Classification of depression tendency from gaze patterns during sentence reading

Depression is a common and disabling mental health disorder, which impacts hundreds of millions of people worldwide. Current diagnosis methods rely almost solely on self-report and are prone to subjectivity and biases. In recent years, computational psychiatry has employed advanced sensing technology, utilizing rich data, to train accurate algorithms to detect depression from passive, non-invasive physiological markers. Gaze-tracking is used to collect cognitive data with high temporal resolution and offers a surrogate to underlying processes such as attention distribution, making it particularly useful for classification of attention-related cognitive abnormalities, including depression. We used data from gaze-tracking while participants were engaged in sentence reading to build a classifier for depression tendency. We created sentences constructed to highlight expected attention biases in depression. We recorded gaze data during reading from a sample of 101 participants and analyzed the data as a raw time-series. We used the validated PHQ-9 questionnaire to obtain depression levels per participant. Using LSTMs (Long Short-Term Memory Artificial Neural Network) and Random Forest analysis techniques we were able to reach above chance classification (60+%) of depression tendency levels from the gaze patterns. Limitations: A replication with more participants is needed. Data was collected among undergraduate students and was conducted only in Hebrew. Individual assessment was not validated against clinical data. The results can lead to potential data-driven and accessible diagnosis tools that will support and monitor depression treatment and rehabilitation.