[Paper Review] Towards Deep Learning Models for Psychological State Prediction using Smartphone Data: Challenges and Opportunities

Gatis Mikelsons, Matthew D. Smith+2 more|arXiv (Cornell University)|Nov 16, 2017

Mental Health Research Topics18 references16 citations

TL;DR

This paper proposes a deep learning model using smartphone GPS data to predict psychological stress levels, leveraging hand-crafted spatio-temporal metrics and temporal features. Evaluated on the StudentLife dataset, the model achieves improved classification performance through neural networks with batch normalization and dropout, highlighting the potential of deep learning for digital mental health monitoring despite challenges in label noise and population heterogeneity.

ABSTRACT

There is an increasing interest in exploiting mobile sensing technologies and machine learning techniques for mental health monitoring and intervention. Researchers have effectively used contextual information, such as mobility, communication and mobile phone usage patterns for quantifying individuals' mood and wellbeing. In this paper, we investigate the effectiveness of neural network models for predicting users' level of stress by using the location information collected by smartphones. We characterize the mobility patterns of individuals using the GPS metrics presented in the literature and employ these metrics as input to the network. We evaluate our approach on the open-source StudentLife dataset. Moreover, we discuss the challenges and trade-offs involved in building machine learning models for digital mental health and highlight potential future work in this direction.

Motivation & Objective

• To develop a deep learning model for predicting psychological stress levels using passive smartphone sensor data.
• To evaluate the effectiveness of neural networks in modeling stress from GPS-derived mobility patterns.
• To address challenges such as noisy self-reported labels and population heterogeneity in mental health modeling.
• To explore feature importance and model robustness using cross-validation on the StudentLife dataset.
• To identify key design trade-offs and future research directions in digital mental health applications.

Proposed method

• The model uses a four-layer fully connected neural network with tanh activations, batch normalization, and dropout regularization (rates 0.35, 0.25, 0.15) for feature learning.
• Input features include eight GPS-based spatio-temporal metrics (e.g., total distance, spatial coverage, entropy) and four one-hot encoded temporal features (weekend, term start/mid/end).
• Features are standardized per user to reduce individual variability and improve model generalization.
• Stress labels are derived from daily-averaged self-reports, rescaled into three classes: below-median, median, and above-median stress.
• The final layer uses softmax activation for multi-class classification into three stress levels.
• Model performance is evaluated using cross-validation to assess feature importance and generalization.

Experimental results

Research questions

• RQ1 Can deep learning models effectively predict psychological stress levels using only smartphone GPS data?
• RQ2 How do hand-crafted spatio-temporal GPS metrics compare in predictive power to temporal context features?
• RQ3 What is the impact of label noise and individual variability on model performance in mental health prediction?
• RQ4 How do regularization techniques like batch normalization and dropout affect model generalization in this context?
• RQ5 What are the key challenges in building robust, scalable models for real-world digital mental health applications?

Key findings

• The deep learning model with batch normalization and dropout achieved improved performance in classifying stress levels compared to baseline approaches, demonstrating the value of neural networks in modeling complex psychological states.
• The inclusion of temporal one-hot features (e.g., term start, mid-term) contributed meaningfully to model performance, indicating that contextual time markers enhance prediction accuracy.
• Spatio-temporal GPS metrics such as total distance covered and spatial coverage via convex hull were among the most informative predictors for stress classification.
• The model showed that individual-level standardization of features significantly improved generalization, reducing the impact of inter-user variability.
• Label noise due to self-reported mood and recall bias remains a major challenge, with classification performance limited by inconsistent user interpretations of stress scales.
• The study identifies population heterogeneity as a key obstacle to building a single generalizable model, suggesting individual tailoring or clustering as necessary future directions.