does. It requires subjective decisions about which components to remove, risking loss of true brain signal.
Artifact Subspace Reconstruction (ASR)
Reference
MAI Filter Comparison
MAI outperformed ASR, when used alone and with further gains in classification accuracy when combining the two.
​
Methods
We performed channel-wise Root-Mean-Square to extract reference data from raw data. Applied IIR filter to reference data to reject artifact components and reconstruct viable data.
​
Conclusion
MAI Filter doesn't require baseline segments but instead uses constraints to reconstruct the desired signal based on what physics that signal should obey. This makes it usable even when no clean stretches of data exist in contrast to what the ASR filter requires for optimal performance. ASR is not as generalizable or commercially viable as MAI Filter.
Common Average Referencing (CAR)
Reference
MAI Filter Comparison
CAR performed better on constant noise, but on all other artifacts and noise, the Minds AI Filter outperformed CAR. Combining the two allows for further gains than either stand alone.
​
Conclusion
CAR reduces global noise by subtracting the average signal across all channels, assuming noise is uniformly distributed, which is a rare occurrence. The MAI Filter adapts to non-uniform interference without flattening the data
Least Mean Square (LMS)
​
Reference
MAI Filter Comparison
The Least Mean Squares (LMS) filter (He et al., 2004) is the most similar to the Minds AI Filter but it relies on a desired signal which is not constant or objective. Combining the two allows for further gains than either stand alone.
​
Conclusion
LMS is used to remove predictable noise but in real time yet very little noise is predictable. The MAI filter uses physics-informed principles, accounting for the intractability of brain activity, to separate true brain signals from noise and artifacts.
In Progress Filter Comparisons
​
​Blind Source Separation (BSS)
​
Embedded Empirical Mode Decomposition (EEMD)
​
Joint Blind Source Separation (JBSS)
​
Root-Mean-Square (RMS)
​
Singular Spectrum Analysis (SSA)
Citations
​
-
Bertrand, Olivier, Françoise Perrin, and Jean Pernier. “A Theoretical Justification of the Average Reference in Topographic Evoked Potential Studies.” Electroencephalography and Clinical Neurophysiology/Evoked Potentials Section, vol. 62, no. 6, 1985, pp. 462–464.
-
de Cheveigné, Alain, and Israel Nelken. Filters: When, Why, and How (Not) to Use Them. Lab. des Systèmes Perceptifs, École Normale Supérieure, Paris; UCL Ear Institute, UK; 2019.
-
Jolliffe, I. T. Principal Component Analysis. Springer, 2002.
-
He, Ping, Guangyuan Wilson, and Cuntai Guan. “Removal of Ocular Artifacts from Electroencephalogram by Adaptive Filtering.” Medical and Biological Engineering and Computing, vol. 42, no. 3, 2004, pp. 407–412.
-
Makeig, Scott, et al. “Independent Component Analysis of Electroencephalographic Data.” Advances in Neural Information Processing Systems, vol. 8, 1996, pp. 145–151.
-
S. Koelstra et al., "DEAP: A Database for Emotion Analysis; Using Physiological Signals," in IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 18-31, Jan.-March 2012, doi: 10.1109/T-AFFC.2011.15
-
Singer, W. (2018), Neuronal oscillations: unavoidable and useful?. Eur J Neurosci, 48: 2389-2398. https://doi.org/10.1111/ejn.13796
-
Wang, DeLiang, "Emergent synchrony in locally coupled neural oscillators," in IEEE Transactions on Neural Networks, vol. 6, no. 4, pp. 941-948, July 1995, doi: 10.1109/72.392256.