Major depressive disorder (MDD) and bipolar disorder (BD) exhibit overlapping depressive presentations, making early differentiation difficult and misdiagnosis clinically consequential. Entropy-based resting-state electroencephalography (EEG) markers were evaluated for differentiating drug-naïve patients with stable MDD, converted BD, and diagnosed BD.

Resting-state EEG was acquired from 92 drug-naïve patients: 50 with stable MDD, 16 who converted to bipolar II disorder over 3 years, and 26 with diagnosed BD. Band-scale entropy and cross-sample entropy (CSE) were analyzed. Group differences were tested by analysis of covariance adjusted for depressive severity; classification was evaluated by nested cross-validation with permutation testing.

Relative to stable MDD, diagnosed BD showed higher theta- and alpha-band entropy in frontal and central regions and lower values for selected intra- and inter-regional CSE measures. Two CSE measures correlated positively with non-planning impulsiveness in diagnosed BD. Converted BD showed intermediate values without significant pairwise differences. Prospective discrimination of stable MDD versus converted BD was modest (balanced accuracy, 65.4%; area under the receiver operating characteristic curve [ROC-AUC], 0.661). Classification was strongest for stable MDD versus diagnosed BD, yielding 78.9% pooled accuracy, 76.6% balanced accuracy, and an ROC-AUC of 0.832 when entropy features were combined with depressive severity; combined entropy outperformed either single-family configuration.

Entropy-derived EEG measures differentiated stable MDD from diagnosed BD and suggested a weaker, intermediate profile in converted BD. They may serve as candidate adjunctive biomarkers for mood disorder differentiation, although larger prospective cohorts are required to clarify their value in predicting conversion to bipolar II disorder.

Late-life suicide is the most serious consequences of late-life depression (LLD). Nevertheless, suicidal behavior is complex and hard to predict. With the help of MRI scans and machine learning algorithm, we aim to examine the neural signatures of suicidality in patients of LLD.

We recruited 83 patients of LLD with a mean age of 68.8 years, where 48 were suicidal (26 with suicidal ideation and 22 with past suicide attempts). Cross-sample entropy (CSE) analysis was employed to analyze the resting-state function MRI data. Three-dimensional CSE volume in 90 region-of-interest of the brain in each participant was input into convolutional neural networks (CNN) to test the classification accuracy of suicidality.

After six-fold cross-validation, we found several regions in the default mode, fronto-parietal, and cingulo-opercular resting-state networks to have a mean accuracy above 75% to predict suicidality. Moreover, the models with right amygdala and left caudate provided the most reliable accuracy in all cross-validation folds, signifying their unique roles in late-life suicide.

Our results provide potential targets for intervention or biomarkers in late-life suicide. More research must be conducted to consolidate our results with scalable implementation in clinical setting.