Automatic Detection of Depression from Stratified Samples of Audio Data

Depression is a common mental disorder which has been affecting millions of people around the world and becoming more severe with the arrival of COVID-19. Nevertheless proper diagnosis is not accessible in many regions due to a severe shortage of psychiatrists. This scarcity is worsened in low-income countries which have a psychiatrist to population ratio 210 times lower than that of countries with better economies. This study aimed to explore applications of deep learning in diagnosing depression from voice samples. We collected data from the DAIC-WOZ database which contained 189 vocal recordings from 154 individuals. Voice samples from a patient with a PHQ-8 score equal or higher than 10 were deemed as depressed and those with a PHQ-8 score lower than 10 were considered healthy. We applied mel-spectrogram to extract relevant features from the audio. Three types of encoders were tested i.e. 1D CNN, 1D CNN-LSTM, and 1D CNN-GRU. After tuning hyperparameters systematically, we found that 1D CNN-GRU encoder with a kernel size of 5 and 15 seconds of recording data appeared to have the best performance with F1 score of 0.75, precision of 0.64, and recall of 0.92.