PaperlayerView exact revision on arXiv

Source-linked AI summary

Impact of Physical Activity on Sleep:A Deep Learning Based Exploration

Aarti Sathyanarayana, Shafiq Joty, Luis Fernandez-Luque, Ferda Ofli, Jaideep Srivastava, Ahmed Elmagarmid, Shahrad Taheri, Teresa Arora

arXiv:1607.07034v1cs.LG

TL;DR

Sleep-quality prediction from physical activity remains difficult despite the growth of wearable actigraphy data. The paper compares deep learning and classical models using both activity-recognition features and raw accelerometer signals, finding that raw-data CNNs improve predictive value while simplifying preprocessing.

  • Problem

    The relationship between sleep quality and physical activity is important but not fully understood, while actigraphy analysis requires improved tools for research and clinical practice.

  • Method

    The study compares deep learning with logistic regression, support vector machines, random forest, and adaboost using RAHAR-derived features and raw accelerometer data.

  • Results

    An additional 8% predictive-value improvement was obtained with a convolutional neural network on raw wearable output compared with state-of-the-art classical approaches.

  • Takeaways & Limitations

    Deep learning on raw actigraphy data can eliminate data preprocessing and simplify the workflow for sleep research.

Abstract

from arXiv · show

The importance of sleep is paramount for maintaining physical, emotional and mental wellbeing. Though the relationship between sleep and physical activity is known to be important, it is not yet fully understood. The explosion in popularity of actigraphy and wearable devices, provides a unique opportunity to understand this relationship. Leveraging this information source requires new tools to be developed to facilitate data-driven research for sleep and activity patient-recommendations. In this paper we explore the use of deep learning to build sleep quality prediction models based on actigraphy data. We first use deep learning as a pure model building device by performing human activity recognition (HAR) on raw sensor data, and using deep learning to build sleep prediction models. We compare the deep learning models with those build using classical approaches, i.e. logistic regression, support vector machines, random forest and adaboost. Secondly, we employ the advantage of deep learning with its ability to handle high dimensional datasets. We explore several deep learning models on the raw wearable sensor output without performing HAR or any other feature extraction. Our results show that using a convolutional neural network on the raw wearables output improves the predictive value of sleep quality from physical activity, by an additional 8% compared to state-of-the-art non-deep learning approaches, which itself shows a 15% improvement over current practice. Moreover, utilizing deep learning on raw data eliminates the need for data pre-processing and simplifies the overall workflow to analyze actigraphy data for sleep and physical activity research.

I. INTRODUCTION

Sleep is essential to wellbeing, yet sleep assessment and the relationship between sleep quality and daily physical activity remain challenging. Wearable actigraphy offers continuous, less subjective activity and sleep monitoring that can support research and clinical analysis.

  • Motivation: Poor sleep can impair physical, emotional, and mental wellbeing and is associated with multiple health complications.Reported complications include insulin resistance, high blood pressure, cardiovascular disease, immune or metabolic compromise, mood disorders, and decreased cognitive function.
  • Current assessment: Polysomnography diagnoses major sleep disorders but is cumbersome, requires manual clinical interpretation, and is typically initiated after self-reported problems worsen.
  • Opportunity: Wearable devices are generating rapidly growing human activity datasets that improved analytic tools could leverage for clinical practice and health research.
  • Study approach: The paper models sleep–physical-activity relationships with deep learning, comparing activity-recognition-derived features with models trained directly on raw accelerometer data.The raw-data approach removes feature-space construction and data preprocessing from the workflow.
  • Research gap: PSG usually captures one night and does not account for daily physical activity, whose relationship with sleep quality remains strong, complex, and incompletely understood.
  • Actigraphy: Actigraphy uses wearable motion sensors to measure sleep and physical activity, enabling longitudinal monitoring and more reliable measurement than subjective diaries or logs.It is particularly useful for large cohorts and long-term disease studies where PSG is infeasible.

B. Research challenges of Actigraphy

Actigraphy analysis is constrained by manual configuration and feature-reduction steps that may discard task-relevant information. Deep learning is presented as an alternative because it can learn representations directly from raw sensor data while modeling temporal structure.

  • Research challenges of Actigraphy: RAHAR addresses a major actigraphy-analysis bottleneck by automatically preprocessing accelerometer data into activity levels for classical model building.Its feature construction improves sleep-quality predictive value by 15%.
  • Research challenges of Actigraphy: RAHAR is unsupervised, so it does not use task labels when constructing features and may learn less task-specific representations.
  • Research challenges of Actigraphy: RAHAR adds preprocessing and aggregates the data, potentially failing to exploit the richness of the original actigraphy signal.
  • Deep learning alternative: Deep learning can automatically learn abstract representations from raw features while optimizing for target prediction tasks.This capability motivates its use for sleep and health research.
  • Prior approaches: Prior wearable-sensor studies adapted convolutional, recurrent, and hybrid architectures to preserve temporal information and model activity dynamics.Examples include CNN–LSTM systems, temporal convolutions, hierarchical RNNs, and bidirectional LSTM-RNNs.

III. PRELIMINARIES

The paper defines activity time-series boundaries and uses sleep efficiency to represent sleep quality. These terms establish how periods of sleep, wakefulness, and sleep performance are interpreted.

  • Time-series terminology: A wearable activity time series contains awake and asleep periods, with sleep onset and awakening marking their boundaries.
  • Time-series terminology: Awake time is the interval from a sleep awakening time until the following sleep onset time.
  • Sleep quality: Sleep efficiency is defined as the ratio of total minutes asleep to total time in bed.The supplied equation passage identifies the denominator components as sleep-period length plus latency.
  • Sleep quality: Total sleep time excludes wakefulness during the sleep period, including wake after sleep onset, from the sleep-period duration.

B. Data Collection

The study uses wrist-worn ActiGraph recordings from adolescents and analyzes both RAHAR-derived activity features and raw triaxial accelerometer data. Sleep-quality labels are computed from actigraphy-derived sleep boundaries and efficiency.

  • B. Data Collection: The QUEST dataset contains actigraphy recordings from 92 adolescents collected over one week.Student volunteers from two high schools wore ActiGraph GT3X+ devices on their non-dominant wrists throughout day and night.
  • B. Data Collection: The clinical-grade ActiGraph GT3X+ samples activity at 30–100 Hertz and was evaluated against ActiLife version 6 results.The device’s effectiveness had previously been validated against clinical polysomnography.
  • B. Data Collection: Experiments compare RAHAR output with triaxial accelerometer data from the actigraph.
  • B. Data Collection: Sleep onset is the first minute of 15 continuous minutes of sleep after reported bedtime, while awakening is the last such minute followed by 30 minutes of activity.These definitions are used to automate sleep-period interpretation from accelerometer output.
  • B. Data Collection: Sleep is labeled Good when sleep efficiency exceeds 85%; otherwise it is labeled Poor.WASO counts movement lasting more than five consecutive minutes during the sleep period, while latency is sedentary time before onset.
  • B. Data Collection: RAHAR represents awake time as percentages across sedentary, light, moderate, and vigorous activity levels, producing a four-feature set.
  • B. Data Collection: Raw accelerometer signals are aggregated into minute-by-minute epochs, with the vertical axis used as a one-dimensional movement summary.The sleep period is detached from each time series to prevent autocorrelation.

B. Deep Learning Models

The paper models sleep quality from physical activity using deep neural networks that transform activity inputs into abstract representations before prediction. It contrasts these models with classical approaches and describes multilayer perceptrons as one such architecture.

  • Deep neural models transform activity sequences through nonlinear hidden layers before predicting sleep quality.The output layer defines a Bernoulli distribution over sleep-quality labels.
  • The models are trained by minimizing cross-entropy between predicted and target label distributions.Cross-entropy is equivalent to minimizing negative log-likelihood.
  • 1) Multi-Layer Perceptrons:: Multi-layer perceptrons use fully connected hidden layers to transform concatenated activity vectors with nonlinear activations.Without hidden layers, the model reduces to logistic regression; hidden layers provide capacity to model complex dependencies.

2) Convolutional Neural Network:

Convolutional neural networks address limitations of fully connected models by learning local, timing-invariant activity patterns. Convolution creates feature maps, while pooling retains prominent features with fewer parameters.

  • 2) Convolutional Neural Network:: Fully connected MLPs use the entire input and separate weights for each input dimension, whereas sleep quality may depend on local activity periods and timing-invariant patterns.The paper gives high-intensity exercise or games as an example of activity whose timing may vary.
  • 2) Convolutional Neural Network:: CNN filters compose activity measures from local time windows into higher-level feature maps.Different filters generate different feature maps across the activity sequence.
  • 2) Convolutional Neural Network:: Max-pooling reduces feature dimensionality while retaining the most important aspects of each feature map.For pooling length p = 2, zero padding preserves the number of features in the feature map.
  • 2) Convolutional Neural Network:: CNN convolution-pooling operations make extracted features invariant to where activities occur during awake time.This design uses fewer parameters than a fully connected counterpart.

3) Recurrent Neural Network:

Recurrent neural networks process activity sequences step by step, preserving temporal information through recurrent hidden states. LSTM extends this approach with gated memory blocks designed to capture long-range dependencies.

  • 3) Recurrent Neural Network:: RNNs process each activity measure sequentially by combining the current input with the previous hidden state.Their hidden states summarize the sequence history and support dynamic temporal behavior.
  • 3) Recurrent Neural Network:: Mean-pooling over recurrent hidden states was more effective than other methods in the experiments.
  • 3) Recurrent Neural Network:: Standard RNN training with backpropagation through time can suffer from vanishing or exploding gradients.Truncated backpropagation limits long-range dependency capture in simple RNNs.
  • 3) Recurrent Neural Network:: LSTM memory blocks combine a self-connected memory cell with input, output, and forget gates.These gates allow LSTM networks to capture long-range dependencies more effectively than simple RNNs.

V. EXPERIMENTS AND RESULTS

The experiments use fixed subject partitions and evaluate model predictions on a held-out test set. Training uses RMSprop with regularization and architecture-specific hyperparameter searches, while limited data constrains model depth.

  • V. EXPERIMENTS AND RESULTS: 70%:15%:15% of the data was partitioned into training, testing, and validation sets, with all reported results from the test set.The two data types used the same subjects in corresponding partitions.
  • V. EXPERIMENTS AND RESULTS: Neural models optimize cross-entropy with RMSprop, ReLU activations, dropout, and early stopping.The maximum number of epochs was 50, and RMSprop parameters used default settings.
  • V. EXPERIMENTS AND RESULTS: Dropout rates from 0.0 to 0.5 and minibatch sizes from 5 to 20 were experimentally evaluated.
  • V. EXPERIMENTS AND RESULTS: Increasing the number of MLP hidden layers worsened results because the training data amount was small.
  • V. EXPERIMENTS AND RESULTS: CNN and recurrent models were evaluated across multiple filter, pooling, unit, and pseudo-sequence configurations.CNN filter lengths and pooling lengths ranged from 2 to 5; recurrent pseudo-sequences used 25 to 100 time-slots.

B. Human Activity Recognition Data Results

The study compares deep learning with classical models on activity features produced by RAHAR, then evaluates deep learning on richer raw accelerometer inputs. MLP improves over classical baselines, while CNN and MLP perform best on raw data.

  • Human Activity Recognition Data Results: RAHAR reduced accelerometer time series to four aggregated activity features, making CNN, RNN, and LSTM unsuitable and motivating an MLP.RAHAR also removes the time component through aggregation.
  • Human Activity Recognition Data Results: MLP improved performance over logistic regression and marginally surpassed random forest on the RAHAR-derived feature space.The best MLP used a hidden layer of 15, minibatch size 5, and dropout ratio 0.3.
  • Human Activity Recognition Data Results: Deep learning on raw accelerometer epochs eliminates RAHAR preprocessing and supplies a richer input dataset for model building.The raw data were aggregated into one-minute epochs.
  • Human Activity Recognition Data Results: 46% AUC improvement from logistic regression was achieved by CNN and MLP on the raw accelerometer data.Both models performed best in this comparison.

2) Multi Layer Perceptron:

The study evaluates deep learning architectures for sleep-quality prediction, comparing feature-reduced and raw accelerometer inputs. MLP performs strongly on limited features, while CNN achieves the best reported raw-data results.

  • Multi Layer Perceptron: The study evaluates deep learning for sleep-quality prediction using both state-of-the-art feature processing and deep learning’s broader predictive capacity.These are the study’s two stated objectives.
  • Multi Layer Perceptron: A moderate AUC improvement over traditional classification models was observed on pre-processed, feature-reduced data.The improvement was described as small, while MLP succeeded despite highly limited features.
  • Multi Layer Perceptron: 7% better performance was achieved by MLP with raw data than with RAHAR data.The comparison used the expanded feature set represented by raw accelerometer data.
  • Multi Layer Perceptron: RNN and LSTM improved overall over logistic regression but were limited by vanishing gradients from data granularity and sequential dependencies.Longer-duration epochs showed preliminary potential to improve both models, but further exploration was outside the paper’s scope.
  • Multi Layer Perceptron: High sensitivity across models indicates strong capability to identify individuals with good sleep patterns from preceding awake activity.MLP and CNN also showed high specificity for identifying poor sleep patterns.
  • Multi Layer Perceptron: CNN performed best, with AUC 0.9456, F1 Score 0.9444, and Accuracy 0.9286.These metrics are reported for the overall best-performing model.

VII. DISCUSSION ON CLINICAL AND HEALTH INFORMATICS IMPORTANCE

Actigraphy and wearable activity data are presented as an opportunity for sleep research, but their growing use increases the need for reliable analytic tools. The study positions deep learning as a feasible approach that may improve predictive performance and simplify actigraphy analysis.

  • Actigraphy could produce a paradigm shift in studying sleep and physical activity, provided new algorithms and tools are developed for the growing data volumes.The paper connects this opportunity to data from patients, healthy populations, epidemiological studies, and consumer wearables.
  • Growing sleep-app use heightens the need for systematic wearable research because the quality and reliability of some activity-sensor applications remain uncertain.
  • 8% additional predictive-value improvement was achieved by a CNN on raw wearable output versus state-of-the-art classical classification approaches.The paper describes these results as demonstrating deep learning’s potential for actigraphy-based sleep research, although they were obtained on a small dataset.
  • Deep learning was feasible for sleep research using actigraphy while eliminating data preprocessing and simplifying the overall analytical workflow.
Loading 1607.07034v1…