NapTune: Efficient Model Tuning for Mood Classification using Previous Night's Sleep Measures along with Wearable Time-series

Debaditya Shome, Nasim Montazeri Ghahjaverestan, Ali Etemad

Abstract

While mood is a long-lasting affective state critical to mental health, its classification using real-time physiological signals remains challenging due to data scarcity and the lack of historical context. We present NapTune, a novel prompt-tuning framework that integrates the previous night’s sleep measures with wearable time-series data (ECG, PPG, EDA). By utilizing a frozen, pre-trained encoder and lightweight learnable "prompts," NapTune achieves significant improvements in F1 scores—up to 11%—while maintaining high performance even with limited training data.

Introduction

Unlike fleeting emotions, "mood" is a persistent state that can last for hours or days, making it notoriously difficult to classify using only instantaneous physiological signals from wearables. Current research acknowledges that sleep is a cornerstone of emotional regulation; however, most automated mood recognition models ignore this vital historical context.

The primary hurdle in this field is data scarcity. Public datasets rarely provide high-frequency wearable signals paired with objective sleep metrics from the same users over time. Standard deep learning models typically require massive amounts of paired data to learn these complex relationships, making them impractical for niche multimodal tasks like mood classification.

Methods: The NapTune Framework

To bridge the gap between sleep history and real-time physiology, we introduce NapTune, a two-stage framework designed for efficiency and robustness:

  1. Self-Supervised Pre-training: A hybrid CNN-Transformer encoder is first trained on a massive corpus of unlabeled wearable data (including ECG, PPG, and EDA) using SimCLR, a contrastive learning framework. This allows the model to learn robust representations of physiological signals without requiring mood labels.
  2. Prompt-Tuning: Instead of traditional fine-tuning—which involves updating millions of parameters and risks overfitting on small datasets—the pre-trained encoder is frozen. Nine sleep-related measures (e.g., REM duration, sleep efficiency, apnea index) are projected into a vector space. This vector acts as a "prompt" inserted into each layer of the Transformer. Only these lightweight prompt parameters are trained for the final mood classification task.
The NapTune architecture: Sleep-related measures are projected as prompts into a frozen Transformer encoder.

Datasets

  • Pre-training: A diverse collection of datasets including WESAD, MIMIC AFib, CAPNO, BIDMC, DALIA, and signals from Fekri Azgomi et al. to cover multiple physiological modalities.
  • Downstream Task: The ECSMP dataset, currently the only known dataset containing physiological time-series paired with both mood labels (via the Profile of Mood States) and objective sleep measures from the previous night.

Key Results and Insights

The inclusion of sleep data via NapTune provided a substantial performance boost across all tested modalities. Most notably, the model treats tabular sleep data as a contextual guide for interpreting physiological signals—interpreting a heart rate signal differently depending on whether the user had a restful night or suffered from sleep apnea.

  • Performance Gains: ECG + Sleep achieved an F1 score of 0.76 (an 11% improvement over the baseline). PPG + Sleep reached an F1 score of 0.72 (a 9% improvement).
  • Sample Efficiency: NapTune is remarkably data-efficient. When training data was reduced to just 25%, NapTune maintained a 62% F1 score, whereas standard baselines saw their performance plummet.
  • The Role of REM: Ablation studies revealed that REM sleep duration had the highest impact on mood recognition accuracy, aligning with neurobiological theories regarding REM's role in emotional memory processing.
Performance comparison across varying training data percentages, highlighting NapTune's sample efficiency compared to traditional baselines.

Discussion and Limitations

While NapTune demonstrates the power of parameter-efficient tuning, some challenges remain. Electrodermal Activity (EDA) results were lower than ECG and PPG, likely due to low sampling rates and environmental noise sensitivity. Furthermore, the current model assumes sleep data is always available. Future work may involve using generative models to estimate missing sleep measures for users who do not wear trackers every night. Additionally, exploring other parameter-efficient methods, such as Adapters, could further optimize the integration of sleep and wearable modalities.

Summary updated

BibTeX
@inproceedings{shome2024naptune,
  author = {Debaditya Shome and Nasim Montazeri Ghahjaverestan and Ali Etemad},
  title = {{NapTune: Efficient Model Tuning for Mood Classification using Previous Night's Sleep Measures along with Wearable Time-series}},
  booktitle = {ICMI},
  year = {2024},
  doi = {10.1145/3678957.3685722},
}
Figure