Bipolar disorder is a mental health condition that affects, according to the French National Authority for Health, at least 1 to 2.5% of the French population. It is characterized by marked mood changes, including depressive episodes and hyperthymic episodes, which are periods of elevated mood and increased energy. Bipolar disorder is mainly divided into two subtypes. Type I involves depressive and manic or hypomanic episodes, while type II involves depressive and hypomanic episodes.
Diagnosis of this disorder is generally based on a clinical assessment performed by a psychiatrist, using reported symptoms and observed clinical signs.
“However, this assessment can be challenging, as bipolar disorder may be confused with depressive disorder," explains Inès Tahir, PhD candidate in AI for healthcare at CEA-Leti.
“Similarly, confirming the subtype can be difficult until more specific characteristic episodes have been diagnosed. This can have serious consequences, as therapeutic strategies differ depending on the case. For example, medication used to treat depression may have deleterious effects in people with bipolar disorder."
Two complementary brain imaging methods
Inès Tahir's PhD thesis specifically aims to help physicians establish a diagnosis. Her work is part of a clinical study led by Professor Mircea Polosan — Professor at Université Grenoble Alpes, Head of the Psychiatry Department at Grenoble Alpes University Hospital, Head of a FondaMental Expert Center, and thesis director — within a broader collaborative project with the Holon Institute of Technology in Tel Aviv, Israel. Named BipoNIRS, this study aims to identify relevant biomarkers for the classification of bipolar disorder and its subtypes. In the longer term, the goal is to support earlier diagnosis.
Inès Tahir contributes to this objective by analyzing the collected data to identify biomarkers of bipolar disorder across the full set of acquired signals. She is also pursuing this work by proposing a frontal EEG-fNIRS configuration based on the most relevant biomarkers. This configuration could facilitate the development of a dedicated clinical device for routine use.
EEG consists of placing electrodes on the scalp to measure the brain's electrical activity. fNIRS, on the other hand, relies on optical sensors, known as optodes, to estimate concentrations of oxygenated and deoxygenated hemoglobin in the brain by analyzing how light is scattered and absorbed through tissue. These two complementary methods have been investigated by CEA-Leti.
AI-based analysis of EEG and fNIRS data
The study conducted by Inès Tahir is based on a cohort of 25 healthy subjects, 21 patients with bipolar disorder type I, and 25 patients with bipolar disorder type II.
All participants wore a headset combining EEG and fNIRS. They completed a protocol involving a “visual emotional task", designed to explore the cognitive and emotional processing of different stimuli. This type of task generally requires more time in people with bipolar disorder, due to their difficulties in regulating emotions.
To process the EEG and fNIRS data, she developed an AI model capable of learning to classify the different subjects.
“This approach represents an innovative contribution, as it leverages the complementarity of EEG and fNIRS to identify the most relevant biomarkers of bipolar disorder and its subtypes, while taking into account the constraints associated with the limited number of subjects in the study," explains Inès Tahir."
The algorithm was designed to account for the limited sample size, reduce overfitting and potential bias, and preserve good generalization ability.
A model that distinguishes bipolar disorder type I from type II
The model was then tested on this sample. It successfully identified relevant biomarkers and distinguished individuals with bipolar disorder type I from those with type II and from healthy subjects. In particular, combining EEG and fNIRS signals significantly improved classification performance compared with EEG alone.
More specifically, in the configuration where EEG covered the whole head, accuracy reached:
- 92% for distinguishing patients with bipolar disorder type I from healthy subjects,
- 85% for distinguishing patients with bipolar disorder type II from healthy subjects,
- 85% for distinguishing between the two subtypes.
The study also confirmed the complementarity of EEG and fNIRS. When the two modalities were combined, classification remained possible using only frontal electrodes and optodes. In this configuration, fNIRS compensates for the difficulty of using frontal EEG signals alone, which are often affected by ocular artifacts.
Performance reached:
- 82% for discriminating patients with bipolar disorder type II from healthy subjects,
- 75% for distinguishing between subtypes.
These results were obtained while maintaining satisfactory overall discrimination performance. They support the relevance of a compact, portable system compatible with clinical practice.
Promising Future Perspectives
These initial results open up several promising perspectives.
“To go further, it would also be important to study patients' relatives, including parents, children, brothers and sisters, in larger cohorts," note Anne Planat-Chrétien, researcher at CEA-Leti, and Professor Polosan.
“It would also be relevant to include other psychiatric disorders, such as depression or schizophrenia. This would help refine the specificity of the approach and enable bipolar disorder to be differentiated from other conditions."
Discover the publication:
https://www.nature.com/articles/s41398-026-03858-1
This work was conducted in collaboration with
CHU Grenoble Alpes and Université Grenoble Alpes.