BD-Transformer: A Transformer-Based Approach for Bipolar Disorder Classification Using Audio

Mohamed Ramadan, Hazem Abdelkawy, Mustaqueem, Alice Othmani

Abstract: Bipolar disorder, named also manic depression, is a mental disorder perceived by extreme mood swings that include abnormally emotional lows (depression) and highs (hypomania or mania episodes). According to the World Health Organization, 46 million people around the world have bipolar disorder, including 2.8% of the U.S. population. The risk of suicide in bipolar disorder over a period of 20 years is high, 6% died by suicide, while 30–40% engaged in self-harm. Bipolar disorder (BD) is a significant public health issue and computer-aided diagnosis systems are needed for the diagnosis and the follow-up of patients. In this paper, a new Transformer-based approach for BD classification based on audio data is proposed. Our proposed approach outperforms all existing approaches for BD classification on the turkish audio-visual bipolar disorder corpus by achieving an accuracy of 88.2% and a F1-score of 87.8%.