TL;DR
This paper introduces a novel approach to predict individuals' emotional states and mental health questionnaire scores using passive data from smartphones and wearables. By combining Hidden Markov Models (HMMs) for data imputation with Transformer networks for time-series forecasting, the method achieves stable and accurate predictions, paving the way for more proactive mental health monitoring and intervention.
What Problem Does This Paper Address?
Monitoring the emotional states of individuals with psychiatric conditions is notoriously challenging. Traditional methods rely on subjective self-reports and infrequent clinical assessments, which fail to capture the dynamic and often subtle fluctuations in mood. This lack of continuous, objective monitoring can lead to delayed interventions, increasing the risk of adverse outcomes. The authors aim to bridge this gap by leveraging passively collected behavioral data from everyday devices to provide a more objective and real-time understanding of emotional well-being.
Methods
The core of this research lies in a three-stage pipeline:
- Hidden Markov Models (HMMs) for Missing Data Imputation: Given the inherent missingness in passively collected sensor data, HMMs are employed to model the underlying temporal structure and infer posterior probabilities of hidden states. This approach is crucial for handling the incomplete nature of the data without requiring explicit imputation beforehand. The authors used a heterogeneous HMM implementation to accommodate both continuous and discrete sensor data.
- Transformer-Based Time-Series Forecasting: A Transformer model, specifically inspired by the Informer architecture, is used for sequence forecasting. Transformers, with their attention mechanisms, are adept at capturing long-range dependencies in sequential data, making them ideal for modeling temporal patterns in behavioral data. This stage uses the HMM's posterior probabilities as input to learn temporal relationships and predict future states.
- Classification for Target Prediction: Finally, the output from the Transformer model (contextual embeddings or predicted future states) is fed into classification models (e.g., XGBoost, Random Forest) to predict specific targets: emotional valence and Patient Health Questionnaire-9 (PHQ-9) scores.
Datasets
The study utilized a large, heterogeneous dataset collected through the Evidence-Based Behavior (eB2) app. This dataset comprises passive behavioral data (step count, distance, sleep hours, app usage, location, time at home) and self-reported data (emotions, PHQ-9 responses) from over 4,400 patients across 8 distinct cohorts, including those with common mental disorders, eating disorders, high suicidal risk, and control groups. The data spans 8 years (2016-2023).
Key Insights and Ideas
- Passive Data as a Powerful Biomarker: The research strongly supports the idea that everyday behavioral data captured passively can serve as reliable biomarkers for emotional states and mental health indicators, reducing reliance on subjective self-reports.
- Transformer's Superiority in Temporal Modeling: The Transformer architecture, with its attention mechanism, demonstrates superior performance in capturing long-range temporal dependencies compared to traditional HMMs for time-series forecasting, leading to more stable predictions.
- HMMs as a Bridge for Missing Data: HMMs effectively handle the pervasive missingness in sensor data, providing a robust foundation for subsequent advanced modeling techniques like Transformers.
- Predicting Future Mental Health States: The ability to predict emotional valence and PHQ-9 scores a day in advance highlights the potential for early risk detection and timely intervention.
- Question-Specific Predictability in PHQ-9: Certain PHQ-9 questions, particularly those related to suicidal ideation, show significantly higher predictability, underscoring the model's potential for targeted risk assessment.
Key Results
- Emotional Valence Forecasting: The XGBoost classifier achieved high accuracy (0.93) and an ROC AUC of 0.98 for predicting emotional valence 1 day in advance.
- Emotional State Change Detection: The model demonstrated an ROC AUC of 0.87 for predicting emotional state changes 1 day in advance.
- PHQ-9 Score Forecasting: For question 9 (suicidal ideation), the model achieved an accuracy of 0.9 and an ROC AUC of 0.77.
- Model Stability: The combination of HMM preprocessing and Transformer forecasting led to enhanced stability in multivariate time-series forecasting compared to other methods.
- Change Detection: Jensen-Shannon (JS) divergence outperformed entropy for emotional change detection, with an ROC AUC of 0.87 for patients with mental disorders.
Limitations and Future Work
The authors acknowledge several limitations:
- High Missing Data Percentage: Despite HMMs, a significant percentage of missing data remains in both passive and active variables, particularly for emotional valence and PHQ-9 responses.
- Data Imbalance: The dataset exhibited an imbalance in emotion reporting, with a higher prevalence of negative emotions, potentially biasing the model.
- Limited PHQ-9 Responses: The PHQ-9 responses were sporadic and limited in number, posing a challenge for robust prediction.
- Short Prediction Horizon: While the model can predict a day in advance, extending this horizon for more proactive intervention remains a future goal.
Future work should focus on:
- Improving Interpretability: Understanding the contribution of individual variables and exploring the impact of adding new sensor data (e.g., heart rate, oxygen saturation).
- Extending Prediction Horizons: Developing methods to forecast further into the future for more comprehensive monitoring.
- Refining Predictive Accuracy: Further optimizing models and exploring more sophisticated feature engineering.
- Addressing Data Imbalance and Missingness: Investigating advanced techniques to mitigate these challenges.
- Real-world Deployment: Exploring the practical implementation of such systems for clinical use.
Summary updated