Iman Azimi’s Post

Wearable devices allow us to collect multimodal data from individuals in everyday settings. In recent years, deep learning methods have been increasingly applied for early detection and prediction tasks, using objective physiological and behavioral data collected over time. But, here is a challenge: Existing deep learning methods are not sufficiently explainable to clarify the relationship between input signals and the model’s outcome. How do we uncover data-driven values that aren’t immediately obvious? How can we extract actionable insights from this data, for example, to enable early intervention? In our latest paper, published by IEEE BSN, we tackle this issue. We present an attention-based xAI method to predict health events and interpret multivariate time series data collected from multiple modalities (devices). The proposed method leverages self-attention mechanisms and graph attention layers to capture both temporal and inter-variable dependencies. Then, we evaluate the method using a case study on longitudinal mental health monitoring collected from college students via wearables over one year. A glance at the findings: The trained model achieves 78.62% accuracy for positive affect events prediction and 76.30% for negative affect events prediction using one week of wearable data preceding the events. For positive affect prediction, a forward-looking dependency was more important: the relationship between data farther from the event (e.g., Day 6 to the event) and data closer to it (e.g., Day 2 to the event). In contrast, backward dependency (e.g., Day 1 to Day 7) was more important for negative affect prediction. Additionally, low-intensity activity and changes in location were examples of important inputs for positive affect prediction, while REM sleep duration and time of being active during the day were examples of important inputs for negative affect prediction. This solution can provide insights into prediction tasks, expanding the potential of AI in interventions and empowering health professionals to make data-driven decisions. Special thanks to Yuning Wang, Zhongqi Yang, Amir M. Rahmani, and Pasi Liljeberg. Read the paper here: https://lnkd.in/d2B82ajR #AI #Wearable #ExplainableAI #HealthMonitoring  #DeepLearning #TimeSeries