Start Your Search
PS06 - Digital Health Innovation Across Canada (ID 41)
- Event: e-Health 2019 Virtual Meeting
- Type: Panel Session
- Presentations: 1
- Coordinates: 5/28/2019, 02:30 PM - 03:30 PM, Area 1
PS06.02 - Detecting Depression via Multimodal Neural Networks with an Automated Evaluation (ID 287)
There is mounting evidence that the technology fueled by machine learning has the potential to detect, and substantially improve treatment of complex mental disorders such as depression. We developed a framework capable of detecting depression with minimal human intervention: AiME (Artificial Intelligence Mental Evaluation). AiME consists of a short human-computer interactive evaluation and artificial intelligence, namely deep learning, and can predict whether the participant is depressed or not with satisfactory performance. Due to its ease of use, this technology can offer a viable tool for mental health professionals to identify symptoms of depression, thus enabling a faster preventative intervention. Furthermore, it may alleviate the challenge of interpreting highly nuanced physiological and behavioral biomarkers of depression by providing a more objective evaluation.
We collected data from 671 participants who performed a human-computer interactive evaluation composed of interview questions where participants were recorded by a webcam and a microphone while they responded to questions relating to their mental well-being. The evaluation also contained an anonymous demographics questionnaire (age, sex, ethnicity, etc.) as well as a brief, multiple-choice, mental health questionnaire in order to provide additional data and ground-truth validation. The evaluation took approximately five minutes, and data from the demographics questionnaire, video responses, and mental health questionnaires were stored and accessed in accordance with HIPAA compliance standards. We developed a multimodal deep learning neural network model that used video data, audio data, and word content from participants responses, as well as demographics and other metadata. These data were used as adjacent inputs to the model to perform binary classification on whether participants were depressed. The scores from PHQ-9 were used as the ground truth such that a PHQ-9 score of 10 was used as a threshold for depression. Computations were implemented using Keras with a TensorFlow backend. We experimented with three variations of our model that allowed us to compare performances within our framework and with results from prior work in the literature. These variations include two binary classification models as well as a regression model. The classification models were trained on 365 exams using a binary cross-entropy loss function and an independent set of 91 exams were left for a testing phase. The output of the model (predicted y) was rounded to construct a binary vector consisting of ones (depressed) and zeros (non-depressed) and was compared against the true values (true y)another binary vector built from the PHQ-9 scores.
We used various metrics to assess the performance of our models, including: accuracy, AUROC (Figure 1), specificity and sensitivity. According to all metrics, our models successfully classified depressed versus non-depressed individuals well above chance level. Two representative epochs reached high specificity and sensitivity values (87.77% and 86.81% respectively) and, it is possible to adjust the threshold value at which a prediction is considered positive to achieve desired levels of specificity and sensitivity.
There are significant physiological differences between individuals with depression and non-depressed individuals and our results suggest effectiveness in detecting depression with a neural network model with minimal human intervention
140 Character Summary:
A deep learning neural network model that observes human audio/visual responses can be used to detect depression without human intervention.