Thalamocortical dysrhythmia detected by machine learning

Thalamocortical dysrhythmia (TCD) is a model proposed to explain divergent neurological disorders. It is characterized by a common oscillatory pattern in which resting-state alpha activity is replaced by cross-frequency coupling of low- and high-frequency oscillations. We undertook a data-driven approach using support vector machine learning for analyzing resting-state electroencephalography oscillatory patterns in patients with Parkinson’s disease, neuropathic pain, tinnitus, and depression. We show a spectrally equivalent but spatially distinct form of TCD that depends on the specific disorder. However, we also identify brain areas that are common to the pathology of Parkinson’s disease, pain, tinnitus, and depression. This study therefore supports the validity of TCD as an oscillatory mechanism underlying diverse neurological disorders.

We used multinomial logistic regression model with ridge parameter of 1.0E-8.

Weighting of Data
In order to determine whether imbalanced class sizes play a role in the accuracy of the model retrieved for tinnitus/control, pain/control, Parkinson's/control, and depression/control data, we randomly select subgroups of the healthy control subjects (N = 264) similar in size to, respectively, the tinnitus group (N = 153), chronic pain group (N = 78), PD group (N = 31), and the major depression group (N = 15). This weighted dataset was then used to generate a prediction model and to model accuracy values. This was done 100 times and the resulting weighted model accuracy statistics were averaged across all trials and compared to our test (i.e. unbalanced/unweighted) model.

Tinnitus.
A comparison between the unweighted model and the weighted model did show a significant different for accuracy rate (F = 33.64, p < .001). However, no significant effect was obtained for the TPR, FPR and ROC between the unweighted and weighted models. A significant effect was obtained for κ-statistic (F = 64.77, p < .001), RMSE (F = 17.47, p < .001), and MAE (F = 13.81, p = .001)(see Fig. 3S).

Pain.
A comparison between the unweighted model and the weighted model did not show a significant different for accuracy rate. A significant effect was obtained for the TPR (F = 6.78, p = .013), the FPR (F = 208.67, p < .001) and the ROC (F =10.74, p = .002) between the unweighted and weighted models. A significant effect was also obtained for κ-statistic (F = 69.69, p < .001) but not for MAE and RMSE (see Fig. 3S).

Parkinson.
A comparison between the unweighted model and the weighted model did not show a significant different for accuracy rate and the TPR. A significant effect was demonstrated for the FPR (F = 55.53, p < .001), the ROC (F = 8.10, p = .007), an RMSE (F = 92.79, p = .007), between the unweighted and weighted models. No significant effect was also obtained for κ-statistic and MAE (see Fig. 3S).
Overall, the differences obtained between the weighted and unweighted models are minor. For tinnitus and depression we only see an increase in correctly classified subjects of 2.54% and 5.60%, respectively, while no significant effect was obtained for pain or Parkinson's disease. For tinnitus, pain, and Parkinson's, the results are not consistent over the different outcome measures (correctly classified, incorrectly classified, TPR, FPR, ROC, κ-statistic, RMSE, and MAE) of the SVMlearning approach. However, for depression, the weighted model shows a weak yet a significant improvement over all outcome measures.