Deep Learning Characterizes Depression and Suicidal Ideation from Eye Movements

Abstract

Identifying physiological and behavioral markers for mental health conditions is a longstanding challenge in psychiatry. Depression and suicidal ideation, in particular, lack objective biomarkers, with screening and diagnosis primarily relying on self-reports and clinical interviews. Here, we investigate eye tracking as a potential marker modality for screening purposes. Eye movements are directly modulated by neuronal networks and have been associated with attentional and mood-related patterns; however, their predictive value for depression and suicidality remains unclear. We recorded eye-tracking sequences from 126 young adults as they read and responded to affective sentences, and subsequently developed a deep learning framework to predict their clinical status. The proposed model included separate branches for trials of positive and negative sentiment, and used 2D time-series representations to account for both intra-trial and inter-trial variations. We were able to identify depression and suicidal ideation with an area under the receiver operating curve (AUC) of 0.793 (95% CI: 0.765-0.819) against healthy controls, and suicidality specifically with 0.826 AUC (95% CI: 0.797-0.852). The model also exhibited moderate, yet significant, accuracy in differentiating depressed from suicidal participants, with 0.609 AUC (95% CI 0.571-0.646). Discriminative patterns emerge more strongly when assessing the data relative to response generation than relative to the onset time of the final word of the sentences. The most pronounced effects were observed for negative-sentiment sentences, that are congruent to depressed and suicidal participants. Our findings highlight eye tracking as an objective tool for mental health assessment and underscore the modulatory impact of emotional stimuli on cognitive processes affecting oculomotor control.

Figures (5)

• Figure 1: Experimental setup and deep learning pipeline: (A) The stimulus sequence presentation for a single trial of the sentence viewing task. (B) Schematic overview of the data collection setup, with only the eye-tracking setup depicted. (C) The two different strategies for segmenting eye-tracking trials. Top: Signals aligned to onset of the final word. Bottom: Signals aligned to the subject's response and reaching back to 1200 ms before the response. (D) Histogram distribution of response times with respect to final-word onset, for the entire dataset. Highlighted are the median and mode of the distribution, which determined the trial segmentation. (E) The process of creating input samples by bootstrapping 200 sets of 30 trials per participant. Each set is input to the deep learning model as a 2-channel 2D input, reflecting both inter-trial and intra-trial variability. Channels correspond to horizontal (x) and vertical (y) movements. (F) The 2D input is processed by an Inception inception (convolutional) block and then concatenated before the classifier. MLP stands for multi-layer perceptron. $P_{=1}$ is the output probability of the sample belonging to the positive (non-control) class. The final subject-level prediction is derived by the proportion of input trial sets scoring $P_{=1}>0.5$.
• Figure 2: Task-wise performance of response-model: (A) Receiver-operating curve (ROC) plots per binary test setting. Each plot compares the performance of four model variants trained on separate group setups, with AUC denoted in the respective legend. For example, the left-most figure depicts and compares the performance of four variants on the same evaluation task, i.e., CvDS. Each of those variants was trained on a different classification objective, as indicated by their naming. (B) Violin plots of group-wise model predictions per training configuration. Each participant is represented by a dot indicating the model-assigned probability of belonging to the positive class, averaged across 10 random seeds.
• Figure 3: Model interpretability and ablations: Analysis is done here for the main response-model, trained on the CvDS task. (A) Ablation study with respect to sentence sentiment (left) and input directionality (right). For each scenario, we train a control model by shuffling the attribute of interest. (B) Time-resolved model attributions over raw gaze trials, computed with the Integrated Gradients sundararajan2017axiomatic method. Positive scores correspond to higher attribution to the positive class, namely DS. Colors reflect participant groups: C: green, D: blue, S: red. Each time-series represents the average model attribution over all available samples from subjects of the respective group. (C) Model weights extracted from the first fully connected (FC) classification layer of the model. Based on the model architecture, this layer's input is a concatenated vector representation of both (x, y) directions and (positive, negative) conditions, thus letting us examine where the model attends to render the output predictions.
• Figure 4: Sensitivity to Modeling Parameters: (A) Ablation study with respect to the duration of the experiment in number of trials. Shading denotes 95% CI derived through bootstrap resampling, see also Table \ref{['tab:classification_performance']}. (B) Analysis of model sensitivity to the number of bootstrapped trials in each constructed set. Error bars denote 95% CI. (C) Analysis of model sensitivity to the number of subject-wise trial sets used in training. $N$ refers to the number of sets created separately for each sentiment category. The evaluation setup was fixed to 200 sets per subject and sentiment category. Shading denotes 95% CI. (D) Group-wise fixation density maps computed for all negative sentences, from 100 ms before response to the moment of the response, and plotted as a histogram in projected screen dimensions. We applied 20% zoom to the image center for better visualization.
• Figure 5: Performance summary for the reading-model: (A) Receiver-operating curve (ROC) plots per binary test setting. Each plot compares the performance of 4 model variants trained on separate group setups, with AUC denoted in the respective legend. (B) Violin plots of group-wise model predictions per training configuration. Each participant is represented by a dot indicating the model-assigned probability of belonging to the positive class, averaged across 10 random seeds. (C) Time-resolved model attributions over raw gaze trials, computed with the Integrated Gradients [23] method. Positive scores correspond to higher attribution to the positive class, namely DS. Colors reflect participant groups: C: green, D: blue, S: red. Each time-series represents the average model attribution over all available samples from subjects of the respective group.