[Paper Review] MHDeep: Mental Health Disorder Detection System based on Body-Area and Deep Neural Networks

Shayan Hassantabar, Joe Zhang+2 more|arXiv (Cornell University)|Feb 20, 2021

Digital Mental Health Interventions40 references4 citations

TL;DR

MHDeep is a novel mental health disorder detection system that leverages wearable sensors from smartwatches and smartphones to collect physiological and behavioral data, using synthetic data augmentation and a grow-and-prune deep neural network (DNN) architecture to enable accurate, edge-deployable diagnosis of schizoaffective, major depressive, and bipolar disorders. It achieves 100% patient-level accuracy for schizoaffective and major depressive disorders using just 16–40 minutes of continuous data.

ABSTRACT

Mental health problems impact quality of life of millions of people around the world. However, diagnosis of mental health disorders is a challenging problem that often relies on self-reporting by patients about their behavioral patterns. Therefore, there is a need for new strategies for diagnosis of mental health problems. The recent introduction of body-area networks consisting of a plethora of accurate sensors embedded in smartwatches and smartphones and deep neural networks (DNNs) points towards a possible solution. However, disease diagnosis based on WMSs and DNNs, and their deployment on edge devices, remains a challenging problem. To this end, we propose a framework called MHDeep that utilizes commercially available WMSs and efficient DNN models to diagnose three important mental health disorders: schizoaffective, major depressive, and bipolar. MHDeep uses eight different categories of data obtained from sensors integrated in a smartwatch and smartphone. Due to limited available data, MHDeep uses a synthetic data generation module to augment real data with synthetic data drawn from the same probability distribution. We use the synthetic dataset to pre-train the DNN models, thus imposing a prior on the weights. We use a grow-and-prune DNN synthesis approach to learn both the architecture and weights during the training process. We use three different data partitions to evaluate the MHDeep models trained with data collected from 74 individuals. We conduct data instance level and patient level evaluations. MHDeep achieves an average test accuracy of 90.4%, 87.3%, and 82.4%, respectively, for classifications between healthy instances and schizoaffective disorder instances, major depressive disorder instances, and bipolar disorder instances. At the patient level, MHDeep DNNs achieve an accuracy of 100%, 100%, and 90.0% for the three mental health disorders, respectively.

Motivation & Objective

• Address the challenge of limited, hard-to-collect mental health datasets by leveraging synthetic data generation to augment real physiological sensor data.
• Develop a scalable, automated system for continuous mental health monitoring using commercially available wearable medical sensors (WMSs) on smartphones and smartwatches.
• Eliminate manual feature engineering by directly processing raw sensor streams through end-to-end deep neural networks.
• Design computationally efficient DNN models suitable for deployment on edge devices like smartphones, enabling real-time, pervasive mental health diagnosis.
• Achieve high diagnostic accuracy at both the data instance and patient level for three major mental health disorders: schizoaffective, major depressive, and bipolar disorders.

Proposed method

• Collects eight categories of physiological, motion, and environmental data from smartwatches and smartphones to form a comprehensive sensor dataset.
• Applies synthetic data generation to augment real data, ensuring the synthetic samples follow the same probability distribution as real data to improve model generalization.
• Uses a grow-and-prune DNN synthesis approach to jointly learn optimal network architecture and weights during training, enhancing model efficiency and accuracy.
• Pre-trains DNN weights using the synthetic dataset to impose a prior on the model parameters, improving performance on small real-world datasets.
• Employs three data partitions for evaluation: instance-level and patient-level classification using majority voting over 15-second data windows.
• Deploys trained models on edge devices for real-time inference, enabling continuous, non-invasive mental health monitoring.

Experimental results

Research questions

• RQ1 Can a deep learning system achieve high diagnostic accuracy for mental health disorders using only raw sensor data from commercial wearables, without manual feature engineering?
• RQ2 To what extent does synthetic data generation improve model performance when real-world mental health datasets are small and imbalanced?
• RQ3 How much continuous sensor data is required for reliable patient-level diagnosis of schizoaffective, major depressive, and bipolar disorders?
• RQ4 Can a grow-and-prune DNN architecture achieve high accuracy while remaining computationally efficient enough for edge deployment on smartphones and smartwatches?
• RQ5 Can the system achieve clinically meaningful accuracy at the patient level, sufficient for real-world clinical decision support?

Key findings

• MHDeep achieves an average test accuracy of 90.4% at the data instance level for distinguishing healthy individuals from those with schizoaffective disorder.
• At the patient level, MHDeep reaches 100% accuracy in classifying healthy individuals versus those with schizoaffective disorder using 40 minutes of continuous sensor data.
• For major depressive disorder, the system achieves 100% patient-level accuracy after 16 minutes of data collection.
• For bipolar disorder, the patient-level accuracy reaches 90.0% after 22 minutes of data, demonstrating strong performance despite higher diagnostic complexity.
• The grow-and-prune DNN synthesis method successfully learns both network architecture and weights, resulting in computationally efficient models suitable for edge deployment.
• Synthetic data generation effectively mitigates data scarcity, enabling robust pre-training and improving model generalization on limited real-world datasets.