Machine learning-based prediction of bipolar II disorder conversion in major depressive disorder using spatiotemporal EEG analysis: a 3-year prospective cohort study

Major depressive disorder (MDD) and bipolar II disorder (BD II) are difficult to distinguish during depressive episodes, and delayed recognition of BD II exposes patients to antidepressant monotherapy and is associated with poorer outcomes. This prospective cohort study examined whether spatiotemporal dynamics of baseline resting-state electroencephalography (EEG) differed among patients with stable MDD, patients who later converted to BD II, and healthy controls (HCs), and whether these dynamics predicted conversion. Ninety-seven drug-naïve patients with MDD were followed for approximately three years and classified as stable MDD ( n = 75) or converted BD II ( n = 22); 120 age-, sex-, and education-matched HCs were assessed at baseline. From 62-channel eyes-open resting-state EEG, inter-regional dissimilarity (brain-region adversarial spatiotemporal information-based similarity [braSTIBS]) and non-randomness indices were derived at word length m = 4. Group differences in these indices were tested with Welch’s analysis of variance, with p -values corrected using the false discovery rate (FDR) method; classification of stable MDD versus converted BD II was evaluated by repeated nested cross-validation with permutation testing. Baseline symptom severity did not differ significantly between the patient groups, and non-randomness indices did not differ significantly among groups. braSTIBS differed significantly among groups for the Left Temporal–Fronto-Central (FDR-adjusted p = 0.03) and Left Temporal–Centro-Parietal (FDR-adjusted p = 0.01) connections. Values were lower in stable MDD than in converted BD II and HCs, with large effects versus converted BD II (Hedges’ g = −0.82 and −0.93), whereas converted BD II and HCs did not differ significantly. In an exploratory analysis within stable MDD, trait anxiety correlated positively with Left Temporal–Centro-Parietal braSTIBS ( r = 0.255; 95% confidence interval [CI]: 0.026–0.452). A multilayer perceptron discriminated the two patient groups with a pooled area under the receiver operating characteristic curve of 0.692 (95% CI: 0.544–0.830; permutation p = 0.01). Despite similar baseline symptom severity between the patient groups, reduced Left Temporal inter-regional dissimilarity characterized stable MDD, whereas converted BD II did not differ significantly from HCs. braSTIBS is a candidate prognostic EEG marker of diagnostic trajectory requiring validation in larger multi-site cohorts.

Authors

Institutions

Publication Details

Journal
International Journal of Bipolar Disorders
Published
2026-11-19
DOI
https://doi.org/10.1186/s40345-026-00448-4
Primary Topic
Bipolar Disorder and Treatment
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Machine learning-based prediction of bipolar II disorder conversion in major depressive disorder using spatiotemporal EEG analysis: a 3-year prospective cohort study

Sungkean Kim, Hyeon-Ho Hwang, Ji Sun Kim
International Journal of Bipolar Disorders
Bipolar Disorder and Treatment
article

Machine learning-based prediction of bipolar II disorder conversion in major depressive disorder using spatiotemporal EEG analysis: a 3-year prospective cohort study

Sungkean Kim, Hyeon-Ho Hwang, Ji Sun Kim
article en

Abstract

Major depressive disorder (MDD) and bipolar II disorder (BD II) are difficult to distinguish during depressive episodes, and delayed recognition of BD II exposes patients to antidepressant monotherapy and is associated with poorer outcomes. This prospective cohort study examined whether spatiotemporal dynamics of baseline resting-state electroencephalography (EEG) differed among patients with stable MDD, patients who later converted to BD II, and healthy controls (HCs), and whether these dynamics predicted conversion. Ninety-seven drug-naïve patients with MDD were followed for approximately three years and classified as stable MDD ( n = 75) or converted BD II ( n = 22); 120 age-, sex-, and education-matched HCs were assessed at baseline. From 62-channel eyes-open resting-state EEG, inter-regional dissimilarity (brain-region adversarial spatiotemporal information-based similarity [braSTIBS]) and non-randomness indices were derived at word length m = 4. Group differences in these indices were tested with Welch’s analysis of variance, with p -values corrected using the false discovery rate (FDR) method; classification of stable MDD versus converted BD II was evaluated by repeated nested cross-validation with permutation testing. Baseline symptom severity did not differ significantly between the patient groups, and non-randomness indices did not differ significantly among groups. braSTIBS differed significantly among groups for the Left Temporal–Fronto-Central (FDR-adjusted p = 0.03) and Left Temporal–Centro-Parietal (FDR-adjusted p = 0.01) connections. Values were lower in stable MDD than in converted BD II and HCs, with large effects versus converted BD II (Hedges’ g = −0.82 and −0.93), whereas converted BD II and HCs did not differ significantly. In an exploratory analysis within stable MDD, trait anxiety correlated positively with Left Temporal–Centro-Parietal braSTIBS ( r = 0.255; 95% confidence interval [CI]: 0.026–0.452). A multilayer perceptron discriminated the two patient groups with a pooled area under the receiver operating characteristic curve of 0.692 (95% CI: 0.544–0.830; permutation p = 0.01). Despite similar baseline symptom severity between the patient groups, reduced Left Temporal inter-regional dissimilarity characterized stable MDD, whereas converted BD II did not differ significantly from HCs. braSTIBS is a candidate prognostic EEG marker of diagnostic trajectory requiring validation in larger multi-site cohorts.

International Journal of Bipolar Disorders
Soonchunhyang University (KR), Soonchunhyang University Hospital Cheonan (KR), Hanyang University (KR)
Bipolar Disorder and Treatment
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon